diff --git "a/out/eval/tinyllama_full_arc_arxiv_mc/2407/results.json" "b/out/eval/tinyllama_full_arc_arxiv_mc/2407/results.json" new file mode 100644--- /dev/null +++ "b/out/eval/tinyllama_full_arc_arxiv_mc/2407/results.json" @@ -0,0 +1,103962 @@ +{ + "results": { + "arxiv_mc": { + "acc,none": 0.16583333333333333, + "acc_stderr,none": 0.01074491240245708, + "acc_norm,none": 0.23916666666666667, + "acc_norm_stderr,none": 0.012288331311441483 + }, + "arxiv_mc_2407": { + "acc,none": 0.2, + "acc_stderr,none": 0.04020151261036846, + "acc_norm,none": 0.33, + "acc_norm_stderr,none": 0.047258156262526045 + }, + "arxiv_mc_2408": { + "acc,none": 0.16, + "acc_stderr,none": 0.036845294917747115, + "acc_norm,none": 0.18, + "acc_norm_stderr,none": 0.03861229196653697 + }, + "arxiv_mc_2409": { + "acc,none": 0.14, + "acc_stderr,none": 0.03487350880197772, + "acc_norm,none": 0.24, + "acc_norm_stderr,none": 0.04292346959909283 + }, + "arxiv_mc_2410": { + "acc,none": 0.2, + "acc_stderr,none": 0.04020151261036846, + "acc_norm,none": 0.23, + "acc_norm_stderr,none": 0.042295258468165044 + }, + "arxiv_mc_2411": { + "acc,none": 0.24, + "acc_stderr,none": 0.04292346959909282, + "acc_norm,none": 0.34, + "acc_norm_stderr,none": 0.04760952285695235 + }, + "arxiv_mc_2412": { + "acc,none": 0.11, + "acc_stderr,none": 0.031446603773522035, + "acc_norm,none": 0.23, + "acc_norm_stderr,none": 0.04229525846816506 + }, + "arxiv_mc_2501": { + "acc,none": 0.14, + "acc_stderr,none": 0.03487350880197771, + "acc_norm,none": 0.24, + "acc_norm_stderr,none": 0.04292346959909284 + }, + "arxiv_mc_2502": { + "acc,none": 0.14, + "acc_stderr,none": 0.034873508801977704, + "acc_norm,none": 0.2, + "acc_norm_stderr,none": 0.040201512610368445 + }, + "arxiv_mc_2503": { + "acc,none": 0.17, + "acc_stderr,none": 0.0377525168068637, + "acc_norm,none": 0.2, + "acc_norm_stderr,none": 0.04020151261036843 + }, + "arxiv_mc_2504": { + "acc,none": 0.15, + "acc_stderr,none": 0.035887028128263734, + "acc_norm,none": 0.17, + "acc_norm_stderr,none": 0.0377525168068637 + }, + "arxiv_mc_2505": { + "acc,none": 0.19, + "acc_stderr,none": 0.03942772444036623, + "acc_norm,none": 0.24, + "acc_norm_stderr,none": 0.042923469599092816 + }, + "arxiv_mc_2506": { + "acc,none": 0.15, + "acc_stderr,none": 0.03588702812826371, + "acc_norm,none": 0.27, + "acc_norm_stderr,none": 0.0446196043338474 + } + }, + "groups": { + "arxiv_mc": { + "acc,none": 0.16583333333333333, + "acc_stderr,none": 0.01074491240245708, + "acc_norm,none": 0.23916666666666667, + "acc_norm_stderr,none": 0.012288331311441483 + } + }, + "group_subtasks": { + "arxiv_mc": [ + "arxiv_mc_2407", + "arxiv_mc_2408", + "arxiv_mc_2409", + "arxiv_mc_2410", + "arxiv_mc_2411", + "arxiv_mc_2412", + "arxiv_mc_2501", + "arxiv_mc_2502", + "arxiv_mc_2503", + "arxiv_mc_2504", + "arxiv_mc_2505", + "arxiv_mc_2506" + ] + }, + "configs": { + "arxiv_mc_2407": { + "task": "arxiv_mc_2407", + "task_alias": "2024-07", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2407", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2407mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2408": { + "task": "arxiv_mc_2408", + "task_alias": "2024-08", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2408", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2408mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2409": { + "task": "arxiv_mc_2409", + "task_alias": "2024-09", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2409", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2409mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2410": { + "task": "arxiv_mc_2410", + "task_alias": "2024-10", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2410", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2410mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2411": { + "task": "arxiv_mc_2411", + "task_alias": "2024-11", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2411", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2411mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2412": { + "task": "arxiv_mc_2412", + "task_alias": "2024-12", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2412", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2412mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2501": { + "task": "arxiv_mc_2501", + "task_alias": "2025-01", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2501", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2501mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2502": { + "task": "arxiv_mc_2502", + "task_alias": "2025-02", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2502", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2502mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2503": { + "task": "arxiv_mc_2503", + "task_alias": "2025-03", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2503", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2503mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2504": { + "task": "arxiv_mc_2504", + "task_alias": "2025-04", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2504", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2504mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2505": { + "task": "arxiv_mc_2505", + "task_alias": "2025-05", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2505", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2505mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + }, + "arxiv_mc_2506": { + "task": "arxiv_mc_2506", + "task_alias": "2025-06", + "tag": "arxiv_mc_tasks", + "dataset_path": "json", + "dataset_name": "arxiv_mc_2506", + "dataset_kwargs": { + "data_files": { + "test": "/mnt/data/lm-evaluation-harness/dataset/arxiv_mc/2506mc.json" + } + }, + "test_split": "test", + "process_docs": "def process_docs(dataset: datasets.Dataset) -> datasets.Dataset:\n def _process_doc(doc):\n answer_map = {\"A\": 0, \"B\": 1, \"C\": 2, \"D\": 3}\n out_doc = {\n \"question\": doc[\"question\"],\n \"choices\": doc[\"choices\"],\n \"answer\": answer_map[doc[\"answer\"]],\n }\n return out_doc\n\n return dataset.map(_process_doc)\n", + "doc_to_text": "Question:{{question.strip()}}\nAnswer:", + "doc_to_target": "answer", + "unsafe_code": false, + "doc_to_choice": "{{choices}}", + "description": "", + "target_delimiter": " ", + "fewshot_delimiter": "\n\n", + "num_fewshot": 0, + "metric_list": [ + { + "metric": "acc", + "aggregation": "mean", + "higher_is_better": true + }, + { + "metric": "acc_norm", + "aggregation": "mean", + "higher_is_better": true + } + ], + "output_type": "multiple_choice", + "repeats": 1, + "should_decontaminate": false, + "metadata": {} + } + }, + "versions": { + "arxiv_mc": null, + "arxiv_mc_2407": "Yaml", + "arxiv_mc_2408": "Yaml", + "arxiv_mc_2409": "Yaml", + "arxiv_mc_2410": "Yaml", + "arxiv_mc_2411": "Yaml", + "arxiv_mc_2412": "Yaml", + "arxiv_mc_2501": "Yaml", + "arxiv_mc_2502": "Yaml", + "arxiv_mc_2503": "Yaml", + "arxiv_mc_2504": "Yaml", + "arxiv_mc_2505": "Yaml", + "arxiv_mc_2506": "Yaml" + }, + "n-shot": { + "arxiv_mc_2407": 0, + "arxiv_mc_2408": 0, + "arxiv_mc_2409": 0, + "arxiv_mc_2410": 0, + "arxiv_mc_2411": 0, + "arxiv_mc_2412": 0, + "arxiv_mc_2501": 0, + "arxiv_mc_2502": 0, + "arxiv_mc_2503": 0, + "arxiv_mc_2504": 0, + "arxiv_mc_2505": 0, + "arxiv_mc_2506": 0 + }, + "higher_is_better": { + "arxiv_mc": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2407": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2408": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2409": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2410": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2411": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2412": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2501": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2502": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2503": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2504": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2505": { + "acc": true, + "acc_norm": true + }, + "arxiv_mc_2506": { + "acc": true, + "acc_norm": true + } + }, + "n-samples": { + "arxiv_mc_2407": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2408": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2409": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2410": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2411": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2412": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2501": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2502": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2503": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2504": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2505": { + "original": 100, + "effective": 100 + }, + "arxiv_mc_2506": { + "original": 100, + "effective": 100 + } + }, + "samples": { + "arxiv_mc_2407": [ + { + "doc_id": 0, + "doc": { + "question": "Which chunking method was chosen in the study for balancing simplicity and semantic preservation?", + "choices": [ + "Token-level Chunking", + "Sentence-level Chunking", + "Semantic-level Chunking", + "Paragraph-level Chunking" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which chunking method was chosen in the study for balancing simplicity and semantic preservation?\nAnswer:", + " Token-level Chunking" + ], + [ + "Question:Which chunking method was chosen in the study for balancing simplicity and semantic preservation?\nAnswer:", + " Sentence-level Chunking" + ], + [ + "Question:Which chunking method was chosen in the study for balancing simplicity and semantic preservation?\nAnswer:", + " Semantic-level Chunking" + ], + [ + "Question:Which chunking method was chosen in the study for balancing simplicity and semantic preservation?\nAnswer:", + " Paragraph-level Chunking" + ] + ], + "resps": [ + [ + [ + -14.86745834350586, + false + ] + ], + [ + [ + -12.75033950805664, + false + ] + ], + [ + [ + -18.18337059020996, + false + ] + ], + [ + [ + -15.440104484558105, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.86745834350586, + false + ], + [ + -12.75033950805664, + false + ], + [ + -18.18337059020996, + false + ], + [ + -15.440104484558105, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "112fbf650834f38f39bc57c35a7c8cc57f706e1daacf1cdd6a832275c5608ac3", + "prompt_hash": "4f21c252401f8a7f83cba5cd45a593221ca2fa1ec916980f2b415e1ac1d60783", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 1, + "doc": { + "question": "Which vector database was identified as the most comprehensive solution, meeting all key criteria including multiple index types, billion-scale support, hybrid search, and cloud-native capabilities?", + "choices": [ + "Weaviate", + "Faiss", + "Qdrant", + "Milvus" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which vector database was identified as the most comprehensive solution, meeting all key criteria including multiple index types, billion-scale support, hybrid search, and cloud-native capabilities?\nAnswer:", + " Weaviate" + ], + [ + "Question:Which vector database was identified as the most comprehensive solution, meeting all key criteria including multiple index types, billion-scale support, hybrid search, and cloud-native capabilities?\nAnswer:", + " Faiss" + ], + [ + "Question:Which vector database was identified as the most comprehensive solution, meeting all key criteria including multiple index types, billion-scale support, hybrid search, and cloud-native capabilities?\nAnswer:", + " Qdrant" + ], + [ + "Question:Which vector database was identified as the most comprehensive solution, meeting all key criteria including multiple index types, billion-scale support, hybrid search, and cloud-native capabilities?\nAnswer:", + " Milvus" + ] + ], + "resps": [ + [ + [ + -16.53425407409668, + false + ] + ], + [ + [ + -7.181585788726807, + false + ] + ], + [ + [ + -9.658040046691895, + false + ] + ], + [ + [ + -8.326963424682617, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.53425407409668, + false + ], + [ + -7.181585788726807, + false + ], + [ + -9.658040046691895, + false + ], + [ + -8.326963424682617, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "dc5641478241a96b0f12c3e9c434abd3b0971c0ed76a108c5f8298504f86d732", + "prompt_hash": "aedd8317881ad284c723d89028876ce36faf93c1b4b32fc269961413c7729440", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 2, + "doc": { + "question": "Which reranking model was recommended as balancing performance and efficiency?", + "choices": [ + "monoT5", + "RankLLaMA", + "TILDEv2", + "monoBERT" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which reranking model was recommended as balancing performance and efficiency?\nAnswer:", + " monoT5" + ], + [ + "Question:Which reranking model was recommended as balancing performance and efficiency?\nAnswer:", + " RankLLaMA" + ], + [ + "Question:Which reranking model was recommended as balancing performance and efficiency?\nAnswer:", + " TILDEv2" + ], + [ + "Question:Which reranking model was recommended as balancing performance and efficiency?\nAnswer:", + " monoBERT" + ] + ], + "resps": [ + [ + [ + -15.6542329788208, + false + ] + ], + [ + [ + -12.929773330688477, + false + ] + ], + [ + [ + -19.209312438964844, + false + ] + ], + [ + [ + -14.345467567443848, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.6542329788208, + false + ], + [ + -12.929773330688477, + false + ], + [ + -19.209312438964844, + false + ], + [ + -14.345467567443848, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ec1881d1f0dd4d7a02eb312674cd46a64e082bcc50727d3d07966e690bcaaee0", + "prompt_hash": "1516fe9f99e098b69b5a14277865693e334db21041e6a088a001c7daf9ed2959", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 3, + "doc": { + "question": "In generator fine-tuning, which training strategy was found to enhance robustness by mixing relevant and random contexts?", + "choices": [ + "Training only with relevant documents (Dg)", + "Training only with random documents (Dr)", + "Training with both relevant and random documents (Dgr)", + "Training with duplicate relevant documents (Dgg)" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In generator fine-tuning, which training strategy was found to enhance robustness by mixing relevant and random contexts?\nAnswer:", + " Training only with relevant documents (Dg)" + ], + [ + "Question:In generator fine-tuning, which training strategy was found to enhance robustness by mixing relevant and random contexts?\nAnswer:", + " Training only with random documents (Dr)" + ], + [ + "Question:In generator fine-tuning, which training strategy was found to enhance robustness by mixing relevant and random contexts?\nAnswer:", + " Training with both relevant and random documents (Dgr)" + ], + [ + "Question:In generator fine-tuning, which training strategy was found to enhance robustness by mixing relevant and random contexts?\nAnswer:", + " Training with duplicate relevant documents (Dgg)" + ] + ], + "resps": [ + [ + [ + -40.50318908691406, + false + ] + ], + [ + [ + -39.652557373046875, + false + ] + ], + [ + [ + -44.193763732910156, + false + ] + ], + [ + [ + -52.69440460205078, + false + ] + ] + ], + "filtered_resps": [ + [ + -40.50318908691406, + false + ], + [ + -39.652557373046875, + false + ], + [ + -44.193763732910156, + false + ], + [ + -52.69440460205078, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7173c9ca783290cdb4b8a10eb3a1dce044a749f837da2139a7a29d92d000cf85", + "prompt_hash": "384c9b52ef5231fd417565f14aea7a6a784241b6b1fb250deee33dbddf87fff7", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 4, + "doc": { + "question": "Which summarization method was recommended for its outstanding performance?", + "choices": [ + "Selective Context", + "LongLLMLingua", + "BM25 Extractive", + "Recomp" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which summarization method was recommended for its outstanding performance?\nAnswer:", + " Selective Context" + ], + [ + "Question:Which summarization method was recommended for its outstanding performance?\nAnswer:", + " LongLLMLingua" + ], + [ + "Question:Which summarization method was recommended for its outstanding performance?\nAnswer:", + " BM25 Extractive" + ], + [ + "Question:Which summarization method was recommended for its outstanding performance?\nAnswer:", + " Recomp" + ] + ], + "resps": [ + [ + [ + -20.990047454833984, + false + ] + ], + [ + [ + -18.77847671508789, + false + ] + ], + [ + [ + -23.54022216796875, + false + ] + ], + [ + [ + -14.204093933105469, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.990047454833984, + false + ], + [ + -18.77847671508789, + false + ], + [ + -23.54022216796875, + false + ], + [ + -14.204093933105469, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a6b581adefddccee39d5247c99e1f0bde00414c0f883a8cf2547bdb539a9aa7b", + "prompt_hash": "9fa3215a12dc1fc19dc1c9d7fd30cd52340f16e534d409f05ea2262435ed51a1", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 5, + "doc": { + "question": "How many visual math problems are included in the proposed dataset?", + "choices": [ + "4.5K", + "5.0K", + "6.5K", + "7.2K" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many visual math problems are included in the proposed dataset?\nAnswer:", + " 4.5K" + ], + [ + "Question:How many visual math problems are included in the proposed dataset?\nAnswer:", + " 5.0K" + ], + [ + "Question:How many visual math problems are included in the proposed dataset?\nAnswer:", + " 6.5K" + ], + [ + "Question:How many visual math problems are included in the proposed dataset?\nAnswer:", + " 7.2K" + ] + ], + "resps": [ + [ + [ + -12.729255676269531, + false + ] + ], + [ + [ + -14.22647476196289, + false + ] + ], + [ + [ + -13.167308807373047, + false + ] + ], + [ + [ + -13.763832092285156, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.729255676269531, + false + ], + [ + -14.22647476196289, + false + ], + [ + -13.167308807373047, + false + ], + [ + -13.763832092285156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fb3604d749f5b7fcc83d359bdf0a18094bf32436d7242c0d61f39bc92a5575ec", + "prompt_hash": "e7b62eec87e03ee8506c86b240c326f3114a1ab0247407850fc920b4c2c2332b", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 6, + "doc": { + "question": "Into how many knowledge concepts are the problems in the dataset categorized?", + "choices": [ + "45", + "57", + "67", + "77" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Into how many knowledge concepts are the problems in the dataset categorized?\nAnswer:", + " 45" + ], + [ + "Question:Into how many knowledge concepts are the problems in the dataset categorized?\nAnswer:", + " 57" + ], + [ + "Question:Into how many knowledge concepts are the problems in the dataset categorized?\nAnswer:", + " 67" + ], + [ + "Question:Into how many knowledge concepts are the problems in the dataset categorized?\nAnswer:", + " 77" + ] + ], + "resps": [ + [ + [ + -8.812484741210938, + false + ] + ], + [ + [ + -9.466009140014648, + false + ] + ], + [ + [ + -9.849740982055664, + false + ] + ], + [ + [ + -10.170186996459961, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.812484741210938, + false + ], + [ + -9.466009140014648, + false + ], + [ + -9.849740982055664, + false + ], + [ + -10.170186996459961, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9dbe00d7ad342d9d61a05b95909c8335e44a65600dbca6821fc9ff31d702aa79", + "prompt_hash": "b8c118c69992d28a3628ec73ea1ac12a01da856f97eef3b95a3339bb4e94a8fb", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 7, + "doc": { + "question": "Which four-dimensional metric is used to evaluate problem-solving abilities?", + "choices": [ + "Precision, Recall, Accuracy, F1", + "IK, IG, CM, RM", + "BLEU, ROUGE, METEOR, CIDEr", + "Speed, Memory, Efficiency, Robustness" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which four-dimensional metric is used to evaluate problem-solving abilities?\nAnswer:", + " Precision, Recall, Accuracy, F1" + ], + [ + "Question:Which four-dimensional metric is used to evaluate problem-solving abilities?\nAnswer:", + " IK, IG, CM, RM" + ], + [ + "Question:Which four-dimensional metric is used to evaluate problem-solving abilities?\nAnswer:", + " BLEU, ROUGE, METEOR, CIDEr" + ], + [ + "Question:Which four-dimensional metric is used to evaluate problem-solving abilities?\nAnswer:", + " Speed, Memory, Efficiency, Robustness" + ] + ], + "resps": [ + [ + [ + -24.218046188354492, + false + ] + ], + [ + [ + -40.70429611206055, + false + ] + ], + [ + [ + -28.812503814697266, + false + ] + ], + [ + [ + -38.256866455078125, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.218046188354492, + false + ], + [ + -40.70429611206055, + false + ], + [ + -28.812503814697266, + false + ], + [ + -38.256866455078125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e7d482ff7ce85c9bb9f81ef47fda540bfc235fd3cdbb8cce2b61775e27c03d22", + "prompt_hash": "f32689e2b31b6aa425d21f4e02e4b2b1a6bfbbb6c61cbb7ddf7f0895b98e939d", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 8, + "doc": { + "question": "Which model achieved the best overall performance across different visual mathematics categories?", + "choices": [ + "GPT-4V", + "Gemini 1.5 Pro", + "GPT-4o", + "LLaVA-NeXT-110B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model achieved the best overall performance across different visual mathematics categories?\nAnswer:", + " GPT-4V" + ], + [ + "Question:Which model achieved the best overall performance across different visual mathematics categories?\nAnswer:", + " Gemini 1.5 Pro" + ], + [ + "Question:Which model achieved the best overall performance across different visual mathematics categories?\nAnswer:", + " GPT-4o" + ], + [ + "Question:Which model achieved the best overall performance across different visual mathematics categories?\nAnswer:", + " LLaVA-NeXT-110B" + ] + ], + "resps": [ + [ + [ + -9.522256851196289, + false + ] + ], + [ + [ + -8.766477584838867, + false + ] + ], + [ + [ + -7.747229099273682, + false + ] + ], + [ + [ + -21.87334442138672, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.522256851196289, + false + ], + [ + -8.766477584838867, + false + ], + [ + -7.747229099273682, + false + ], + [ + -21.87334442138672, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8c452c6f79e5150462edd67e4195af4bd8b30a326a1993e322675fd8487a60fe", + "prompt_hash": "6c857a8e695bf66745f3a947b038b9297beda2f70ee0d8d18dbdf51c67e38e55", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 9, + "doc": { + "question": "From where were the 67 knowledge concept descriptions for Knowledge Concept Augmentation (KCA) constructed?", + "choices": [ + "Only Wikipedia", + "Only Euclid's Elements", + "Wikipedia and textbooks", + "Research papers only" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:From where were the 67 knowledge concept descriptions for Knowledge Concept Augmentation (KCA) constructed?\nAnswer:", + " Only Wikipedia" + ], + [ + "Question:From where were the 67 knowledge concept descriptions for Knowledge Concept Augmentation (KCA) constructed?\nAnswer:", + " Only Euclid's Elements" + ], + [ + "Question:From where were the 67 knowledge concept descriptions for Knowledge Concept Augmentation (KCA) constructed?\nAnswer:", + " Wikipedia and textbooks" + ], + [ + "Question:From where were the 67 knowledge concept descriptions for Knowledge Concept Augmentation (KCA) constructed?\nAnswer:", + " Research papers only" + ] + ], + "resps": [ + [ + [ + -14.755115509033203, + false + ] + ], + [ + [ + -23.898681640625, + false + ] + ], + [ + [ + -15.336736679077148, + false + ] + ], + [ + [ + -14.167567253112793, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.755115509033203, + false + ], + [ + -23.898681640625, + false + ], + [ + -15.336736679077148, + false + ], + [ + -14.167567253112793, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "aac932a6dc506406fdc0373f66d9e474f5ea83351d26ca76893eba8dacb3de96", + "prompt_hash": "c41ce6c6bd641fdcf1c75204b06887e517a2df335ace0c502170795e268a6228", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 10, + "doc": { + "question": "Which humanoid robot used in the experiments is equipped with multi-finger hands?", + "choices": [ + "Fourier GR-1", + "Unitree H1", + "ALOHA", + "AnyTeleop" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which humanoid robot used in the experiments is equipped with multi-finger hands?\nAnswer:", + " Fourier GR-1" + ], + [ + "Question:Which humanoid robot used in the experiments is equipped with multi-finger hands?\nAnswer:", + " Unitree H1" + ], + [ + "Question:Which humanoid robot used in the experiments is equipped with multi-finger hands?\nAnswer:", + " ALOHA" + ], + [ + "Question:Which humanoid robot used in the experiments is equipped with multi-finger hands?\nAnswer:", + " AnyTeleop" + ] + ], + "resps": [ + [ + [ + -19.240524291992188, + false + ] + ], + [ + [ + -8.106757164001465, + false + ] + ], + [ + [ + -14.45346450805664, + false + ] + ], + [ + [ + -17.957672119140625, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.240524291992188, + false + ], + [ + -8.106757164001465, + false + ], + [ + -14.45346450805664, + false + ], + [ + -17.957672119140625, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "03549faa0210f17c230f5e924730bcd50c2c8b8d86e30dd769837a95380f5495", + "prompt_hash": "513ea08ab58365f9c8d0dbe8082c93d5d5beaa6661c0cee8334ebe899b02c226", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 11, + "doc": { + "question": "What stereo camera is used with both robots to provide RGB streaming?", + "choices": [ + "Intel RealSense D435", + "ZED Mini", + "Azure Kinect", + "Leap Motion" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What stereo camera is used with both robots to provide RGB streaming?\nAnswer:", + " Intel RealSense D435" + ], + [ + "Question:What stereo camera is used with both robots to provide RGB streaming?\nAnswer:", + " ZED Mini" + ], + [ + "Question:What stereo camera is used with both robots to provide RGB streaming?\nAnswer:", + " Azure Kinect" + ], + [ + "Question:What stereo camera is used with both robots to provide RGB streaming?\nAnswer:", + " Leap Motion" + ] + ], + "resps": [ + [ + [ + -7.248941898345947, + false + ] + ], + [ + [ + -8.554323196411133, + false + ] + ], + [ + [ + -10.395401000976562, + false + ] + ], + [ + [ + -17.37678337097168, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.248941898345947, + false + ], + [ + -8.554323196411133, + false + ], + [ + -10.395401000976562, + false + ], + [ + -17.37678337097168, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "097bac5cbc7f405d0d2c5d54bb87a36fbc0b869e5ca0fb38460550e54b3c1deb", + "prompt_hash": "6e647c4b2048258098420c8acb5aed5a05834763fdf430aab1b860c12a68b87b", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 12, + "doc": { + "question": "In the Can Insertion task, what is the approximate diameter of the soda can slot?", + "choices": [ + "3.6 cm", + "5.6 cm", + "7.6 cm", + "9.6 cm" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In the Can Insertion task, what is the approximate diameter of the soda can slot?\nAnswer:", + " 3.6 cm" + ], + [ + "Question:In the Can Insertion task, what is the approximate diameter of the soda can slot?\nAnswer:", + " 5.6 cm" + ], + [ + "Question:In the Can Insertion task, what is the approximate diameter of the soda can slot?\nAnswer:", + " 7.6 cm" + ], + [ + "Question:In the Can Insertion task, what is the approximate diameter of the soda can slot?\nAnswer:", + " 9.6 cm" + ] + ], + "resps": [ + [ + [ + -9.950165748596191, + false + ] + ], + [ + [ + -10.258798599243164, + false + ] + ], + [ + [ + -10.89891242980957, + false + ] + ], + [ + [ + -11.146050453186035, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.950165748596191, + false + ], + [ + -10.258798599243164, + false + ], + [ + -10.89891242980957, + false + ], + [ + -11.146050453186035, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9a1a06df62b817e903f1359d4c025997f64cf35d891deaac93377a160228984a", + "prompt_hash": "6059a3748f485ec2ce37de889c2df78a6f8f5e469d1c7156c1c872b8278b5edd", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 13, + "doc": { + "question": "Which visual backbone replaced ResNet in the modified ACT algorithm?", + "choices": [ + "ViT-B/16", + "DinoV2", + "CLIP-ViT", + "Swin Transformer" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which visual backbone replaced ResNet in the modified ACT algorithm?\nAnswer:", + " ViT-B/16" + ], + [ + "Question:Which visual backbone replaced ResNet in the modified ACT algorithm?\nAnswer:", + " DinoV2" + ], + [ + "Question:Which visual backbone replaced ResNet in the modified ACT algorithm?\nAnswer:", + " CLIP-ViT" + ], + [ + "Question:Which visual backbone replaced ResNet in the modified ACT algorithm?\nAnswer:", + " Swin Transformer" + ] + ], + "resps": [ + [ + [ + -10.351160049438477, + false + ] + ], + [ + [ + -9.934405326843262, + false + ] + ], + [ + [ + -12.724019050598145, + false + ] + ], + [ + [ + -9.91457748413086, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.351160049438477, + false + ], + [ + -9.934405326843262, + false + ], + [ + -12.724019050598145, + false + ], + [ + -9.91457748413086, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7335d1d88d4bdc8ea82abb121524726484026f6508ee853f1e27fb73070421c3", + "prompt_hash": "55413e378ac34f5ac9b87eb75e3a8375c6d12a88b184f3919918f9cbeb5ffd03", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 14, + "doc": { + "question": "What is the action dimension of the H1 robot in the imitation learning experiments?", + "choices": [ + "19", + "22", + "28", + "36" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the action dimension of the H1 robot in the imitation learning experiments?\nAnswer:", + " 19" + ], + [ + "Question:What is the action dimension of the H1 robot in the imitation learning experiments?\nAnswer:", + " 22" + ], + [ + "Question:What is the action dimension of the H1 robot in the imitation learning experiments?\nAnswer:", + " 28" + ], + [ + "Question:What is the action dimension of the H1 robot in the imitation learning experiments?\nAnswer:", + " 36" + ] + ], + "resps": [ + [ + [ + -7.213255882263184, + false + ] + ], + [ + [ + -7.448844909667969, + false + ] + ], + [ + [ + -7.832391738891602, + false + ] + ], + [ + [ + -7.059266567230225, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.213255882263184, + false + ], + [ + -7.448844909667969, + false + ], + [ + -7.832391738891602, + false + ], + [ + -7.059266567230225, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "814c5e04b3c52aee4763235c2a6be2412ce07724c00036756ee1a253a02d9677", + "prompt_hash": "afce1cbfccf449d047baf9ab43d4dfee10be79063fe8399c469b1df801b2e8a9", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 15, + "doc": { + "question": "What is the primary goal of Kimi as a Model-as-a-Service (MaaS) provider?", + "choices": [ + "To maximize overall effective throughput while meeting latency-related SLO requirements", + "To minimize GPU usage regardless of service quality", + "To reject as many requests as possible to save resources", + "To increase output length of all requests without constraints" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the primary goal of Kimi as a Model-as-a-Service (MaaS) provider?\nAnswer:", + " To maximize overall effective throughput while meeting latency-related SLO requirements" + ], + [ + "Question:What is the primary goal of Kimi as a Model-as-a-Service (MaaS) provider?\nAnswer:", + " To minimize GPU usage regardless of service quality" + ], + [ + "Question:What is the primary goal of Kimi as a Model-as-a-Service (MaaS) provider?\nAnswer:", + " To reject as many requests as possible to save resources" + ], + [ + "Question:What is the primary goal of Kimi as a Model-as-a-Service (MaaS) provider?\nAnswer:", + " To increase output length of all requests without constraints" + ] + ], + "resps": [ + [ + [ + -46.09083938598633, + false + ] + ], + [ + [ + -44.75458526611328, + false + ] + ], + [ + [ + -32.21599197387695, + false + ] + ], + [ + [ + -50.386390686035156, + false + ] + ] + ], + "filtered_resps": [ + [ + -46.09083938598633, + false + ], + [ + -44.75458526611328, + false + ], + [ + -32.21599197387695, + false + ], + [ + -50.386390686035156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "357101440addaf6419788f163efe79352930f15a794edcda0260253127f3980e", + "prompt_hash": "61cb0bfc2625337ab48a23c909c4203ddc219f9d2ade56062d4f2d2ddc02cbc6", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 16, + "doc": { + "question": "In Mooncake's architecture, what does the prefill stage primarily focus on?", + "choices": [ + "Generating all output tokens at once", + "Processing all input tokens in parallel and generating the first output token", + "Rejecting requests before execution", + "Managing memory transfers between CPU and GPU" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:In Mooncake's architecture, what does the prefill stage primarily focus on?\nAnswer:", + " Generating all output tokens at once" + ], + [ + "Question:In Mooncake's architecture, what does the prefill stage primarily focus on?\nAnswer:", + " Processing all input tokens in parallel and generating the first output token" + ], + [ + "Question:In Mooncake's architecture, what does the prefill stage primarily focus on?\nAnswer:", + " Rejecting requests before execution" + ], + [ + "Question:In Mooncake's architecture, what does the prefill stage primarily focus on?\nAnswer:", + " Managing memory transfers between CPU and GPU" + ] + ], + "resps": [ + [ + [ + -23.025474548339844, + false + ] + ], + [ + [ + -34.40630340576172, + false + ] + ], + [ + [ + -33.791385650634766, + false + ] + ], + [ + [ + -24.586078643798828, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.025474548339844, + false + ], + [ + -34.40630340576172, + false + ], + [ + -33.791385650634766, + false + ], + [ + -24.586078643798828, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "38ca07f4adc667f8133e57528e618095826332c29aa1c3319cdae66a99c6378a", + "prompt_hash": "18182cffc75f3b11fb711616fe233f9e643833bc77701108f1a5e3995d185645", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 17, + "doc": { + "question": "What is the main advantage of Early Rejection in Mooncake's overload scheduling?", + "choices": [ + "It increases the output length of requests", + "It reduces ineffective computations by rejecting requests before wasting prefill resources", + "It doubles the number of decoding nodes required", + "It guarantees zero fluctuations in system load" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main advantage of Early Rejection in Mooncake's overload scheduling?\nAnswer:", + " It increases the output length of requests" + ], + [ + "Question:What is the main advantage of Early Rejection in Mooncake's overload scheduling?\nAnswer:", + " It reduces ineffective computations by rejecting requests before wasting prefill resources" + ], + [ + "Question:What is the main advantage of Early Rejection in Mooncake's overload scheduling?\nAnswer:", + " It doubles the number of decoding nodes required" + ], + [ + "Question:What is the main advantage of Early Rejection in Mooncake's overload scheduling?\nAnswer:", + " It guarantees zero fluctuations in system load" + ] + ], + "resps": [ + [ + [ + -27.2811222076416, + false + ] + ], + [ + [ + -61.370513916015625, + false + ] + ], + [ + [ + -38.169864654541016, + false + ] + ], + [ + [ + -31.666122436523438, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.2811222076416, + false + ], + [ + -61.370513916015625, + false + ], + [ + -38.169864654541016, + false + ], + [ + -31.666122436523438, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "34b971e8ea67c807f3df941c16dce19d1c799012d68e6bc92335d7c2803a550e", + "prompt_hash": "1a16f6563d0113375ee1a225e7fa2fde0dd9fbeb6b356447607f396f66e4d5fd", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 18, + "doc": { + "question": "Which cache eviction policy achieved the best performance under Mooncake's request trace dataset?", + "choices": [ + "FIFO (First In First Out)", + "LRU (Least Recently Used)", + "LFU (Least Frequently Used)", + "LengthAwareCache" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which cache eviction policy achieved the best performance under Mooncake's request trace dataset?\nAnswer:", + " FIFO (First In First Out)" + ], + [ + "Question:Which cache eviction policy achieved the best performance under Mooncake's request trace dataset?\nAnswer:", + " LRU (Least Recently Used)" + ], + [ + "Question:Which cache eviction policy achieved the best performance under Mooncake's request trace dataset?\nAnswer:", + " LFU (Least Frequently Used)" + ], + [ + "Question:Which cache eviction policy achieved the best performance under Mooncake's request trace dataset?\nAnswer:", + " LengthAwareCache" + ] + ], + "resps": [ + [ + [ + -12.678746223449707, + false + ] + ], + [ + [ + -10.518775939941406, + false + ] + ], + [ + [ + -17.02437973022461, + false + ] + ], + [ + [ + -18.03860092163086, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.678746223449707, + false + ], + [ + -10.518775939941406, + false + ], + [ + -17.02437973022461, + false + ], + [ + -18.03860092163086, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fc4857a67b3c44de917f0962153f5561de3784f4837615bf36cf05223ce42f3e", + "prompt_hash": "fbd26fbc05e64de5a13e17eb8ddaac22222e0f3ecd4d93f15822c5194dc0f8eb", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 19, + "doc": { + "question": "In the evaluation of Mooncake versus vLLM on simulated long-context data, what was Mooncake's throughput improvement range?", + "choices": [ + "10% to 25%", + "30% to 75%", + "50% to 525%", + "600% to 1000%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In the evaluation of Mooncake versus vLLM on simulated long-context data, what was Mooncake's throughput improvement range?\nAnswer:", + " 10% to 25%" + ], + [ + "Question:In the evaluation of Mooncake versus vLLM on simulated long-context data, what was Mooncake's throughput improvement range?\nAnswer:", + " 30% to 75%" + ], + [ + "Question:In the evaluation of Mooncake versus vLLM on simulated long-context data, what was Mooncake's throughput improvement range?\nAnswer:", + " 50% to 525%" + ], + [ + "Question:In the evaluation of Mooncake versus vLLM on simulated long-context data, what was Mooncake's throughput improvement range?\nAnswer:", + " 600% to 1000%" + ] + ], + "resps": [ + [ + [ + -20.490446090698242, + false + ] + ], + [ + [ + -21.809951782226562, + false + ] + ], + [ + [ + -28.085920333862305, + false + ] + ], + [ + [ + -22.367385864257812, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.490446090698242, + false + ], + [ + -21.809951782226562, + false + ], + [ + -28.085920333862305, + false + ], + [ + -22.367385864257812, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d29b4d3b3018678369da35496faf2781c1cd2413e158c32855218e1cc61af667", + "prompt_hash": "996d24f8f92ece7fb356f59ef65a55f3015130f5e67c6565b61d20902e9c837c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 20, + "doc": { + "question": "What is the primary purpose of the FEVER dataset?", + "choices": [ + "To verify factual claims using evidence from Wikipedia", + "To evaluate conversational abilities of AI assistants", + "To generate multi-hop reasoning chains", + "To benchmark cross-lingual retrieval tasks" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the primary purpose of the FEVER dataset?\nAnswer:", + " To verify factual claims using evidence from Wikipedia" + ], + [ + "Question:What is the primary purpose of the FEVER dataset?\nAnswer:", + " To evaluate conversational abilities of AI assistants" + ], + [ + "Question:What is the primary purpose of the FEVER dataset?\nAnswer:", + " To generate multi-hop reasoning chains" + ], + [ + "Question:What is the primary purpose of the FEVER dataset?\nAnswer:", + " To benchmark cross-lingual retrieval tasks" + ] + ], + "resps": [ + [ + [ + -27.64024543762207, + false + ] + ], + [ + [ + -29.950502395629883, + false + ] + ], + [ + [ + -21.078857421875, + false + ] + ], + [ + [ + -24.49301528930664, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.64024543762207, + false + ], + [ + -29.950502395629883, + false + ], + [ + -21.078857421875, + false + ], + [ + -24.49301528930664, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "cbe34ffff85183cd109608a0dcc923042e88190c33685761b9a3196d82235d94", + "prompt_hash": "7711b67fd53913f5056986c72fa2a4002ba3a489c2518ce1ec791fcccf202504", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 21, + "doc": { + "question": "How many questions are included in the MMLU-med dataset?", + "choices": [ + "1,273", + "4,183", + "1,089", + "1,000" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many questions are included in the MMLU-med dataset?\nAnswer:", + " 1,273" + ], + [ + "Question:How many questions are included in the MMLU-med dataset?\nAnswer:", + " 4,183" + ], + [ + "Question:How many questions are included in the MMLU-med dataset?\nAnswer:", + " 1,089" + ], + [ + "Question:How many questions are included in the MMLU-med dataset?\nAnswer:", + " 1,000" + ] + ], + "resps": [ + [ + [ + -13.105728149414062, + false + ] + ], + [ + [ + -14.59390926361084, + false + ] + ], + [ + [ + -12.9570894241333, + false + ] + ], + [ + [ + -8.63340950012207, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.105728149414062, + false + ], + [ + -14.59390926361084, + false + ], + [ + -12.9570894241333, + false + ], + [ + -8.63340950012207, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1c2a3feddd0a24b3633fc2cfc1e6137d92c681e6bcb5ac792da9785f22329b9d", + "prompt_hash": "2849090bd6bbb607b5e19bb26891b41e60931e3fb66a3eb6384318e73d63ef7e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 22, + "doc": { + "question": "What is the main characteristic of the PopQA dataset?", + "choices": [ + "It focuses on cross-lingual question answering", + "It targets long-tail entities with low Wikipedia page views", + "It verifies biomedical hypotheses using abstracts", + "It tests multi-hop reasoning across documents" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main characteristic of the PopQA dataset?\nAnswer:", + " It focuses on cross-lingual question answering" + ], + [ + "Question:What is the main characteristic of the PopQA dataset?\nAnswer:", + " It targets long-tail entities with low Wikipedia page views" + ], + [ + "Question:What is the main characteristic of the PopQA dataset?\nAnswer:", + " It verifies biomedical hypotheses using abstracts" + ], + [ + "Question:What is the main characteristic of the PopQA dataset?\nAnswer:", + " It tests multi-hop reasoning across documents" + ] + ], + "resps": [ + [ + [ + -21.004573822021484, + false + ] + ], + [ + [ + -43.161983489990234, + false + ] + ], + [ + [ + -46.05027770996094, + false + ] + ], + [ + [ + -30.60395050048828, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.004573822021484, + false + ], + [ + -43.161983489990234, + false + ], + [ + -46.05027770996094, + false + ], + [ + -30.60395050048828, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "402c31403bb0f6373c60bdbb8ae1740ddc46d18ede9eeb6c93717377641878ba", + "prompt_hash": "a55011ae2acadc35c2994a6c5e1267c86e1a7cd3c7062ac6d9d09d7d06d4ea21", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 23, + "doc": { + "question": "Which dataset comprises four-option multiple-choice questions derived from the US Medical Licensing Examination?", + "choices": [ + "MedQA", + "MedMCQA", + "PubMedQA", + "BioASQ" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which dataset comprises four-option multiple-choice questions derived from the US Medical Licensing Examination?\nAnswer:", + " MedQA" + ], + [ + "Question:Which dataset comprises four-option multiple-choice questions derived from the US Medical Licensing Examination?\nAnswer:", + " MedMCQA" + ], + [ + "Question:Which dataset comprises four-option multiple-choice questions derived from the US Medical Licensing Examination?\nAnswer:", + " PubMedQA" + ], + [ + "Question:Which dataset comprises four-option multiple-choice questions derived from the US Medical Licensing Examination?\nAnswer:", + " BioASQ" + ] + ], + "resps": [ + [ + [ + -12.405452728271484, + false + ] + ], + [ + [ + -13.451678276062012, + false + ] + ], + [ + [ + -11.517139434814453, + false + ] + ], + [ + [ + -15.579611778259277, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.405452728271484, + false + ], + [ + -13.451678276062012, + false + ], + [ + -11.517139434814453, + false + ], + [ + -15.579611778259277, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ff39eacbadfde0220c466c7c4682e2f8dea522c56595934cfd1e44f27d98ab7e", + "prompt_hash": "b326f1ed4d75097c5ee290924fd998b1cd403d4c50c0f32066da74d9b18bfab0", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 24, + "doc": { + "question": "Which dataset consists of 4,183 questions from Indian medical entrance exams?", + "choices": [ + "MedQA", + "MedMCQA", + "MMLU-med", + "BioASQ" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset consists of 4,183 questions from Indian medical entrance exams?\nAnswer:", + " MedQA" + ], + [ + "Question:Which dataset consists of 4,183 questions from Indian medical entrance exams?\nAnswer:", + " MedMCQA" + ], + [ + "Question:Which dataset consists of 4,183 questions from Indian medical entrance exams?\nAnswer:", + " MMLU-med" + ], + [ + "Question:Which dataset consists of 4,183 questions from Indian medical entrance exams?\nAnswer:", + " BioASQ" + ] + ], + "resps": [ + [ + [ + -10.04144287109375, + false + ] + ], + [ + [ + -11.477560997009277, + false + ] + ], + [ + [ + -17.35746955871582, + false + ] + ], + [ + [ + -13.449769973754883, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.04144287109375, + false + ], + [ + -11.477560997009277, + false + ], + [ + -17.35746955871582, + false + ], + [ + -13.449769973754883, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7d73c5a4d61656c245d61a96508c21f0938457c79aadfb3371e93093c71724c4", + "prompt_hash": "4aa67d3abbe709aca8112d6fa1f1b6c58886840955ee28a445dc0bfff11d2f1c", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 25, + "doc": { + "question": "What is the maximum context window size supported by InternLM-XComposer-2.5 (IXC-2.5) through positional encoding extrapolation?", + "choices": [ + "32K", + "64K", + "96K", + "128K" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the maximum context window size supported by InternLM-XComposer-2.5 (IXC-2.5) through positional encoding extrapolation?\nAnswer:", + " 32K" + ], + [ + "Question:What is the maximum context window size supported by InternLM-XComposer-2.5 (IXC-2.5) through positional encoding extrapolation?\nAnswer:", + " 64K" + ], + [ + "Question:What is the maximum context window size supported by InternLM-XComposer-2.5 (IXC-2.5) through positional encoding extrapolation?\nAnswer:", + " 96K" + ], + [ + "Question:What is the maximum context window size supported by InternLM-XComposer-2.5 (IXC-2.5) through positional encoding extrapolation?\nAnswer:", + " 128K" + ] + ], + "resps": [ + [ + [ + -8.073480606079102, + false + ] + ], + [ + [ + -7.711792945861816, + false + ] + ], + [ + [ + -10.791585922241211, + false + ] + ], + [ + [ + -8.099313735961914, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.073480606079102, + false + ], + [ + -7.711792945861816, + false + ], + [ + -10.791585922241211, + false + ], + [ + -8.099313735961914, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0f2c2f84db9ec7c913471865cd0b8c41e8a1692ee8beaca699c03fc401f5a85d", + "prompt_hash": "3dfe147b1ecd03739fc0f37cd3ad9e7b56c59a19ca2448842d4d8fee8295cf70", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 26, + "doc": { + "question": "Which vision encoder resolution does IXC-2.5 employ to enhance ultra-high-resolution image understanding?", + "choices": [ + "448×448", + "490×490", + "512×512", + "560×560" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which vision encoder resolution does IXC-2.5 employ to enhance ultra-high-resolution image understanding?\nAnswer:", + " 448×448" + ], + [ + "Question:Which vision encoder resolution does IXC-2.5 employ to enhance ultra-high-resolution image understanding?\nAnswer:", + " 490×490" + ], + [ + "Question:Which vision encoder resolution does IXC-2.5 employ to enhance ultra-high-resolution image understanding?\nAnswer:", + " 512×512" + ], + [ + "Question:Which vision encoder resolution does IXC-2.5 employ to enhance ultra-high-resolution image understanding?\nAnswer:", + " 560×560" + ] + ], + "resps": [ + [ + [ + -10.27751350402832, + false + ] + ], + [ + [ + -17.577922821044922, + false + ] + ], + [ + [ + -10.534107208251953, + false + ] + ], + [ + [ + -14.884693145751953, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.27751350402832, + false + ], + [ + -17.577922821044922, + false + ], + [ + -10.534107208251953, + false + ], + [ + -14.884693145751953, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "48f11e12cf84425b320fe029b209d3d5bfe7b85c4a83f61df33fc5f241d880e4", + "prompt_hash": "1de30132ffb210ada2cef34d8a4a9f774dec474bff03da29088e916500d1200a", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 27, + "doc": { + "question": "On the MVBench benchmark, how much higher is IXC-2.5's score compared to VideoChat2-7B, the previous SOTA method?", + "choices": [ + "+5.3%", + "+8.7%", + "+12.4%", + "+15.1%" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:On the MVBench benchmark, how much higher is IXC-2.5's score compared to VideoChat2-7B, the previous SOTA method?\nAnswer:", + " +5.3%" + ], + [ + "Question:On the MVBench benchmark, how much higher is IXC-2.5's score compared to VideoChat2-7B, the previous SOTA method?\nAnswer:", + " +8.7%" + ], + [ + "Question:On the MVBench benchmark, how much higher is IXC-2.5's score compared to VideoChat2-7B, the previous SOTA method?\nAnswer:", + " +12.4%" + ], + [ + "Question:On the MVBench benchmark, how much higher is IXC-2.5's score compared to VideoChat2-7B, the previous SOTA method?\nAnswer:", + " +15.1%" + ] + ], + "resps": [ + [ + [ + -19.748226165771484, + false + ] + ], + [ + [ + -20.179574966430664, + false + ] + ], + [ + [ + -20.353487014770508, + false + ] + ], + [ + [ + -20.995635986328125, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.748226165771484, + false + ], + [ + -20.179574966430664, + false + ], + [ + -20.353487014770508, + false + ], + [ + -20.995635986328125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f84ae4d8ba1bd54bf160f301466be73ae54e58ea8ba4276dab4c2f131148aaca", + "prompt_hash": "d5e8d82f36af7daf364fd683f1b027a61bf29c34d889c1fd0a4112ed02cdc5be", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 28, + "doc": { + "question": "Which datasets were used in IXC-2.5's pre-training phase for general semantic alignment?", + "choices": [ + "ShareGPT4V-PT, COCO, Nocaps, TextCaps, LAION, SBU, CC 3M, ALLaVA", + "VQAv2, GQA, OK-VQA, ALLaVA-QA, MMDU", + "WebSight v0.1/v0.2, Stack v2, Tailwind CSS dataset", + "DVQA, ChartQA, CLEVR-MATH, DeepForm" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which datasets were used in IXC-2.5's pre-training phase for general semantic alignment?\nAnswer:", + " ShareGPT4V-PT, COCO, Nocaps, TextCaps, LAION, SBU, CC 3M, ALLaVA" + ], + [ + "Question:Which datasets were used in IXC-2.5's pre-training phase for general semantic alignment?\nAnswer:", + " VQAv2, GQA, OK-VQA, ALLaVA-QA, MMDU" + ], + [ + "Question:Which datasets were used in IXC-2.5's pre-training phase for general semantic alignment?\nAnswer:", + " WebSight v0.1/v0.2, Stack v2, Tailwind CSS dataset" + ], + [ + "Question:Which datasets were used in IXC-2.5's pre-training phase for general semantic alignment?\nAnswer:", + " DVQA, ChartQA, CLEVR-MATH, DeepForm" + ] + ], + "resps": [ + [ + [ + -93.22857666015625, + false + ] + ], + [ + [ + -47.38764953613281, + false + ] + ], + [ + [ + -66.16090393066406, + false + ] + ], + [ + [ + -52.975765228271484, + false + ] + ] + ], + "filtered_resps": [ + [ + -93.22857666015625, + false + ], + [ + -47.38764953613281, + false + ], + [ + -66.16090393066406, + false + ], + [ + -52.975765228271484, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "91e14a9eff8495d2aee9943461e779417887d1ecb4f0cc5754ae98a45ead996d", + "prompt_hash": "0fba08fe9bf2110d9ebd4867bd6f6281c78969c2efa62433145f3c25e8eaaff2", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 29, + "doc": { + "question": "What is the LoRA rank used in IXC-2.5's screenshot-to-code training phase?", + "choices": [ + "128", + "256", + "384", + "512" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the LoRA rank used in IXC-2.5's screenshot-to-code training phase?\nAnswer:", + " 128" + ], + [ + "Question:What is the LoRA rank used in IXC-2.5's screenshot-to-code training phase?\nAnswer:", + " 256" + ], + [ + "Question:What is the LoRA rank used in IXC-2.5's screenshot-to-code training phase?\nAnswer:", + " 384" + ], + [ + "Question:What is the LoRA rank used in IXC-2.5's screenshot-to-code training phase?\nAnswer:", + " 512" + ] + ], + "resps": [ + [ + [ + -4.809566497802734, + false + ] + ], + [ + [ + -5.269222259521484, + false + ] + ], + [ + [ + -8.795157432556152, + false + ] + ], + [ + [ + -5.778667449951172, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.809566497802734, + false + ], + [ + -5.269222259521484, + false + ], + [ + -8.795157432556152, + false + ], + [ + -5.778667449951172, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2b02af2e42eeeec4a8bfe1ebed1524093c46feab1666cf52aa334ed26fe24d71", + "prompt_hash": "d92d5eeefe98aef00d3cf8cf5a50a1e9838edc655619525802339940a0371418", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 30, + "doc": { + "question": "What is the primary purpose of the RelD discriminator introduced in the paper?", + "choices": [ + "To generate bilingual question-answering datasets", + "To detect hallucinations in LLM-generated answers", + "To improve sliding window segmentation of long texts", + "To train large language models from scratch" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the primary purpose of the RelD discriminator introduced in the paper?\nAnswer:", + " To generate bilingual question-answering datasets" + ], + [ + "Question:What is the primary purpose of the RelD discriminator introduced in the paper?\nAnswer:", + " To detect hallucinations in LLM-generated answers" + ], + [ + "Question:What is the primary purpose of the RelD discriminator introduced in the paper?\nAnswer:", + " To improve sliding window segmentation of long texts" + ], + [ + "Question:What is the primary purpose of the RelD discriminator introduced in the paper?\nAnswer:", + " To train large language models from scratch" + ] + ], + "resps": [ + [ + [ + -31.954849243164062, + false + ] + ], + [ + [ + -20.652280807495117, + false + ] + ], + [ + [ + -38.18414306640625, + false + ] + ], + [ + [ + -22.242000579833984, + false + ] + ] + ], + "filtered_resps": [ + [ + -31.954849243164062, + false + ], + [ + -20.652280807495117, + false + ], + [ + -38.18414306640625, + false + ], + [ + -22.242000579833984, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f8715d4609cb5bd0ab47ee0f64b797ae754515a5d357c96e9d8936537c91a1e7", + "prompt_hash": "b5f243aca35069a0a609e6a4b62a5094d9167bd082cf70e213cd81152e70a692", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 31, + "doc": { + "question": "How many samples does the RelQA dataset contain in total?", + "choices": [ + "274,426 samples", + "1,372,130 samples", + "743,910 samples", + "628,220 samples" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many samples does the RelQA dataset contain in total?\nAnswer:", + " 274,426 samples" + ], + [ + "Question:How many samples does the RelQA dataset contain in total?\nAnswer:", + " 1,372,130 samples" + ], + [ + "Question:How many samples does the RelQA dataset contain in total?\nAnswer:", + " 743,910 samples" + ], + [ + "Question:How many samples does the RelQA dataset contain in total?\nAnswer:", + " 628,220 samples" + ] + ], + "resps": [ + [ + [ + -21.363178253173828, + false + ] + ], + [ + [ + -22.405603408813477, + false + ] + ], + [ + [ + -23.08317756652832, + false + ] + ], + [ + [ + -22.19939613342285, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.363178253173828, + false + ], + [ + -22.405603408813477, + false + ], + [ + -23.08317756652832, + false + ], + [ + -22.19939613342285, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d3f58d5d62b77f94ea2d5697f862acb17f8f7c1e25b69c1d8c43e08af54e2f04", + "prompt_hash": "38f2ec70ddb9d7d428e15946c1d8056264f914f972aab69e57de31b99df853b3", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 32, + "doc": { + "question": "Which pre-trained language model serves as the backbone of RelD?", + "choices": [ + "BERT", + "RoBERTa", + "ELECTRA", + "DeBERTa" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which pre-trained language model serves as the backbone of RelD?\nAnswer:", + " BERT" + ], + [ + "Question:Which pre-trained language model serves as the backbone of RelD?\nAnswer:", + " RoBERTa" + ], + [ + "Question:Which pre-trained language model serves as the backbone of RelD?\nAnswer:", + " ELECTRA" + ], + [ + "Question:Which pre-trained language model serves as the backbone of RelD?\nAnswer:", + " DeBERTa" + ] + ], + "resps": [ + [ + [ + -4.401993751525879, + false + ] + ], + [ + [ + -5.599353790283203, + false + ] + ], + [ + [ + -8.7052583694458, + false + ] + ], + [ + [ + -7.099686145782471, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.401993751525879, + false + ], + [ + -5.599353790283203, + false + ], + [ + -8.7052583694458, + false + ], + [ + -7.099686145782471, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7e3932664074d45608cc62d85b4eaf400fcbe60f05f32f31d1ccad4f3a7268e0", + "prompt_hash": "ce210da2c71a35dbf6414602d2fe6c56508da12047c61004b12c3c94875361cb", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 33, + "doc": { + "question": "Which four types of metrics are used to evaluate the reliability of LLM-generated answers in RelQA?", + "choices": [ + "Accuracy, F1, BLEU, and ROUGE", + "LLM-assessment, human, machine, and composite metrics", + "Similarity, diversity, overlap, and embedding metrics", + "Precision, recall, diversity, and entropy metrics" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which four types of metrics are used to evaluate the reliability of LLM-generated answers in RelQA?\nAnswer:", + " Accuracy, F1, BLEU, and ROUGE" + ], + [ + "Question:Which four types of metrics are used to evaluate the reliability of LLM-generated answers in RelQA?\nAnswer:", + " LLM-assessment, human, machine, and composite metrics" + ], + [ + "Question:Which four types of metrics are used to evaluate the reliability of LLM-generated answers in RelQA?\nAnswer:", + " Similarity, diversity, overlap, and embedding metrics" + ], + [ + "Question:Which four types of metrics are used to evaluate the reliability of LLM-generated answers in RelQA?\nAnswer:", + " Precision, recall, diversity, and entropy metrics" + ] + ], + "resps": [ + [ + [ + -15.617055892944336, + false + ] + ], + [ + [ + -41.019859313964844, + false + ] + ], + [ + [ + -42.50098419189453, + false + ] + ], + [ + [ + -33.57671356201172, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.617055892944336, + false + ], + [ + -41.019859313964844, + false + ], + [ + -42.50098419189453, + false + ], + [ + -33.57671356201172, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c00e773de89a2acc29af1502161379c84980fbbb886884125ab8cbda482d5ba2", + "prompt_hash": "c7f2336b716039b45f75c7bee32aa77a60218ce7f2aca6f12585b3c778aa3337", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 34, + "doc": { + "question": "What is the learning rate set for training RelD?", + "choices": [ + "1e-05", + "5e-05", + "2e-05", + "3e-04" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the learning rate set for training RelD?\nAnswer:", + " 1e-05" + ], + [ + "Question:What is the learning rate set for training RelD?\nAnswer:", + " 5e-05" + ], + [ + "Question:What is the learning rate set for training RelD?\nAnswer:", + " 2e-05" + ], + [ + "Question:What is the learning rate set for training RelD?\nAnswer:", + " 3e-04" + ] + ], + "resps": [ + [ + [ + -9.810700416564941, + false + ] + ], + [ + [ + -11.288911819458008, + false + ] + ], + [ + [ + -11.6030912399292, + false + ] + ], + [ + [ + -12.463712692260742, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.810700416564941, + false + ], + [ + -11.288911819458008, + false + ], + [ + -11.6030912399292, + false + ], + [ + -12.463712692260742, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "86aed700016830d9aeb0a64b4d9f1b0b79c0d532f85d4b81a31553de5527b86e", + "prompt_hash": "a227d0d87d2da3b0353bc9d72add06c0edb62611e79dcf170ccae53537461e24", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 35, + "doc": { + "question": "What is the main advantage of using supervised semantic tokens in CosyVoice?", + "choices": [ + "They reduce the size of the model", + "They improve content consistency and speaker similarity", + "They speed up the training process", + "They eliminate the need for Mel spectrograms" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main advantage of using supervised semantic tokens in CosyVoice?\nAnswer:", + " They reduce the size of the model" + ], + [ + "Question:What is the main advantage of using supervised semantic tokens in CosyVoice?\nAnswer:", + " They improve content consistency and speaker similarity" + ], + [ + "Question:What is the main advantage of using supervised semantic tokens in CosyVoice?\nAnswer:", + " They speed up the training process" + ], + [ + "Question:What is the main advantage of using supervised semantic tokens in CosyVoice?\nAnswer:", + " They eliminate the need for Mel spectrograms" + ] + ], + "resps": [ + [ + [ + -16.160953521728516, + false + ] + ], + [ + [ + -23.697269439697266, + false + ] + ], + [ + [ + -14.877523422241211, + false + ] + ], + [ + [ + -23.25432777404785, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.160953521728516, + false + ], + [ + -23.697269439697266, + false + ], + [ + -14.877523422241211, + false + ], + [ + -23.25432777404785, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "32f9666275762353507dd197f4616bf5e893a724a0fdf6fbdae89cf1ad4752b6", + "prompt_hash": "a2fd5519d6723f41f66469ab69bdc64db9b62dee24c06d3f546dfa8ddd5dbe22", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 36, + "doc": { + "question": "Which technique does CosyVoice employ instead of diffusion probabilistic models (DDPM) for faster training and inference?", + "choices": [ + "Generative Adversarial Networks", + "Conditional Flow Matching", + "Recurrent Neural Networks", + "Variational Autoencoders" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which technique does CosyVoice employ instead of diffusion probabilistic models (DDPM) for faster training and inference?\nAnswer:", + " Generative Adversarial Networks" + ], + [ + "Question:Which technique does CosyVoice employ instead of diffusion probabilistic models (DDPM) for faster training and inference?\nAnswer:", + " Conditional Flow Matching" + ], + [ + "Question:Which technique does CosyVoice employ instead of diffusion probabilistic models (DDPM) for faster training and inference?\nAnswer:", + " Recurrent Neural Networks" + ], + [ + "Question:Which technique does CosyVoice employ instead of diffusion probabilistic models (DDPM) for faster training and inference?\nAnswer:", + " Variational Autoencoders" + ] + ], + "resps": [ + [ + [ + -7.09904146194458, + false + ] + ], + [ + [ + -17.3647518157959, + false + ] + ], + [ + [ + -10.57017707824707, + false + ] + ], + [ + [ + -10.066808700561523, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.09904146194458, + false + ], + [ + -17.3647518157959, + false + ], + [ + -10.57017707824707, + false + ], + [ + -10.066808700561523, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f90748ba30d980c60f68438996820d4d143ace3468c3d3363b6e9f8b67a555f8", + "prompt_hash": "b4be1101c4959c8e45f1584f97ab24176d0b2cf2ab9983d2f030b79fff120f29", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 37, + "doc": { + "question": "How many codes are included in the single codebook used by the supervised semantic speech tokenizer?", + "choices": [ + "1,024", + "2,048", + "4,096", + "8,192" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many codes are included in the single codebook used by the supervised semantic speech tokenizer?\nAnswer:", + " 1,024" + ], + [ + "Question:How many codes are included in the single codebook used by the supervised semantic speech tokenizer?\nAnswer:", + " 2,048" + ], + [ + "Question:How many codes are included in the single codebook used by the supervised semantic speech tokenizer?\nAnswer:", + " 4,096" + ], + [ + "Question:How many codes are included in the single codebook used by the supervised semantic speech tokenizer?\nAnswer:", + " 8,192" + ] + ], + "resps": [ + [ + [ + -10.077766418457031, + false + ] + ], + [ + [ + -11.457889556884766, + false + ] + ], + [ + [ + -10.916120529174805, + false + ] + ], + [ + [ + -14.03486442565918, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.077766418457031, + false + ], + [ + -11.457889556884766, + false + ], + [ + -10.916120529174805, + false + ], + [ + -14.03486442565918, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6edd2ab81831c1f356526e910ee0f90df1e2589eb38db06f14ab76ab986e0f65", + "prompt_hash": "797df02bbcfc8d676a055ee518485d78dc7d9013519df369e7bfd8ddc9ad3112", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 38, + "doc": { + "question": "On the LibriTTS test-clean set, what word error rate (WER) does CosyVoice achieve after ASR re-ranking?", + "choices": [ + "2.89%", + "1.51%", + "3.17%", + "8.32%" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:On the LibriTTS test-clean set, what word error rate (WER) does CosyVoice achieve after ASR re-ranking?\nAnswer:", + " 2.89%" + ], + [ + "Question:On the LibriTTS test-clean set, what word error rate (WER) does CosyVoice achieve after ASR re-ranking?\nAnswer:", + " 1.51%" + ], + [ + "Question:On the LibriTTS test-clean set, what word error rate (WER) does CosyVoice achieve after ASR re-ranking?\nAnswer:", + " 3.17%" + ], + [ + "Question:On the LibriTTS test-clean set, what word error rate (WER) does CosyVoice achieve after ASR re-ranking?\nAnswer:", + " 8.32%" + ] + ], + "resps": [ + [ + [ + -14.48992919921875, + false + ] + ], + [ + [ + -14.496801376342773, + false + ] + ], + [ + [ + -14.643253326416016, + false + ] + ], + [ + [ + -14.531301498413086, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.48992919921875, + false + ], + [ + -14.496801376342773, + false + ], + [ + -14.643253326416016, + false + ], + [ + -14.531301498413086, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d52e2878a69c9abff4521937fa3f88ff1cc0b707d0dd5a8a8e6cf741df6a6ceb", + "prompt_hash": "47292ce2c33805e39769dd3f1458ca53301685a873d6e66320ebd6d129b198bf", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 39, + "doc": { + "question": "Which emotion showed the greatest improvement in control accuracy when using CosyVoice-instruct compared to CosyVoice-base?", + "choices": [ + "Happy", + "Sad", + "Disgusted", + "Fearful" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which emotion showed the greatest improvement in control accuracy when using CosyVoice-instruct compared to CosyVoice-base?\nAnswer:", + " Happy" + ], + [ + "Question:Which emotion showed the greatest improvement in control accuracy when using CosyVoice-instruct compared to CosyVoice-base?\nAnswer:", + " Sad" + ], + [ + "Question:Which emotion showed the greatest improvement in control accuracy when using CosyVoice-instruct compared to CosyVoice-base?\nAnswer:", + " Disgusted" + ], + [ + "Question:Which emotion showed the greatest improvement in control accuracy when using CosyVoice-instruct compared to CosyVoice-base?\nAnswer:", + " Fearful" + ] + ], + "resps": [ + [ + [ + -7.383396148681641, + false + ] + ], + [ + [ + -8.395179748535156, + false + ] + ], + [ + [ + -15.370882987976074, + false + ] + ], + [ + [ + -10.768058776855469, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.383396148681641, + false + ], + [ + -8.395179748535156, + false + ], + [ + -15.370882987976074, + false + ], + [ + -10.768058776855469, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b74c37d3171aa5ca64392e549cde202d5fb7ea554fc7574d3e69b67e5e0e0ff6", + "prompt_hash": "6ab38ced6f896b1247e9a87a0a715bb24483db0e6ba01b45e0d480ef27562f5b", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 40, + "doc": { + "question": "What is the main advantage of WTConv over standard depth-wise convolutions?", + "choices": [ + "It reduces the number of trainable parameters quadratically", + "It achieves a larger receptive field with logarithmic parameter growth", + "It uses attention mechanisms instead of convolutions", + "It converts the input entirely into the frequency domain" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main advantage of WTConv over standard depth-wise convolutions?\nAnswer:", + " It reduces the number of trainable parameters quadratically" + ], + [ + "Question:What is the main advantage of WTConv over standard depth-wise convolutions?\nAnswer:", + " It achieves a larger receptive field with logarithmic parameter growth" + ], + [ + "Question:What is the main advantage of WTConv over standard depth-wise convolutions?\nAnswer:", + " It uses attention mechanisms instead of convolutions" + ], + [ + "Question:What is the main advantage of WTConv over standard depth-wise convolutions?\nAnswer:", + " It converts the input entirely into the frequency domain" + ] + ], + "resps": [ + [ + [ + -22.36676025390625, + false + ] + ], + [ + [ + -45.17839050292969, + false + ] + ], + [ + [ + -18.9272518157959, + false + ] + ], + [ + [ + -33.368980407714844, + false + ] + ] + ], + "filtered_resps": [ + [ + -22.36676025390625, + false + ], + [ + -45.17839050292969, + false + ], + [ + -18.9272518157959, + false + ], + [ + -33.368980407714844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "bb9319f6f6523e6ef30fa02d44e6c00f46a83477073de301e0af4660f8332c1d", + "prompt_hash": "7c42e02ea55d1ec0818a0a5ad17920e6c29f4796dd7632d58084bb510f9888df", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 41, + "doc": { + "question": "Which wavelet basis is used in WTConv for efficiency?", + "choices": [ + "Fourier Transform", + "Haar Wavelet Transform", + "Daubechies Wavelet Transform", + "Discrete Cosine Transform" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which wavelet basis is used in WTConv for efficiency?\nAnswer:", + " Fourier Transform" + ], + [ + "Question:Which wavelet basis is used in WTConv for efficiency?\nAnswer:", + " Haar Wavelet Transform" + ], + [ + "Question:Which wavelet basis is used in WTConv for efficiency?\nAnswer:", + " Daubechies Wavelet Transform" + ], + [ + "Question:Which wavelet basis is used in WTConv for efficiency?\nAnswer:", + " Discrete Cosine Transform" + ] + ], + "resps": [ + [ + [ + -10.868124008178711, + false + ] + ], + [ + [ + -13.273340225219727, + false + ] + ], + [ + [ + -20.367637634277344, + false + ] + ], + [ + [ + -7.96662712097168, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.868124008178711, + false + ], + [ + -13.273340225219727, + false + ], + [ + -20.367637634277344, + false + ], + [ + -7.96662712097168, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8600813d0c358ab4185c7d49a3cb37e183c975e849c39676fde3a2728aec51e2", + "prompt_hash": "ec9b5a9e71d2ce56edf7cbb8827ec8527ac3518a2695db961815d6c6df323004", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 42, + "doc": { + "question": "What accuracy improvement does WTConvNeXt-T achieve over ConvNeXt-T in ImageNet-C mean corruption error (mCE)?", + "choices": [ + "53.2% to 52.0%", + "41.6% to 39.0%", + "55.0% to 50.0%", + "60.5% to 58.0%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What accuracy improvement does WTConvNeXt-T achieve over ConvNeXt-T in ImageNet-C mean corruption error (mCE)?\nAnswer:", + " 53.2% to 52.0%" + ], + [ + "Question:What accuracy improvement does WTConvNeXt-T achieve over ConvNeXt-T in ImageNet-C mean corruption error (mCE)?\nAnswer:", + " 41.6% to 39.0%" + ], + [ + "Question:What accuracy improvement does WTConvNeXt-T achieve over ConvNeXt-T in ImageNet-C mean corruption error (mCE)?\nAnswer:", + " 55.0% to 50.0%" + ], + [ + "Question:What accuracy improvement does WTConvNeXt-T achieve over ConvNeXt-T in ImageNet-C mean corruption error (mCE)?\nAnswer:", + " 60.5% to 58.0%" + ] + ], + "resps": [ + [ + [ + -36.595645904541016, + false + ] + ], + [ + [ + -38.15071105957031, + false + ] + ], + [ + [ + -34.00459289550781, + false + ] + ], + [ + [ + -36.84385681152344, + false + ] + ] + ], + "filtered_resps": [ + [ + -36.595645904541016, + false + ], + [ + -38.15071105957031, + false + ], + [ + -34.00459289550781, + false + ], + [ + -36.84385681152344, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6da62bf2f48179bd7ba682edf10f2964d3717e3a6cd687e2adf2a18abcdc82e7", + "prompt_hash": "c5bffff41cccf839939a88f2eba6f43dbda372e284f2e9876b56ee8d8ea5de3e", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 43, + "doc": { + "question": "How does WTConvNeXt affect shape-bias compared to ConvNeXt?", + "choices": [ + "It decreases the network's shape-bias", + "It keeps the shape-bias unchanged", + "It significantly increases the shape-bias", + "It eliminates texture-based decisions entirely" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How does WTConvNeXt affect shape-bias compared to ConvNeXt?\nAnswer:", + " It decreases the network's shape-bias" + ], + [ + "Question:How does WTConvNeXt affect shape-bias compared to ConvNeXt?\nAnswer:", + " It keeps the shape-bias unchanged" + ], + [ + "Question:How does WTConvNeXt affect shape-bias compared to ConvNeXt?\nAnswer:", + " It significantly increases the shape-bias" + ], + [ + "Question:How does WTConvNeXt affect shape-bias compared to ConvNeXt?\nAnswer:", + " It eliminates texture-based decisions entirely" + ] + ], + "resps": [ + [ + [ + -20.318492889404297, + false + ] + ], + [ + [ + -19.004671096801758, + false + ] + ], + [ + [ + -15.258134841918945, + false + ] + ], + [ + [ + -38.38514709472656, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.318492889404297, + false + ], + [ + -19.004671096801758, + false + ], + [ + -15.258134841918945, + false + ], + [ + -38.38514709472656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "163c68f81c202eaa477d5551c1c8c8d183e8e1dc6ba0b7e692cf5f2546bd3469", + "prompt_hash": "aa579b89dca17c357b86632fbae490cab4afe1b2802815f55b17e112975080cd", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 44, + "doc": { + "question": "What is the primary reason WTConv achieves better robustness under image corruptions?", + "choices": [ + "It focuses entirely on high-frequency information", + "It integrates Vision Transformer attention mechanisms", + "It emphasizes low-frequency components through wavelet decomposition", + "It replaces convolutional layers with pooling layers" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the primary reason WTConv achieves better robustness under image corruptions?\nAnswer:", + " It focuses entirely on high-frequency information" + ], + [ + "Question:What is the primary reason WTConv achieves better robustness under image corruptions?\nAnswer:", + " It integrates Vision Transformer attention mechanisms" + ], + [ + "Question:What is the primary reason WTConv achieves better robustness under image corruptions?\nAnswer:", + " It emphasizes low-frequency components through wavelet decomposition" + ], + [ + "Question:What is the primary reason WTConv achieves better robustness under image corruptions?\nAnswer:", + " It replaces convolutional layers with pooling layers" + ] + ], + "resps": [ + [ + [ + -25.85192108154297, + false + ] + ], + [ + [ + -28.19028091430664, + false + ] + ], + [ + [ + -27.371051788330078, + false + ] + ], + [ + [ + -21.658126831054688, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.85192108154297, + false + ], + [ + -28.19028091430664, + false + ], + [ + -27.371051788330078, + false + ], + [ + -21.658126831054688, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6bc68414e1b2a06374963af51331f539c7fad1173ed20e9990a150fe18ffd896", + "prompt_hash": "3e8d462f256feaf421b3dee7d20bac60366c94311ce2b0a57250257b68dbc4d0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 45, + "doc": { + "question": "Which model was used as the base model for all experiments in the study?", + "choices": [ + "GPT-4-0125-preview", + "Llama-2-70B-chat", + "PaLM-2", + "Claude-3" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which model was used as the base model for all experiments in the study?\nAnswer:", + " GPT-4-0125-preview" + ], + [ + "Question:Which model was used as the base model for all experiments in the study?\nAnswer:", + " Llama-2-70B-chat" + ], + [ + "Question:Which model was used as the base model for all experiments in the study?\nAnswer:", + " PaLM-2" + ], + [ + "Question:Which model was used as the base model for all experiments in the study?\nAnswer:", + " Claude-3" + ] + ], + "resps": [ + [ + [ + -11.283348083496094, + false + ] + ], + [ + [ + -16.040966033935547, + false + ] + ], + [ + [ + -14.582185745239258, + false + ] + ], + [ + [ + -12.037026405334473, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.283348083496094, + false + ], + [ + -16.040966033935547, + false + ], + [ + -14.582185745239258, + false + ], + [ + -12.037026405334473, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1383145086897e76bb0d4776b6c1e6e80782890a7ed056a1f72b9c1316943d93", + "prompt_hash": "c7d8e5ee70ecb3ea4ac98b838f06b89f9fe0824e8eb73f877a757d10423cbcf7", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 46, + "doc": { + "question": "What was the achieved accuracy after distilling the 2-Step Rephrase and Respond (RaR) method into a System 1 Llama-2-70B-chat model on the last letter concatenation task?", + "choices": [ + "44.5%", + "56.1%", + "98.0%", + "69.5%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What was the achieved accuracy after distilling the 2-Step Rephrase and Respond (RaR) method into a System 1 Llama-2-70B-chat model on the last letter concatenation task?\nAnswer:", + " 44.5%" + ], + [ + "Question:What was the achieved accuracy after distilling the 2-Step Rephrase and Respond (RaR) method into a System 1 Llama-2-70B-chat model on the last letter concatenation task?\nAnswer:", + " 56.1%" + ], + [ + "Question:What was the achieved accuracy after distilling the 2-Step Rephrase and Respond (RaR) method into a System 1 Llama-2-70B-chat model on the last letter concatenation task?\nAnswer:", + " 98.0%" + ], + [ + "Question:What was the achieved accuracy after distilling the 2-Step Rephrase and Respond (RaR) method into a System 1 Llama-2-70B-chat model on the last letter concatenation task?\nAnswer:", + " 69.5%" + ] + ], + "resps": [ + [ + [ + -15.09747314453125, + false + ] + ], + [ + [ + -14.72747802734375, + false + ] + ], + [ + [ + -13.923602104187012, + false + ] + ], + [ + [ + -14.50871467590332, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.09747314453125, + false + ], + [ + -14.72747802734375, + false + ], + [ + -13.923602104187012, + false + ], + [ + -14.50871467590332, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e60644694d85dc759b2c978a6ff884c55b4a5e79508100ae0e6c472092b69dec", + "prompt_hash": "fb69b33d2d19e03a16ce11134296fc4d84787079fc51f3d1a2112d672dab78c3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 47, + "doc": { + "question": "How does System 2 Attention (S2A) primarily improve reasoning performance?", + "choices": [ + "By adding more training data to reduce overfitting", + "By rewriting the input to remove biases or irrelevant information", + "By introducing more intermediate reasoning steps", + "By increasing the model size and token limit" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How does System 2 Attention (S2A) primarily improve reasoning performance?\nAnswer:", + " By adding more training data to reduce overfitting" + ], + [ + "Question:How does System 2 Attention (S2A) primarily improve reasoning performance?\nAnswer:", + " By rewriting the input to remove biases or irrelevant information" + ], + [ + "Question:How does System 2 Attention (S2A) primarily improve reasoning performance?\nAnswer:", + " By introducing more intermediate reasoning steps" + ], + [ + "Question:How does System 2 Attention (S2A) primarily improve reasoning performance?\nAnswer:", + " By increasing the model size and token limit" + ] + ], + "resps": [ + [ + [ + -30.416637420654297, + false + ] + ], + [ + [ + -42.379478454589844, + false + ] + ], + [ + [ + -22.114347457885742, + false + ] + ], + [ + [ + -35.356021881103516, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.416637420654297, + false + ], + [ + -42.379478454589844, + false + ], + [ + -22.114347457885742, + false + ], + [ + -35.356021881103516, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f6fc0d851d2fdff1716418a5330d89655fae039bfe7cfebe205dac52103a1eea", + "prompt_hash": "947a69f16e4b2bc4a7c5f9520c71e19f88c02112f65ee62363a8077b994d2cd5", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 48, + "doc": { + "question": "What technique was used to ensure the quality of distillation data for Branch-Solve-Merge (BSM)?", + "choices": [ + "Cross-entropy minimization", + "Majority voting across outputs", + "Self-consistency under input perturbations", + "Manual annotation of all samples" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What technique was used to ensure the quality of distillation data for Branch-Solve-Merge (BSM)?\nAnswer:", + " Cross-entropy minimization" + ], + [ + "Question:What technique was used to ensure the quality of distillation data for Branch-Solve-Merge (BSM)?\nAnswer:", + " Majority voting across outputs" + ], + [ + "Question:What technique was used to ensure the quality of distillation data for Branch-Solve-Merge (BSM)?\nAnswer:", + " Self-consistency under input perturbations" + ], + [ + "Question:What technique was used to ensure the quality of distillation data for Branch-Solve-Merge (BSM)?\nAnswer:", + " Manual annotation of all samples" + ] + ], + "resps": [ + [ + [ + -17.37027359008789, + false + ] + ], + [ + [ + -24.0562744140625, + false + ] + ], + [ + [ + -29.128202438354492, + false + ] + ], + [ + [ + -19.985612869262695, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.37027359008789, + false + ], + [ + -24.0562744140625, + false + ], + [ + -29.128202438354492, + false + ], + [ + -19.985612869262695, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "36344b17e47b97610171de862ccd12375a23e6f878dc0a1d2cfe111f23838953", + "prompt_hash": "87290dac7772ef7346b3ef0d92b9f5bf5574ad93388b41a8ad56a49a56d96e45", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 49, + "doc": { + "question": "Which reasoning task showed poor performance when applying System 2 distillation, indicating that it could not be effectively distilled?", + "choices": [ + "Last letter concatenation task", + "Coin flip reasoning task", + "SycophancyEval biased QA task", + "GSM8k math problem-solving task" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which reasoning task showed poor performance when applying System 2 distillation, indicating that it could not be effectively distilled?\nAnswer:", + " Last letter concatenation task" + ], + [ + "Question:Which reasoning task showed poor performance when applying System 2 distillation, indicating that it could not be effectively distilled?\nAnswer:", + " Coin flip reasoning task" + ], + [ + "Question:Which reasoning task showed poor performance when applying System 2 distillation, indicating that it could not be effectively distilled?\nAnswer:", + " SycophancyEval biased QA task" + ], + [ + "Question:Which reasoning task showed poor performance when applying System 2 distillation, indicating that it could not be effectively distilled?\nAnswer:", + " GSM8k math problem-solving task" + ] + ], + "resps": [ + [ + [ + -26.82872772216797, + false + ] + ], + [ + [ + -20.605525970458984, + false + ] + ], + [ + [ + -49.651554107666016, + false + ] + ], + [ + [ + -26.793594360351562, + false + ] + ] + ], + "filtered_resps": [ + [ + -26.82872772216797, + false + ], + [ + -20.605525970458984, + false + ], + [ + -49.651554107666016, + false + ], + [ + -26.793594360351562, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0c5a64651e593f7334dc76475ccce742ed0be1ae580af4895bb8dc289b0cda68", + "prompt_hash": "2c9437c3ff992b9dd1cb1a469649859a1c6719d05eb3060708ccfffa956091b7", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 50, + "doc": { + "question": "What is the precisely measured distance to pulsar PSR J0437−4715 based on radio timing observations?", + "choices": [ + "156.98 ± 0.15 pc", + "129.4 ± 0.5 pc", + "174.0 ± 0.2 pc", + "200.15 ± 1.0 pc" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the precisely measured distance to pulsar PSR J0437−4715 based on radio timing observations?\nAnswer:", + " 156.98 ± 0.15 pc" + ], + [ + "Question:What is the precisely measured distance to pulsar PSR J0437−4715 based on radio timing observations?\nAnswer:", + " 129.4 ± 0.5 pc" + ], + [ + "Question:What is the precisely measured distance to pulsar PSR J0437−4715 based on radio timing observations?\nAnswer:", + " 174.0 ± 0.2 pc" + ], + [ + "Question:What is the precisely measured distance to pulsar PSR J0437−4715 based on radio timing observations?\nAnswer:", + " 200.15 ± 1.0 pc" + ] + ], + "resps": [ + [ + [ + -23.42844009399414, + false + ] + ], + [ + [ + -19.135929107666016, + false + ] + ], + [ + [ + -20.670236587524414, + false + ] + ], + [ + [ + -28.07659149169922, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.42844009399414, + false + ], + [ + -19.135929107666016, + false + ], + [ + -20.670236587524414, + false + ], + [ + -28.07659149169922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9136d143f6fbec7e4ca9165be4d452f05f6ad1a30173e9d1e6a91fe102584000", + "prompt_hash": "a50819b0366d5f70b4b331b6f66c1549975ffb36782f46e713520207c6b2cfda", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 51, + "doc": { + "question": "Which background modeling method used by NICER provides quantified statistical uncertainties in the net background spectrum?", + "choices": [ + "Space Weather estimate", + "3C50 background model", + "XMM-MOS1 background fit", + "Delta dataset filtering" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which background modeling method used by NICER provides quantified statistical uncertainties in the net background spectrum?\nAnswer:", + " Space Weather estimate" + ], + [ + "Question:Which background modeling method used by NICER provides quantified statistical uncertainties in the net background spectrum?\nAnswer:", + " 3C50 background model" + ], + [ + "Question:Which background modeling method used by NICER provides quantified statistical uncertainties in the net background spectrum?\nAnswer:", + " XMM-MOS1 background fit" + ], + [ + "Question:Which background modeling method used by NICER provides quantified statistical uncertainties in the net background spectrum?\nAnswer:", + " Delta dataset filtering" + ] + ], + "resps": [ + [ + [ + -23.84143829345703, + false + ] + ], + [ + [ + -15.207002639770508, + false + ] + ], + [ + [ + -34.21397018432617, + false + ] + ], + [ + [ + -29.099979400634766, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.84143829345703, + false + ], + [ + -15.207002639770508, + false + ], + [ + -34.21397018432617, + false + ], + [ + -29.099979400634766, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d29a9f90f4dd47e62853c3571f7346e637325b11296cc948f9f1ecdbda8a1ae4", + "prompt_hash": "7bd2c5d239ce1590ef0c12d0ebdcae61e4f703fada9defdf904df81a3da6d226", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 52, + "doc": { + "question": "What is the measured spin frequency of PSR J0437−4715?", + "choices": [ + "174 Hz", + "129 Hz", + "210 Hz", + "98 Hz" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the measured spin frequency of PSR J0437−4715?\nAnswer:", + " 174 Hz" + ], + [ + "Question:What is the measured spin frequency of PSR J0437−4715?\nAnswer:", + " 129 Hz" + ], + [ + "Question:What is the measured spin frequency of PSR J0437−4715?\nAnswer:", + " 210 Hz" + ], + [ + "Question:What is the measured spin frequency of PSR J0437−4715?\nAnswer:", + " 98 Hz" + ] + ], + "resps": [ + [ + [ + -11.427151679992676, + false + ] + ], + [ + [ + -10.656230926513672, + false + ] + ], + [ + [ + -10.553150177001953, + false + ] + ], + [ + [ + -10.735501289367676, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.427151679992676, + false + ], + [ + -10.656230926513672, + false + ], + [ + -10.553150177001953, + false + ], + [ + -10.735501289367676, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a3317615d62bc3dac178890e48f3b4bac1a63795c3104864a252cde455139a6b", + "prompt_hash": "7b73b6789f9631e1f0f035a051058ba0607731cd7fb355c18416ec15184b1031", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 53, + "doc": { + "question": "Which software package is used to perform Bayesian parameter estimation by modeling X-ray emission from PSR J0437−4715?", + "choices": [ + "HEASoft", + "X-PSI", + "SAS", + "NSX" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which software package is used to perform Bayesian parameter estimation by modeling X-ray emission from PSR J0437−4715?\nAnswer:", + " HEASoft" + ], + [ + "Question:Which software package is used to perform Bayesian parameter estimation by modeling X-ray emission from PSR J0437−4715?\nAnswer:", + " X-PSI" + ], + [ + "Question:Which software package is used to perform Bayesian parameter estimation by modeling X-ray emission from PSR J0437−4715?\nAnswer:", + " SAS" + ], + [ + "Question:Which software package is used to perform Bayesian parameter estimation by modeling X-ray emission from PSR J0437−4715?\nAnswer:", + " NSX" + ] + ], + "resps": [ + [ + [ + -11.748787879943848, + false + ] + ], + [ + [ + -11.375994682312012, + false + ] + ], + [ + [ + -8.618194580078125, + false + ] + ], + [ + [ + -13.982128143310547, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.748787879943848, + false + ], + [ + -11.375994682312012, + false + ], + [ + -8.618194580078125, + false + ], + [ + -13.982128143310547, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "40c129c953387082b60e7d8a841f9c300b1e201c6b80864f06642485aa6ff0a4", + "prompt_hash": "17722b0305f01c9020f7b3f55e579eacc373f25594dd32a17c1693d7bcaa088c", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 54, + "doc": { + "question": "In the NICER observations, what is the approximate contribution of the nearby AGN RX J0437.4−4711 to the 0.3–3 keV band?", + "choices": [ + "About 0.2 counts/s", + "About 1.5 counts/s", + "About 3.0 counts/s", + "About 5.0 counts/s" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:In the NICER observations, what is the approximate contribution of the nearby AGN RX J0437.4−4711 to the 0.3–3 keV band?\nAnswer:", + " About 0.2 counts/s" + ], + [ + "Question:In the NICER observations, what is the approximate contribution of the nearby AGN RX J0437.4−4711 to the 0.3–3 keV band?\nAnswer:", + " About 1.5 counts/s" + ], + [ + "Question:In the NICER observations, what is the approximate contribution of the nearby AGN RX J0437.4−4711 to the 0.3–3 keV band?\nAnswer:", + " About 3.0 counts/s" + ], + [ + "Question:In the NICER observations, what is the approximate contribution of the nearby AGN RX J0437.4−4711 to the 0.3–3 keV band?\nAnswer:", + " About 5.0 counts/s" + ] + ], + "resps": [ + [ + [ + -22.581018447875977, + false + ] + ], + [ + [ + -24.277639389038086, + false + ] + ], + [ + [ + -27.225370407104492, + false + ] + ], + [ + [ + -27.68436050415039, + false + ] + ] + ], + "filtered_resps": [ + [ + -22.581018447875977, + false + ], + [ + -24.277639389038086, + false + ], + [ + -27.225370407104492, + false + ], + [ + -27.68436050415039, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b32f1db2cede76d6f2cdf7c473fd967f1e38b84c0c10ee1f13c6bb56c16c79ba", + "prompt_hash": "7347866172412d992f39e909a8c48662828ecc18ea5b707099c207bbeb787234", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 55, + "doc": { + "question": "What is the primary purpose of EchoMimic in portrait animation?", + "choices": [ + "To generate portrait videos using both audio signals and facial landmarks", + "To synthesize 3D avatars from textual descriptions", + "To perform real-time lip synchronization using only 3D morphable models", + "To replace GAN-based approaches with Wav2Lip mechanisms" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the primary purpose of EchoMimic in portrait animation?\nAnswer:", + " To generate portrait videos using both audio signals and facial landmarks" + ], + [ + "Question:What is the primary purpose of EchoMimic in portrait animation?\nAnswer:", + " To synthesize 3D avatars from textual descriptions" + ], + [ + "Question:What is the primary purpose of EchoMimic in portrait animation?\nAnswer:", + " To perform real-time lip synchronization using only 3D morphable models" + ], + [ + "Question:What is the primary purpose of EchoMimic in portrait animation?\nAnswer:", + " To replace GAN-based approaches with Wav2Lip mechanisms" + ] + ], + "resps": [ + [ + [ + -35.45629119873047, + false + ] + ], + [ + [ + -22.518705368041992, + false + ] + ], + [ + [ + -41.64351272583008, + false + ] + ], + [ + [ + -52.94198226928711, + false + ] + ] + ], + "filtered_resps": [ + [ + -35.45629119873047, + false + ], + [ + -22.518705368041992, + false + ], + [ + -41.64351272583008, + false + ], + [ + -52.94198226928711, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "09fcb207ffa83b036d34b90fd8eb0c632f12ba8cfdc9f1187a854d65968ec3a8", + "prompt_hash": "86fb8321e982423113547efe693b5dcd70111ccc80a3075f4773c97e3a4aae1a", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 56, + "doc": { + "question": "Which architecture does EchoMimic's Denoising U-Net draw inspiration from?", + "choices": [ + "GAN-based MegaPortraits", + "Stable Diffusion v1.5", + "Transformer-based GPT architecture", + "3D Morphable Models" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which architecture does EchoMimic's Denoising U-Net draw inspiration from?\nAnswer:", + " GAN-based MegaPortraits" + ], + [ + "Question:Which architecture does EchoMimic's Denoising U-Net draw inspiration from?\nAnswer:", + " Stable Diffusion v1.5" + ], + [ + "Question:Which architecture does EchoMimic's Denoising U-Net draw inspiration from?\nAnswer:", + " Transformer-based GPT architecture" + ], + [ + "Question:Which architecture does EchoMimic's Denoising U-Net draw inspiration from?\nAnswer:", + " 3D Morphable Models" + ] + ], + "resps": [ + [ + [ + -43.4454345703125, + false + ] + ], + [ + [ + -12.865447044372559, + false + ] + ], + [ + [ + -22.89692497253418, + false + ] + ], + [ + [ + -18.237205505371094, + false + ] + ] + ], + "filtered_resps": [ + [ + -43.4454345703125, + false + ], + [ + -12.865447044372559, + false + ], + [ + -22.89692497253418, + false + ], + [ + -18.237205505371094, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e6976e2db7c66944c31a7d101b4a1fc4100abb4916f120c3eca5653f40f362a9", + "prompt_hash": "6e50180c89a59c87f3e94f877679ad6ef88b3aec3d9cd05c5b8d2210e0ea89e6", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 57, + "doc": { + "question": "What role does the Audio Encoder play in the EchoMimic framework?", + "choices": [ + "It extracts motion vectors from reference images", + "It encodes facial landmarks into latent vectors", + "It captures pronunciation and tonality features using Wav2Vec embeddings", + "It generates synthetic audio for video narration" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What role does the Audio Encoder play in the EchoMimic framework?\nAnswer:", + " It extracts motion vectors from reference images" + ], + [ + "Question:What role does the Audio Encoder play in the EchoMimic framework?\nAnswer:", + " It encodes facial landmarks into latent vectors" + ], + [ + "Question:What role does the Audio Encoder play in the EchoMimic framework?\nAnswer:", + " It captures pronunciation and tonality features using Wav2Vec embeddings" + ], + [ + "Question:What role does the Audio Encoder play in the EchoMimic framework?\nAnswer:", + " It generates synthetic audio for video narration" + ] + ], + "resps": [ + [ + [ + -34.38471984863281, + false + ] + ], + [ + [ + -26.634109497070312, + false + ] + ], + [ + [ + -49.42478561401367, + false + ] + ], + [ + [ + -31.105899810791016, + false + ] + ] + ], + "filtered_resps": [ + [ + -34.38471984863281, + false + ], + [ + -26.634109497070312, + false + ], + [ + -49.42478561401367, + false + ], + [ + -31.105899810791016, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9d677c7e270dbe2b16981def7e4de61fddd0be1ea7456d9731b5a34dee522568", + "prompt_hash": "b7e0dd728da0f3b6498eef1a70257eaff55685def4651d494e2ca39bfc11f50f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 58, + "doc": { + "question": "Which metric evaluates the structural similarity between generated and ground truth videos?", + "choices": [ + "FID", + "SSIM", + "E-FID", + "FVD" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which metric evaluates the structural similarity between generated and ground truth videos?\nAnswer:", + " FID" + ], + [ + "Question:Which metric evaluates the structural similarity between generated and ground truth videos?\nAnswer:", + " SSIM" + ], + [ + "Question:Which metric evaluates the structural similarity between generated and ground truth videos?\nAnswer:", + " E-FID" + ], + [ + "Question:Which metric evaluates the structural similarity between generated and ground truth videos?\nAnswer:", + " FVD" + ] + ], + "resps": [ + [ + [ + -6.555878639221191, + false + ] + ], + [ + [ + -6.250504493713379, + false + ] + ], + [ + [ + -14.620190620422363, + false + ] + ], + [ + [ + -7.614960670471191, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.555878639221191, + false + ], + [ + -6.250504493713379, + false + ], + [ + -14.620190620422363, + false + ], + [ + -7.614960670471191, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "886503ac67e97bca5a9de5a56393902b6df58329bb12d6031ef7d4013109082d", + "prompt_hash": "811a868f67d2ec860fff8a94061f5ec88bf6b45090c1b838da2763ff4bfd75ae", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 59, + "doc": { + "question": "On the HDTF dataset, what is EchoMimic's Fréchet Inception Distance (FID) score?", + "choices": [ + "41.535", + "37.659", + "29.136", + "53.143" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:On the HDTF dataset, what is EchoMimic's Fréchet Inception Distance (FID) score?\nAnswer:", + " 41.535" + ], + [ + "Question:On the HDTF dataset, what is EchoMimic's Fréchet Inception Distance (FID) score?\nAnswer:", + " 37.659" + ], + [ + "Question:On the HDTF dataset, what is EchoMimic's Fréchet Inception Distance (FID) score?\nAnswer:", + " 29.136" + ], + [ + "Question:On the HDTF dataset, what is EchoMimic's Fréchet Inception Distance (FID) score?\nAnswer:", + " 53.143" + ] + ], + "resps": [ + [ + [ + -15.374874114990234, + false + ] + ], + [ + [ + -15.217829704284668, + false + ] + ], + [ + [ + -15.107488632202148, + false + ] + ], + [ + [ + -16.111392974853516, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.374874114990234, + false + ], + [ + -15.217829704284668, + false + ], + [ + -15.107488632202148, + false + ], + [ + -16.111392974853516, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "91387f5df300c340c54a96119253cb331876a51247f16e1ee4a9b9cf43338efa", + "prompt_hash": "49cfff5c4848fe783889bc41c7a1a73832c0b168020a74022d71ead9be7badae", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 60, + "doc": { + "question": "Which of the following partial differential equations (PDEs) is included in the systematic review's scope?", + "choices": [ + "Korteweg–de Vries (KdV)", + "Darcy flow", + "Reaction-diffusion", + "Schrodinger equation" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which of the following partial differential equations (PDEs) is included in the systematic review's scope?\nAnswer:", + " Korteweg–de Vries (KdV)" + ], + [ + "Question:Which of the following partial differential equations (PDEs) is included in the systematic review's scope?\nAnswer:", + " Darcy flow" + ], + [ + "Question:Which of the following partial differential equations (PDEs) is included in the systematic review's scope?\nAnswer:", + " Reaction-diffusion" + ], + [ + "Question:Which of the following partial differential equations (PDEs) is included in the systematic review's scope?\nAnswer:", + " Schrodinger equation" + ] + ], + "resps": [ + [ + [ + -15.173296928405762, + false + ] + ], + [ + [ + -12.69983196258545, + false + ] + ], + [ + [ + -12.843111991882324, + false + ] + ], + [ + [ + -11.093381881713867, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.173296928405762, + false + ], + [ + -12.69983196258545, + false + ], + [ + -12.843111991882324, + false + ], + [ + -11.093381881713867, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7ae96dc5fbdda19d7b2f4085705065a1176fc56e5bd7815308596acc0bbad20e", + "prompt_hash": "5ef7dc1b26033a68d8db77326c0f3859286daabd58adfabec08b0aa7135b612c", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 61, + "doc": { + "question": "Why are physics-informed neural network (PINN) methods excluded from the systematic review?", + "choices": [ + "They are too slow and inefficient for any PDEs", + "The literature is too large and they rarely outperform standard methods", + "They cannot solve fluid-related PDEs", + "They only work for inverse problems" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Why are physics-informed neural network (PINN) methods excluded from the systematic review?\nAnswer:", + " They are too slow and inefficient for any PDEs" + ], + [ + "Question:Why are physics-informed neural network (PINN) methods excluded from the systematic review?\nAnswer:", + " The literature is too large and they rarely outperform standard methods" + ], + [ + "Question:Why are physics-informed neural network (PINN) methods excluded from the systematic review?\nAnswer:", + " They cannot solve fluid-related PDEs" + ], + [ + "Question:Why are physics-informed neural network (PINN) methods excluded from the systematic review?\nAnswer:", + " They only work for inverse problems" + ] + ], + "resps": [ + [ + [ + -35.40483093261719, + false + ] + ], + [ + [ + -45.47227096557617, + false + ] + ], + [ + [ + -27.10033416748047, + false + ] + ], + [ + [ + -22.625762939453125, + false + ] + ] + ], + "filtered_resps": [ + [ + -35.40483093261719, + false + ], + [ + -45.47227096557617, + false + ], + [ + -27.10033416748047, + false + ], + [ + -22.625762939453125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d15fcbecf4c315a4ff5d8f95930e335c8a24e7646c955eaf0dfa6d648e779793", + "prompt_hash": "39f1f8d7d75d7a19e4298d2162bf7be88ded1d3acbaa453c06721139e58bd246", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 62, + "doc": { + "question": "Which hardware comparison is generally considered the fairest when evaluating ML-based PDE solvers?", + "choices": [ + "GPU-to-GPU or TPU-to-TPU comparisons", + "CPU-to-TPU comparisons", + "CPU-to-GPU comparisons only", + "There is no fair comparison method" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which hardware comparison is generally considered the fairest when evaluating ML-based PDE solvers?\nAnswer:", + " GPU-to-GPU or TPU-to-TPU comparisons" + ], + [ + "Question:Which hardware comparison is generally considered the fairest when evaluating ML-based PDE solvers?\nAnswer:", + " CPU-to-TPU comparisons" + ], + [ + "Question:Which hardware comparison is generally considered the fairest when evaluating ML-based PDE solvers?\nAnswer:", + " CPU-to-GPU comparisons only" + ], + [ + "Question:Which hardware comparison is generally considered the fairest when evaluating ML-based PDE solvers?\nAnswer:", + " There is no fair comparison method" + ] + ], + "resps": [ + [ + [ + -31.690475463867188, + false + ] + ], + [ + [ + -28.010419845581055, + false + ] + ], + [ + [ + -24.789653778076172, + false + ] + ], + [ + [ + -16.941194534301758, + false + ] + ] + ], + "filtered_resps": [ + [ + -31.690475463867188, + false + ], + [ + -28.010419845581055, + false + ], + [ + -24.789653778076172, + false + ], + [ + -16.941194534301758, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6193794c56683e9761d433a69015d026a6ca3835797950eff3d94e5111d1347e", + "prompt_hash": "eb207fbf62331da84da86793e4428fc9b8395d436696beb3a47488680ba4da88", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 63, + "doc": { + "question": "What is the recommended numerical method for solving large elliptic PDE problems according to the review?", + "choices": [ + "Finite element methods (FEM)", + "Multigrid solvers", + "Pseudo-spectral methods", + "First-order finite volume methods" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the recommended numerical method for solving large elliptic PDE problems according to the review?\nAnswer:", + " Finite element methods (FEM)" + ], + [ + "Question:What is the recommended numerical method for solving large elliptic PDE problems according to the review?\nAnswer:", + " Multigrid solvers" + ], + [ + "Question:What is the recommended numerical method for solving large elliptic PDE problems according to the review?\nAnswer:", + " Pseudo-spectral methods" + ], + [ + "Question:What is the recommended numerical method for solving large elliptic PDE problems according to the review?\nAnswer:", + " First-order finite volume methods" + ] + ], + "resps": [ + [ + [ + -12.821867942810059, + false + ] + ], + [ + [ + -15.478704452514648, + false + ] + ], + [ + [ + -11.922616004943848, + false + ] + ], + [ + [ + -19.134483337402344, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.821867942810059, + false + ], + [ + -15.478704452514648, + false + ], + [ + -11.922616004943848, + false + ], + [ + -19.134483337402344, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "58520cf717a0b780213d59c994f481739f3e52565e02e4eb0daad4a7a72b0cfe", + "prompt_hash": "00e9209e4c905666d275b0a8d0a74e6aded50b08c2920900a20481a5baecdec2", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 64, + "doc": { + "question": "According to the random sample of PINN articles, what is the typical runtime difference between PINN-based solvers and standard numerical methods?", + "choices": [ + "PINNs usually solve PDEs faster than standard methods", + "PINNs and standard methods take about the same time", + "PINNs take hours to days, while standard methods take fractions of a second to minutes", + "PINNs are always unusable for PDEs" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:According to the random sample of PINN articles, what is the typical runtime difference between PINN-based solvers and standard numerical methods?\nAnswer:", + " PINNs usually solve PDEs faster than standard methods" + ], + [ + "Question:According to the random sample of PINN articles, what is the typical runtime difference between PINN-based solvers and standard numerical methods?\nAnswer:", + " PINNs and standard methods take about the same time" + ], + [ + "Question:According to the random sample of PINN articles, what is the typical runtime difference between PINN-based solvers and standard numerical methods?\nAnswer:", + " PINNs take hours to days, while standard methods take fractions of a second to minutes" + ], + [ + "Question:According to the random sample of PINN articles, what is the typical runtime difference between PINN-based solvers and standard numerical methods?\nAnswer:", + " PINNs are always unusable for PDEs" + ] + ], + "resps": [ + [ + [ + -23.198780059814453, + false + ] + ], + [ + [ + -23.7254581451416, + false + ] + ], + [ + [ + -40.41664505004883, + false + ] + ], + [ + [ + -35.51586151123047, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.198780059814453, + false + ], + [ + -23.7254581451416, + false + ], + [ + -40.41664505004883, + false + ], + [ + -35.51586151123047, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0781565615a4c1705bf91c1eb93c8ab0ddcb3fba972b5dd172d9240e3eff6815", + "prompt_hash": "ef3ffa2c12f24e51f8b0ad1128e0de7ebde64d0918297ca1ad76056e276d0fd0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 65, + "doc": { + "question": "What is the primary computational bottleneck in the Transformer architecture?", + "choices": [ + "The embedding layer", + "The attention mechanism", + "The feed-forward network", + "The positional encoding" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the primary computational bottleneck in the Transformer architecture?\nAnswer:", + " The embedding layer" + ], + [ + "Question:What is the primary computational bottleneck in the Transformer architecture?\nAnswer:", + " The attention mechanism" + ], + [ + "Question:What is the primary computational bottleneck in the Transformer architecture?\nAnswer:", + " The feed-forward network" + ], + [ + "Question:What is the primary computational bottleneck in the Transformer architecture?\nAnswer:", + " The positional encoding" + ] + ], + "resps": [ + [ + [ + -9.642364501953125, + false + ] + ], + [ + [ + -5.322152137756348, + false + ] + ], + [ + [ + -10.473891258239746, + false + ] + ], + [ + [ + -10.967617988586426, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.642364501953125, + false + ], + [ + -5.322152137756348, + false + ], + [ + -10.473891258239746, + false + ], + [ + -10.967617988586426, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7e3ba293ac28ebcb6bc67aa7c4670e4575db05560c9ad4da3828fec161c63b26", + "prompt_hash": "06fdde3cc2dae51beb96619eea06c16242ba38200b0bcea9af91338a8c5eb969", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 66, + "doc": { + "question": "Which hardware feature of the NVIDIA Hopper architecture allows asynchronous data transfer between GMEM and SMEM?", + "choices": [ + "Tensor Core Units", + "Warpgroup Synchronizer", + "Tensor Memory Accelerator (TMA)", + "Shared Memory Controller" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which hardware feature of the NVIDIA Hopper architecture allows asynchronous data transfer between GMEM and SMEM?\nAnswer:", + " Tensor Core Units" + ], + [ + "Question:Which hardware feature of the NVIDIA Hopper architecture allows asynchronous data transfer between GMEM and SMEM?\nAnswer:", + " Warpgroup Synchronizer" + ], + [ + "Question:Which hardware feature of the NVIDIA Hopper architecture allows asynchronous data transfer between GMEM and SMEM?\nAnswer:", + " Tensor Memory Accelerator (TMA)" + ], + [ + "Question:Which hardware feature of the NVIDIA Hopper architecture allows asynchronous data transfer between GMEM and SMEM?\nAnswer:", + " Shared Memory Controller" + ] + ], + "resps": [ + [ + [ + -19.239320755004883, + false + ] + ], + [ + [ + -22.14812469482422, + false + ] + ], + [ + [ + -17.409900665283203, + false + ] + ], + [ + [ + -12.405678749084473, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.239320755004883, + false + ], + [ + -22.14812469482422, + false + ], + [ + -17.409900665283203, + false + ], + [ + -12.405678749084473, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "52aacdf905d0e61171afc685cc7773cafc53632d7a01b4b11ccc86819e553951", + "prompt_hash": "d6a79f677f76cb627347b56319183e66e84f27fcddac5fc25e47ee201666e73f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 67, + "doc": { + "question": "In FP8 WGMMA, which data layout is required for input operands stored in shared memory?", + "choices": [ + "mn-major or k-major format", + "k-major format only", + "mn-major format only", + "Any layout is acceptable" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:In FP8 WGMMA, which data layout is required for input operands stored in shared memory?\nAnswer:", + " mn-major or k-major format" + ], + [ + "Question:In FP8 WGMMA, which data layout is required for input operands stored in shared memory?\nAnswer:", + " k-major format only" + ], + [ + "Question:In FP8 WGMMA, which data layout is required for input operands stored in shared memory?\nAnswer:", + " mn-major format only" + ], + [ + "Question:In FP8 WGMMA, which data layout is required for input operands stored in shared memory?\nAnswer:", + " Any layout is acceptable" + ] + ], + "resps": [ + [ + [ + -44.76097869873047, + false + ] + ], + [ + [ + -30.443592071533203, + false + ] + ], + [ + [ + -39.042701721191406, + false + ] + ], + [ + [ + -16.37194061279297, + false + ] + ] + ], + "filtered_resps": [ + [ + -44.76097869873047, + false + ], + [ + -30.443592071533203, + false + ], + [ + -39.042701721191406, + false + ], + [ + -16.37194061279297, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3df4c138fca5f6f0e388e4f554b918ab379e9b676b3a53b59f1c0365fd71a244", + "prompt_hash": "ce26067df1fa7fce65bb596a35a024471813ff953d7445adf72449a7edd5aaa7", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 68, + "doc": { + "question": "Which technique is used in FlashAttention-3 to reduce quantization error when working with outlier features in FP8?", + "choices": [ + "Per-tensor scaling", + "Block quantization and incoherent processing", + "Global normalization", + "Gradient clipping" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which technique is used in FlashAttention-3 to reduce quantization error when working with outlier features in FP8?\nAnswer:", + " Per-tensor scaling" + ], + [ + "Question:Which technique is used in FlashAttention-3 to reduce quantization error when working with outlier features in FP8?\nAnswer:", + " Block quantization and incoherent processing" + ], + [ + "Question:Which technique is used in FlashAttention-3 to reduce quantization error when working with outlier features in FP8?\nAnswer:", + " Global normalization" + ], + [ + "Question:Which technique is used in FlashAttention-3 to reduce quantization error when working with outlier features in FP8?\nAnswer:", + " Gradient clipping" + ] + ], + "resps": [ + [ + [ + -11.715656280517578, + false + ] + ], + [ + [ + -30.47322654724121, + false + ] + ], + [ + [ + -12.290731430053711, + false + ] + ], + [ + [ + -9.121477127075195, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.715656280517578, + false + ], + [ + -30.47322654724121, + false + ], + [ + -12.290731430053711, + false + ], + [ + -9.121477127075195, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "659b27d6a215f0d1bd0a042e0e9f87d82c34e968562612e66ab4102f92140f4e", + "prompt_hash": "c0392c6a289c354ed702f9046c58b831c42e59519727b65c3659310965b1cae2", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 69, + "doc": { + "question": "What is the reported forward pass speedup of FP16 FlashAttention-3 compared to FlashAttention-2?", + "choices": [ + "1.1-1.3×", + "1.5-2.0×", + "2.5-3.0×", + "3.5-4.0×" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the reported forward pass speedup of FP16 FlashAttention-3 compared to FlashAttention-2?\nAnswer:", + " 1.1-1.3×" + ], + [ + "Question:What is the reported forward pass speedup of FP16 FlashAttention-3 compared to FlashAttention-2?\nAnswer:", + " 1.5-2.0×" + ], + [ + "Question:What is the reported forward pass speedup of FP16 FlashAttention-3 compared to FlashAttention-2?\nAnswer:", + " 2.5-3.0×" + ], + [ + "Question:What is the reported forward pass speedup of FP16 FlashAttention-3 compared to FlashAttention-2?\nAnswer:", + " 3.5-4.0×" + ] + ], + "resps": [ + [ + [ + -18.02304458618164, + false + ] + ], + [ + [ + -16.681089401245117, + false + ] + ], + [ + [ + -17.597490310668945, + false + ] + ], + [ + [ + -18.239065170288086, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.02304458618164, + false + ], + [ + -16.681089401245117, + false + ], + [ + -17.597490310668945, + false + ], + [ + -18.239065170288086, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e17537da619793a1d35dcb4e674e82c52561bb7ccde7d0107b2dc1b5b43f5f34", + "prompt_hash": "ba403d403f84b992092e620a1958457ffee7449f194e2c948118cabd50d95bdf", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 70, + "doc": { + "question": "What is the total number of parameters of Qwen2-Audio?", + "choices": [ + "7.0B", + "8.2B", + "6.5B", + "9.1B" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the total number of parameters of Qwen2-Audio?\nAnswer:", + " 7.0B" + ], + [ + "Question:What is the total number of parameters of Qwen2-Audio?\nAnswer:", + " 8.2B" + ], + [ + "Question:What is the total number of parameters of Qwen2-Audio?\nAnswer:", + " 6.5B" + ], + [ + "Question:What is the total number of parameters of Qwen2-Audio?\nAnswer:", + " 9.1B" + ] + ], + "resps": [ + [ + [ + -12.461862564086914, + false + ] + ], + [ + [ + -11.596549987792969, + false + ] + ], + [ + [ + -11.168014526367188, + false + ] + ], + [ + [ + -12.721345901489258, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.461862564086914, + false + ], + [ + -11.596549987792969, + false + ], + [ + -11.168014526367188, + false + ], + [ + -12.721345901489258, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "36a5445979717a321cc191b2cdd44409d3c6cbc42665d07539fe017c4fece54c", + "prompt_hash": "8c8ae3b282fe9e0927e9272d4db4e1edbb49e97b0ce12de5197d8ca0751518ff", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 71, + "doc": { + "question": "Which model serves as the initialization for the audio encoder of Qwen2-Audio?", + "choices": [ + "Whisper-large-v3", + "Qwen-7B", + "SpeechVerse", + "Gemini-1.5-pro" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which model serves as the initialization for the audio encoder of Qwen2-Audio?\nAnswer:", + " Whisper-large-v3" + ], + [ + "Question:Which model serves as the initialization for the audio encoder of Qwen2-Audio?\nAnswer:", + " Qwen-7B" + ], + [ + "Question:Which model serves as the initialization for the audio encoder of Qwen2-Audio?\nAnswer:", + " SpeechVerse" + ], + [ + "Question:Which model serves as the initialization for the audio encoder of Qwen2-Audio?\nAnswer:", + " Gemini-1.5-pro" + ] + ], + "resps": [ + [ + [ + -16.041715621948242, + false + ] + ], + [ + [ + -15.517387390136719, + false + ] + ], + [ + [ + -23.270618438720703, + false + ] + ], + [ + [ + -16.757564544677734, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.041715621948242, + false + ], + [ + -15.517387390136719, + false + ], + [ + -23.270618438720703, + false + ], + [ + -16.757564544677734, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d4189a0ea2d13c666df4cfc81425a4981f563d9bbfbde9b32211dc1ab1f5b690", + "prompt_hash": "8cde8791295b635430528cc664bb33b82b4f4f77628c2a110026f29abd1b9917", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 72, + "doc": { + "question": "On the CoVoST2 dataset, what was Qwen2-Audio’s BLEU score for English-to-Chinese (en-zh) speech-to-text translation?", + "choices": [ + "33.1", + "41.5", + "45.2", + "24.4" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:On the CoVoST2 dataset, what was Qwen2-Audio’s BLEU score for English-to-Chinese (en-zh) speech-to-text translation?\nAnswer:", + " 33.1" + ], + [ + "Question:On the CoVoST2 dataset, what was Qwen2-Audio’s BLEU score for English-to-Chinese (en-zh) speech-to-text translation?\nAnswer:", + " 41.5" + ], + [ + "Question:On the CoVoST2 dataset, what was Qwen2-Audio’s BLEU score for English-to-Chinese (en-zh) speech-to-text translation?\nAnswer:", + " 45.2" + ], + [ + "Question:On the CoVoST2 dataset, what was Qwen2-Audio’s BLEU score for English-to-Chinese (en-zh) speech-to-text translation?\nAnswer:", + " 24.4" + ] + ], + "resps": [ + [ + [ + -9.055087089538574, + false + ] + ], + [ + [ + -9.025057792663574, + false + ] + ], + [ + [ + -9.17648696899414, + false + ] + ], + [ + [ + -8.53309154510498, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.055087089538574, + false + ], + [ + -9.025057792663574, + false + ], + [ + -9.17648696899414, + false + ], + [ + -8.53309154510498, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e255050272f4c1de3100e6a9e2c84d550470786f60c389f2b16edf1748416d92", + "prompt_hash": "d7110a0ee4e475f98bf211259b08ab7d2c5dfc613990469e965d97520b9882f0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 73, + "doc": { + "question": "In voice chat mode, how are users expected to interact with Qwen2-Audio?", + "choices": [ + "By issuing commands through structured tags only", + "By engaging in voice conversations with unrestricted dialogue", + "By providing only textual commands for offline analysis", + "By uploading audio files for automatic transcription" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:In voice chat mode, how are users expected to interact with Qwen2-Audio?\nAnswer:", + " By issuing commands through structured tags only" + ], + [ + "Question:In voice chat mode, how are users expected to interact with Qwen2-Audio?\nAnswer:", + " By engaging in voice conversations with unrestricted dialogue" + ], + [ + "Question:In voice chat mode, how are users expected to interact with Qwen2-Audio?\nAnswer:", + " By providing only textual commands for offline analysis" + ], + [ + "Question:In voice chat mode, how are users expected to interact with Qwen2-Audio?\nAnswer:", + " By uploading audio files for automatic transcription" + ] + ], + "resps": [ + [ + [ + -41.54837417602539, + false + ] + ], + [ + [ + -37.847511291503906, + false + ] + ], + [ + [ + -46.505123138427734, + false + ] + ], + [ + [ + -26.315853118896484, + false + ] + ] + ], + "filtered_resps": [ + [ + -41.54837417602539, + false + ], + [ + -37.847511291503906, + false + ], + [ + -46.505123138427734, + false + ], + [ + -26.315853118896484, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ea51e35a78949b693f0927adcdcbb4d74c2265f706f1dae0b0078b0df21494c6", + "prompt_hash": "7b6f1647be3f0fab5db5482826d1fadd78909d114faee55c3f711217f56dd859", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 74, + "doc": { + "question": "On the Librispeech test-clean dataset, what Word Error Rate (WER) did Qwen2-Audio achieve?", + "choices": [ + "2.0%", + "1.6%", + "3.4%", + "2.4%" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:On the Librispeech test-clean dataset, what Word Error Rate (WER) did Qwen2-Audio achieve?\nAnswer:", + " 2.0%" + ], + [ + "Question:On the Librispeech test-clean dataset, what Word Error Rate (WER) did Qwen2-Audio achieve?\nAnswer:", + " 1.6%" + ], + [ + "Question:On the Librispeech test-clean dataset, what Word Error Rate (WER) did Qwen2-Audio achieve?\nAnswer:", + " 3.4%" + ], + [ + "Question:On the Librispeech test-clean dataset, what Word Error Rate (WER) did Qwen2-Audio achieve?\nAnswer:", + " 2.4%" + ] + ], + "resps": [ + [ + [ + -14.759432792663574, + false + ] + ], + [ + [ + -14.206562042236328, + false + ] + ], + [ + [ + -14.24560546875, + false + ] + ], + [ + [ + -14.26672649383545, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.759432792663574, + false + ], + [ + -14.206562042236328, + false + ], + [ + -14.24560546875, + false + ], + [ + -14.26672649383545, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "da4aa04a709b7560efdf38c0d06ed4427fe1f08aa21e2f69fc46ce5c3a2c86ee", + "prompt_hash": "7c7c13464cd82ac0be19496dfcb8324f810a6fab804a26e7a03a1b7b13fd66f2", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 75, + "doc": { + "question": "What is the surrogate LLM used in the experiments?", + "choices": [ + "gpt-2", + "LLaMA3-70b", + "GPT3.5", + "BERT" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the surrogate LLM used in the experiments?\nAnswer:", + " gpt-2" + ], + [ + "Question:What is the surrogate LLM used in the experiments?\nAnswer:", + " LLaMA3-70b" + ], + [ + "Question:What is the surrogate LLM used in the experiments?\nAnswer:", + " GPT3.5" + ], + [ + "Question:What is the surrogate LLM used in the experiments?\nAnswer:", + " BERT" + ] + ], + "resps": [ + [ + [ + -8.955815315246582, + false + ] + ], + [ + [ + -11.751947402954102, + false + ] + ], + [ + [ + -6.5450286865234375, + false + ] + ], + [ + [ + -7.920544147491455, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.955815315246582, + false + ], + [ + -11.751947402954102, + false + ], + [ + -6.5450286865234375, + false + ], + [ + -7.920544147491455, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4a88247828c8f65da2cd023e803e88a1c6608c741bde2221dfd34031cdfd2f42", + "prompt_hash": "dc64d8747fe9a2cb0affa9df7dac972f5670662c4c675aec047337d950ba6669", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 76, + "doc": { + "question": "How many poisoned instances were injected for Agent-Driver in the experiments?", + "choices": [ + "20", + "10", + "4", + "2" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:How many poisoned instances were injected for Agent-Driver in the experiments?\nAnswer:", + " 20" + ], + [ + "Question:How many poisoned instances were injected for Agent-Driver in the experiments?\nAnswer:", + " 10" + ], + [ + "Question:How many poisoned instances were injected for Agent-Driver in the experiments?\nAnswer:", + " 4" + ], + [ + "Question:How many poisoned instances were injected for Agent-Driver in the experiments?\nAnswer:", + " 2" + ] + ], + "resps": [ + [ + [ + -4.250162124633789, + false + ] + ], + [ + [ + -3.488814353942871, + false + ] + ], + [ + [ + -3.920569896697998, + false + ] + ], + [ + [ + -3.1337132453918457, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.250162124633789, + false + ], + [ + -3.488814353942871, + false + ], + [ + -3.920569896697998, + false + ], + [ + -3.1337132453918457, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "20464913b9534ff408468a1f95f8958de14c3cd42204762ef42250e30d8d1651", + "prompt_hash": "bdcd08af0934c3af61aae0936cdf4fad05cbe1058f4b25e99c350f19b2f7d7dc", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 77, + "doc": { + "question": "Which retriever model is adopted as the contrastive retriever for ReAct-StrategyQA?", + "choices": [ + "DPR", + "REALM", + "BM25", + "ORQA" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which retriever model is adopted as the contrastive retriever for ReAct-StrategyQA?\nAnswer:", + " DPR" + ], + [ + "Question:Which retriever model is adopted as the contrastive retriever for ReAct-StrategyQA?\nAnswer:", + " REALM" + ], + [ + "Question:Which retriever model is adopted as the contrastive retriever for ReAct-StrategyQA?\nAnswer:", + " BM25" + ], + [ + "Question:Which retriever model is adopted as the contrastive retriever for ReAct-StrategyQA?\nAnswer:", + " ORQA" + ] + ], + "resps": [ + [ + [ + -7.363696098327637, + false + ] + ], + [ + [ + -10.3474760055542, + false + ] + ], + [ + [ + -5.98929500579834, + false + ] + ], + [ + [ + -15.886300086975098, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.363696098327637, + false + ], + [ + -10.3474760055542, + false + ], + [ + -5.98929500579834, + false + ], + [ + -15.886300086975098, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "19c0ce4a7df7efcbf4a36fbaf8f3bfff57ad8a7c82889a59823d14ecd7b9adbb", + "prompt_hash": "a30cd2e4ac0962c29654038025c04b00bca127955e10339d2bb80e7c8a8b36a2", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 78, + "doc": { + "question": "What is the attack target action for EHRAgent in the experiments?", + "choices": [ + "delete data of patient ID", + "sudden stop", + "I don't know", + "output random text" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the attack target action for EHRAgent in the experiments?\nAnswer:", + " delete data of patient ID" + ], + [ + "Question:What is the attack target action for EHRAgent in the experiments?\nAnswer:", + " sudden stop" + ], + [ + "Question:What is the attack target action for EHRAgent in the experiments?\nAnswer:", + " I don't know" + ], + [ + "Question:What is the attack target action for EHRAgent in the experiments?\nAnswer:", + " output random text" + ] + ], + "resps": [ + [ + [ + -23.995281219482422, + false + ] + ], + [ + [ + -16.279922485351562, + false + ] + ], + [ + [ + -12.055635452270508, + false + ] + ], + [ + [ + -20.101879119873047, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.995281219482422, + false + ], + [ + -16.279922485351562, + false + ], + [ + -12.055635452270508, + false + ], + [ + -20.101879119873047, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5d1ccd809307220bac91b010edf697bb9722ebf7ae3ab6259bf2cedb88970cb1", + "prompt_hash": "27a600525351da0238a0a909ca0511e3f9f89742b1be812a583a12b1fa3c8c77", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 79, + "doc": { + "question": "What type of poisoning strategy is used for Agent-Driver?", + "choices": [ + "spurious correlation", + "adversarial backdoor", + "jailbreak prompts", + "gradient masking" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What type of poisoning strategy is used for Agent-Driver?\nAnswer:", + " spurious correlation" + ], + [ + "Question:What type of poisoning strategy is used for Agent-Driver?\nAnswer:", + " adversarial backdoor" + ], + [ + "Question:What type of poisoning strategy is used for Agent-Driver?\nAnswer:", + " jailbreak prompts" + ], + [ + "Question:What type of poisoning strategy is used for Agent-Driver?\nAnswer:", + " gradient masking" + ] + ], + "resps": [ + [ + [ + -13.441458702087402, + false + ] + ], + [ + [ + -13.96750545501709, + false + ] + ], + [ + [ + -14.220865249633789, + false + ] + ], + [ + [ + -12.348712921142578, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.441458702087402, + false + ], + [ + -13.96750545501709, + false + ], + [ + -14.220865249633789, + false + ], + [ + -12.348712921142578, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d0c29abf8504cc8a487eec700cbc60e9c4a4c5875b48c80d52c13f41baf401f9", + "prompt_hash": "685d77301b39a898b2f0d0096f143aaf107f546574081bbae97b4bd9d7e3029b", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 80, + "doc": { + "question": "What is the core representation used to model dynamic scenes in the described method?", + "choices": [ + "Volumetric voxel grids", + "Canonical 3D Gaussians", + "Polygonal meshes", + "Point clouds with implicit surfaces" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the core representation used to model dynamic scenes in the described method?\nAnswer:", + " Volumetric voxel grids" + ], + [ + "Question:What is the core representation used to model dynamic scenes in the described method?\nAnswer:", + " Canonical 3D Gaussians" + ], + [ + "Question:What is the core representation used to model dynamic scenes in the described method?\nAnswer:", + " Polygonal meshes" + ], + [ + "Question:What is the core representation used to model dynamic scenes in the described method?\nAnswer:", + " Point clouds with implicit surfaces" + ] + ], + "resps": [ + [ + [ + -17.52205467224121, + false + ] + ], + [ + [ + -18.49441146850586, + false + ] + ], + [ + [ + -12.976879119873047, + false + ] + ], + [ + [ + -21.736690521240234, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.52205467224121, + false + ], + [ + -18.49441146850586, + false + ], + [ + -12.976879119873047, + false + ], + [ + -21.736690521240234, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4a71a8710b5b4843c89b5e6094a6155d44f27f5426dd622bfa594c5ed1e7ac15", + "prompt_hash": "2a45fd7aa1fbba28a6cd62c8d01844656671ea39f3b50e8c8d38853203338ad5", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 81, + "doc": { + "question": "Which dataset provides 14 sequences of 200-500 frames recorded using a handheld moving camera?", + "choices": [ + "KITTI", + "iPhone dataset", + "Kubric MOVi-F", + "TAP-Vid benchmark" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset provides 14 sequences of 200-500 frames recorded using a handheld moving camera?\nAnswer:", + " KITTI" + ], + [ + "Question:Which dataset provides 14 sequences of 200-500 frames recorded using a handheld moving camera?\nAnswer:", + " iPhone dataset" + ], + [ + "Question:Which dataset provides 14 sequences of 200-500 frames recorded using a handheld moving camera?\nAnswer:", + " Kubric MOVi-F" + ], + [ + "Question:Which dataset provides 14 sequences of 200-500 frames recorded using a handheld moving camera?\nAnswer:", + " TAP-Vid benchmark" + ] + ], + "resps": [ + [ + [ + -7.701709747314453, + false + ] + ], + [ + [ + -14.98783016204834, + false + ] + ], + [ + [ + -36.20477294921875, + false + ] + ], + [ + [ + -25.44028663635254, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.701709747314453, + false + ], + [ + -14.98783016204834, + false + ], + [ + -36.20477294921875, + false + ], + [ + -25.44028663635254, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0556d92c2af3244fbc95362eba3d1d35ddc141fbb94f3200ca54e53f30798f02", + "prompt_hash": "daf1ca27eafd86ec618f917cdd82261a07bb49b6fcc6093bfa3a1df9aa0ed739", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 82, + "doc": { + "question": "How many SE(3) motion bases are used in the experiments?", + "choices": [ + "10", + "15", + "20", + "25" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many SE(3) motion bases are used in the experiments?\nAnswer:", + " 10" + ], + [ + "Question:How many SE(3) motion bases are used in the experiments?\nAnswer:", + " 15" + ], + [ + "Question:How many SE(3) motion bases are used in the experiments?\nAnswer:", + " 20" + ], + [ + "Question:How many SE(3) motion bases are used in the experiments?\nAnswer:", + " 25" + ] + ], + "resps": [ + [ + [ + -4.920566082000732, + false + ] + ], + [ + [ + -6.513179302215576, + false + ] + ], + [ + [ + -5.890079021453857, + false + ] + ], + [ + [ + -7.041357517242432, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.920566082000732, + false + ], + [ + -6.513179302215576, + false + ], + [ + -5.890079021453857, + false + ], + [ + -7.041357517242432, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "410a101894c872b47e15db365f4cf6dee0284104c5d11a6e94f06759bca584ac", + "prompt_hash": "003ef24bc5a450560bbc48db84d2484a00eaf229698184c67258246b368be939", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 83, + "doc": { + "question": "Which optimizer is used to optimize the model parameters?", + "choices": [ + "SGD with momentum", + "Adam Optimizer", + "RMSProp", + "AdaGrad" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which optimizer is used to optimize the model parameters?\nAnswer:", + " SGD with momentum" + ], + [ + "Question:Which optimizer is used to optimize the model parameters?\nAnswer:", + " Adam Optimizer" + ], + [ + "Question:Which optimizer is used to optimize the model parameters?\nAnswer:", + " RMSProp" + ], + [ + "Question:Which optimizer is used to optimize the model parameters?\nAnswer:", + " AdaGrad" + ] + ], + "resps": [ + [ + [ + -12.31170654296875, + false + ] + ], + [ + [ + -5.3426055908203125, + false + ] + ], + [ + [ + -7.934563636779785, + false + ] + ], + [ + [ + -9.722707748413086, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.31170654296875, + false + ], + [ + -5.3426055908203125, + false + ], + [ + -7.934563636779785, + false + ], + [ + -9.722707748413086, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "92bffbd3914d4baa00750110dfc81d6a3872c735eb4ad90041cadcdad0acb517", + "prompt_hash": "57f4207bf268aed9b8544b57c46fce191db3a2e7fcf6f2909282f3f48f230172", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 84, + "doc": { + "question": "What frame resolution and GPU are mentioned for training a 300-frame sequence?", + "choices": [ + "720p on RTX 3090", + "1080p on Tesla V100", + "960×720 on A100", + "640×480 on T4" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What frame resolution and GPU are mentioned for training a 300-frame sequence?\nAnswer:", + " 720p on RTX 3090" + ], + [ + "Question:What frame resolution and GPU are mentioned for training a 300-frame sequence?\nAnswer:", + " 1080p on Tesla V100" + ], + [ + "Question:What frame resolution and GPU are mentioned for training a 300-frame sequence?\nAnswer:", + " 960×720 on A100" + ], + [ + "Question:What frame resolution and GPU are mentioned for training a 300-frame sequence?\nAnswer:", + " 640×480 on T4" + ] + ], + "resps": [ + [ + [ + -18.086414337158203, + false + ] + ], + [ + [ + -19.665401458740234, + false + ] + ], + [ + [ + -20.820384979248047, + false + ] + ], + [ + [ + -22.602657318115234, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.086414337158203, + false + ], + [ + -19.665401458740234, + false + ], + [ + -20.820384979248047, + false + ], + [ + -22.602657318115234, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c3733ea2e70f8366500e2478d3ace4710fcc86588a565cfcc3b92b76e836e154", + "prompt_hash": "b0d8157bc04fad12b6a8e0e7582c6e21007152788beeec7dcdaaacab10e65098", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 85, + "doc": { + "question": "What is the total number of multiple-choice questions included in the benchmark?", + "choices": [ + "5,412", + "6,678", + "7,235", + "4,950" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the total number of multiple-choice questions included in the benchmark?\nAnswer:", + " 5,412" + ], + [ + "Question:What is the total number of multiple-choice questions included in the benchmark?\nAnswer:", + " 6,678" + ], + [ + "Question:What is the total number of multiple-choice questions included in the benchmark?\nAnswer:", + " 7,235" + ], + [ + "Question:What is the total number of multiple-choice questions included in the benchmark?\nAnswer:", + " 4,950" + ] + ], + "resps": [ + [ + [ + -16.877378463745117, + false + ] + ], + [ + [ + -16.803268432617188, + false + ] + ], + [ + [ + -17.370201110839844, + false + ] + ], + [ + [ + -15.794563293457031, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.877378463745117, + false + ], + [ + -16.803268432617188, + false + ], + [ + -17.370201110839844, + false + ], + [ + -15.794563293457031, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9eefefe26ce7dd38d28a8386f844287bf0dda24eba0aedec041f58140a26cbe1", + "prompt_hash": "3b8576c5b0841afdce8201794aafedd83491500189ebb11a38c439c1477b1e62", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 86, + "doc": { + "question": "Which task was introduced to address the issue of single-frame bias in video understanding metrics?", + "choices": [ + "Sequential Tracking", + "Referring Reasoning", + "Contextual Retrieval", + "Visual Summarization" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which task was introduced to address the issue of single-frame bias in video understanding metrics?\nAnswer:", + " Sequential Tracking" + ], + [ + "Question:Which task was introduced to address the issue of single-frame bias in video understanding metrics?\nAnswer:", + " Referring Reasoning" + ], + [ + "Question:Which task was introduced to address the issue of single-frame bias in video understanding metrics?\nAnswer:", + " Contextual Retrieval" + ], + [ + "Question:Which task was introduced to address the issue of single-frame bias in video understanding metrics?\nAnswer:", + " Visual Summarization" + ] + ], + "resps": [ + [ + [ + -20.925914764404297, + false + ] + ], + [ + [ + -22.508527755737305, + false + ] + ], + [ + [ + -18.34051513671875, + false + ] + ], + [ + [ + -18.849546432495117, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.925914764404297, + false + ], + [ + -22.508527755737305, + false + ], + [ + -18.34051513671875, + false + ], + [ + -18.849546432495117, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0a74c55939e0ee0aef7576105a07b339281c8eebc617f9de10a2ddfc75391010", + "prompt_hash": "8aca42956d2d1211359e4faefbe33be74f7456a7be81b279e4d0a92537f6b520", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 87, + "doc": { + "question": "How many progressive duration groups of videos are included in the benchmark?", + "choices": [ + "Three", + "Four", + "Five", + "Six" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many progressive duration groups of videos are included in the benchmark?\nAnswer:", + " Three" + ], + [ + "Question:How many progressive duration groups of videos are included in the benchmark?\nAnswer:", + " Four" + ], + [ + "Question:How many progressive duration groups of videos are included in the benchmark?\nAnswer:", + " Five" + ], + [ + "Question:How many progressive duration groups of videos are included in the benchmark?\nAnswer:", + " Six" + ] + ], + "resps": [ + [ + [ + -5.014612197875977, + false + ] + ], + [ + [ + -5.665154933929443, + false + ] + ], + [ + [ + -5.873777389526367, + false + ] + ], + [ + [ + -6.582821846008301, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.014612197875977, + false + ], + [ + -5.665154933929443, + false + ], + [ + -5.873777389526367, + false + ], + [ + -6.582821846008301, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4ddb076c5236ac1d85766f7f715661e83a2f723b33a91965d1114ece439c600e", + "prompt_hash": "75c18a03dceda279b144d6687ea54a3c5c0de57431fa510734136918a426ca79", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 88, + "doc": { + "question": "Which open-source model faced severe accuracy degradation when processing 64 input frames?", + "choices": [ + "Idefics2", + "GPT-4o", + "Gemini-1.5-Pro", + "Phi-3-Vision-Instruct" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which open-source model faced severe accuracy degradation when processing 64 input frames?\nAnswer:", + " Idefics2" + ], + [ + "Question:Which open-source model faced severe accuracy degradation when processing 64 input frames?\nAnswer:", + " GPT-4o" + ], + [ + "Question:Which open-source model faced severe accuracy degradation when processing 64 input frames?\nAnswer:", + " Gemini-1.5-Pro" + ], + [ + "Question:Which open-source model faced severe accuracy degradation when processing 64 input frames?\nAnswer:", + " Phi-3-Vision-Instruct" + ] + ], + "resps": [ + [ + [ + -14.050100326538086, + false + ] + ], + [ + [ + -7.555898189544678, + false + ] + ], + [ + [ + -10.493349075317383, + false + ] + ], + [ + [ + -18.406396865844727, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.050100326538086, + false + ], + [ + -7.555898189544678, + false + ], + [ + -10.493349075317383, + false + ], + [ + -18.406396865844727, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "72ecfa49ab93693acca940d13fad99666b07e54dae8fbeee5b8695b848f68432", + "prompt_hash": "a319e9c070d1ffd3109994fa4594e8e66c9c932cf2ae4db25fe0b41536899ab4", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 89, + "doc": { + "question": "What is the average length of a question in the benchmark?", + "choices": [ + "27.15 words", + "35.82 words", + "43.53 words", + "50.20 words" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the average length of a question in the benchmark?\nAnswer:", + " 27.15 words" + ], + [ + "Question:What is the average length of a question in the benchmark?\nAnswer:", + " 35.82 words" + ], + [ + "Question:What is the average length of a question in the benchmark?\nAnswer:", + " 43.53 words" + ], + [ + "Question:What is the average length of a question in the benchmark?\nAnswer:", + " 50.20 words" + ] + ], + "resps": [ + [ + [ + -15.660734176635742, + false + ] + ], + [ + [ + -15.121524810791016, + false + ] + ], + [ + [ + -15.357699394226074, + false + ] + ], + [ + [ + -17.811992645263672, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.660734176635742, + false + ], + [ + -15.121524810791016, + false + ], + [ + -15.357699394226074, + false + ], + [ + -17.811992645263672, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f6ea5f4b873d21defd6d4012ffe39fc0fe8c93d275d7f6bbd55ca96239576e3a", + "prompt_hash": "085e2913fade8b060f9c9f02a85206298f94060133217cab224c91b002010bfc", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 90, + "doc": { + "question": "What does the RISE algorithm stand for?", + "choices": [ + "Recursive Introspection for Self-Improvement", + "Reinforced Inference for Sequential Evaluation", + "Reward-based Incremental Sampling Enhancement", + "Regression Induced System Evaluation" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What does the RISE algorithm stand for?\nAnswer:", + " Recursive Introspection for Self-Improvement" + ], + [ + "Question:What does the RISE algorithm stand for?\nAnswer:", + " Reinforced Inference for Sequential Evaluation" + ], + [ + "Question:What does the RISE algorithm stand for?\nAnswer:", + " Reward-based Incremental Sampling Enhancement" + ], + [ + "Question:What does the RISE algorithm stand for?\nAnswer:", + " Regression Induced System Evaluation" + ] + ], + "resps": [ + [ + [ + -18.053543090820312, + false + ] + ], + [ + [ + -22.69965362548828, + false + ] + ], + [ + [ + -24.5802059173584, + false + ] + ], + [ + [ + -22.592044830322266, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.053543090820312, + false + ], + [ + -22.69965362548828, + false + ], + [ + -24.5802059173584, + false + ], + [ + -22.592044830322266, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "447cedc70ee958e43686d7c7a4918f12ce062181a998804a729c7fd95aad7a1b", + "prompt_hash": "56a91389443b53bba1b14ecaa15c800a07be68f9373e6a155dc743fcc90e78d1", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 91, + "doc": { + "question": "Which dataset showed a 23.9% improvement for Mistral-7B when using RISE over five turns?", + "choices": [ + "SVAMP", + "GSM8K", + "MATH", + "ARC" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset showed a 23.9% improvement for Mistral-7B when using RISE over five turns?\nAnswer:", + " SVAMP" + ], + [ + "Question:Which dataset showed a 23.9% improvement for Mistral-7B when using RISE over five turns?\nAnswer:", + " GSM8K" + ], + [ + "Question:Which dataset showed a 23.9% improvement for Mistral-7B when using RISE over five turns?\nAnswer:", + " MATH" + ], + [ + "Question:Which dataset showed a 23.9% improvement for Mistral-7B when using RISE over five turns?\nAnswer:", + " ARC" + ] + ], + "resps": [ + [ + [ + -12.474897384643555, + false + ] + ], + [ + [ + -6.9842848777771, + false + ] + ], + [ + [ + -8.997828483581543, + false + ] + ], + [ + [ + -8.794258117675781, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.474897384643555, + false + ], + [ + -6.9842848777771, + false + ], + [ + -8.997828483581543, + false + ], + [ + -8.794258117675781, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "68b8dc73dcb9f561ec8b17b5ec220eb99c6f176dc768fce36d551a221af66114", + "prompt_hash": "47c072b1d8d8e66607250531175d735cd6aadaaeb74c1a9e232137cdcc37b57f", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 92, + "doc": { + "question": "What is the purpose of the 'knowledge boosting' stage in RISE training?", + "choices": [ + "To improve model performance on out-of-distribution data", + "To pre-train the model on diverse internet-scale corpora", + "To teach response style and instruction-following before on-policy rollouts", + "To reduce computational costs during reinforcement learning" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the purpose of the 'knowledge boosting' stage in RISE training?\nAnswer:", + " To improve model performance on out-of-distribution data" + ], + [ + "Question:What is the purpose of the 'knowledge boosting' stage in RISE training?\nAnswer:", + " To pre-train the model on diverse internet-scale corpora" + ], + [ + "Question:What is the purpose of the 'knowledge boosting' stage in RISE training?\nAnswer:", + " To teach response style and instruction-following before on-policy rollouts" + ], + [ + "Question:What is the purpose of the 'knowledge boosting' stage in RISE training?\nAnswer:", + " To reduce computational costs during reinforcement learning" + ] + ], + "resps": [ + [ + [ + -19.200429916381836, + false + ] + ], + [ + [ + -32.36049270629883, + false + ] + ], + [ + [ + -58.351783752441406, + false + ] + ], + [ + [ + -25.660242080688477, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.200429916381836, + false + ], + [ + -32.36049270629883, + false + ], + [ + -58.351783752441406, + false + ], + [ + -25.660242080688477, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "691fc113bfe9da94b48453014c9d0e8e30f5ad527803e97f11a8a3d27eca1efd", + "prompt_hash": "e53d5e7ebff513231466da7a62f31d10ec7f1751bb32499edc7d8f04afc51efb", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 93, + "doc": { + "question": "How does RISE decide the final response when running 'without oracle' mode at inference time?", + "choices": [ + "It selects the response with the highest confidence score", + "It uses majority voting over candidate outputs from different turns", + "It always chooses the response from the last turn", + "It queries a stronger external model for validation" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How does RISE decide the final response when running 'without oracle' mode at inference time?\nAnswer:", + " It selects the response with the highest confidence score" + ], + [ + "Question:How does RISE decide the final response when running 'without oracle' mode at inference time?\nAnswer:", + " It uses majority voting over candidate outputs from different turns" + ], + [ + "Question:How does RISE decide the final response when running 'without oracle' mode at inference time?\nAnswer:", + " It always chooses the response from the last turn" + ], + [ + "Question:How does RISE decide the final response when running 'without oracle' mode at inference time?\nAnswer:", + " It queries a stronger external model for validation" + ] + ], + "resps": [ + [ + [ + -14.699235916137695, + false + ] + ], + [ + [ + -37.79487609863281, + false + ] + ], + [ + [ + -23.1787109375, + false + ] + ], + [ + [ + -37.22762680053711, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.699235916137695, + false + ], + [ + -37.79487609863281, + false + ], + [ + -23.1787109375, + false + ], + [ + -37.22762680053711, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "49446d10ea114fe85f650e62cbe696e3ea6acf8592931f6f15b48a6e90ac48d8", + "prompt_hash": "189a98bbc8bce5e6af01dfbae399f1fe09d20c3d95d2bb4c3740f2334aa23892", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 94, + "doc": { + "question": "Which method did RISE outperform by solving problems unsolved by larger budgets at the first turn?", + "choices": [ + "Self-Refine", + "GLoRE", + "Parallel pass@B sampling", + "Chain-of-Thought prompting" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which method did RISE outperform by solving problems unsolved by larger budgets at the first turn?\nAnswer:", + " Self-Refine" + ], + [ + "Question:Which method did RISE outperform by solving problems unsolved by larger budgets at the first turn?\nAnswer:", + " GLoRE" + ], + [ + "Question:Which method did RISE outperform by solving problems unsolved by larger budgets at the first turn?\nAnswer:", + " Parallel pass@B sampling" + ], + [ + "Question:Which method did RISE outperform by solving problems unsolved by larger budgets at the first turn?\nAnswer:", + " Chain-of-Thought prompting" + ] + ], + "resps": [ + [ + [ + -14.033670425415039, + false + ] + ], + [ + [ + -17.320749282836914, + false + ] + ], + [ + [ + -44.752960205078125, + false + ] + ], + [ + [ + -16.29299545288086, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.033670425415039, + false + ], + [ + -17.320749282836914, + false + ], + [ + -44.752960205078125, + false + ], + [ + -16.29299545288086, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5be7b9aac29cff178ab3bfa0c1f27fedbcbdf9a6e094059fa58fed8b6f0d8b3c", + "prompt_hash": "ee21eb2bda8b580daea4f8db981b89e8aca39a5c8ffdf31bf5c5c81239103415", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 95, + "doc": { + "question": "What model is used as the initial seed model in the experiments?", + "choices": [ + "Llama-3-8B-Instruct", + "Llama-2-70B-Chat", + "GPT-4-1106-preview", + "Starling-RM-34B" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What model is used as the initial seed model in the experiments?\nAnswer:", + " Llama-3-8B-Instruct" + ], + [ + "Question:What model is used as the initial seed model in the experiments?\nAnswer:", + " Llama-2-70B-Chat" + ], + [ + "Question:What model is used as the initial seed model in the experiments?\nAnswer:", + " GPT-4-1106-preview" + ], + [ + "Question:What model is used as the initial seed model in the experiments?\nAnswer:", + " Starling-RM-34B" + ] + ], + "resps": [ + [ + [ + -12.638608932495117, + false + ] + ], + [ + [ + -15.421242713928223, + false + ] + ], + [ + [ + -9.687981605529785, + false + ] + ], + [ + [ + -31.80432891845703, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.638608932495117, + false + ], + [ + -15.421242713928223, + false + ], + [ + -9.687981605529785, + false + ], + [ + -31.80432891845703, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "decef1964b2e2a5dd8554539fdf609f75ca52d60a5b511362d93b3606e83d977", + "prompt_hash": "5def577581854c72012d82f7d550508d28c90b62f0d3d9e78ebcc1de138bea1c", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 96, + "doc": { + "question": "How many response variations are generated per prompt in each iteration during training?", + "choices": [ + "5", + "7", + "9", + "11" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many response variations are generated per prompt in each iteration during training?\nAnswer:", + " 5" + ], + [ + "Question:How many response variations are generated per prompt in each iteration during training?\nAnswer:", + " 7" + ], + [ + "Question:How many response variations are generated per prompt in each iteration during training?\nAnswer:", + " 9" + ], + [ + "Question:How many response variations are generated per prompt in each iteration during training?\nAnswer:", + " 11" + ] + ], + "resps": [ + [ + [ + -4.11230993270874, + false + ] + ], + [ + [ + -6.176117420196533, + false + ] + ], + [ + [ + -6.593822956085205, + false + ] + ], + [ + [ + -7.345574378967285, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.11230993270874, + false + ], + [ + -6.176117420196533, + false + ], + [ + -6.593822956085205, + false + ], + [ + -7.345574378967285, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5938e09b994ec80f367a369388816c68f9946c9c16ee409c43cd4c90f9b6da07", + "prompt_hash": "863a5b0d309f6953bac0812d35bcb80f2c0def65dfe1a0df4ace5bdc9f2e5b7f", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 97, + "doc": { + "question": "Which benchmark evaluates the model's ability in multi-turn conversations?", + "choices": [ + "AlpacaEval 2", + "Arena-Hard", + "MT-Bench", + "Open Assistant" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which benchmark evaluates the model's ability in multi-turn conversations?\nAnswer:", + " AlpacaEval 2" + ], + [ + "Question:Which benchmark evaluates the model's ability in multi-turn conversations?\nAnswer:", + " Arena-Hard" + ], + [ + "Question:Which benchmark evaluates the model's ability in multi-turn conversations?\nAnswer:", + " MT-Bench" + ], + [ + "Question:Which benchmark evaluates the model's ability in multi-turn conversations?\nAnswer:", + " Open Assistant" + ] + ], + "resps": [ + [ + [ + -12.953899383544922, + false + ] + ], + [ + [ + -10.193154335021973, + false + ] + ], + [ + [ + -5.521928787231445, + false + ] + ], + [ + [ + -9.941776275634766, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.953899383544922, + false + ], + [ + -10.193154335021973, + false + ], + [ + -5.521928787231445, + false + ], + [ + -9.941776275634766, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1d0b2ec747493415aa8998fb6c5d640e86cd06eaededa404995f36837b0af203", + "prompt_hash": "fb8092185a1e5cd834f0e4266ceac170583b388675f40a41987e3bbdc5577481", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 98, + "doc": { + "question": "What was the length-controlled (LC) win rate achieved on AlpacaEval 2 after the fourth iteration of training?", + "choices": [ + "22.92%", + "35.49%", + "39.44%", + "41.20%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What was the length-controlled (LC) win rate achieved on AlpacaEval 2 after the fourth iteration of training?\nAnswer:", + " 22.92%" + ], + [ + "Question:What was the length-controlled (LC) win rate achieved on AlpacaEval 2 after the fourth iteration of training?\nAnswer:", + " 35.49%" + ], + [ + "Question:What was the length-controlled (LC) win rate achieved on AlpacaEval 2 after the fourth iteration of training?\nAnswer:", + " 39.44%" + ], + [ + "Question:What was the length-controlled (LC) win rate achieved on AlpacaEval 2 after the fourth iteration of training?\nAnswer:", + " 41.20%" + ] + ], + "resps": [ + [ + [ + -16.45294189453125, + false + ] + ], + [ + [ + -16.92536163330078, + false + ] + ], + [ + [ + -16.5533390045166, + false + ] + ], + [ + [ + -16.253902435302734, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.45294189453125, + false + ], + [ + -16.92536163330078, + false + ], + [ + -16.5533390045166, + false + ], + [ + -16.253902435302734, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7ebd37c8db4919b509d761fba7404258b206bf0ef716354b999a7cac85aef19b", + "prompt_hash": "55db4484c93ad2de95b643524b062dfe42f48efd8075611645e98ddb832df074", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 99, + "doc": { + "question": "Which strong external reward model failed to improve the LC win rate in experiments?", + "choices": [ + "Starling-RM-34B", + "Claude-Opus", + "SPPO", + "GPT-4-0314" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which strong external reward model failed to improve the LC win rate in experiments?\nAnswer:", + " Starling-RM-34B" + ], + [ + "Question:Which strong external reward model failed to improve the LC win rate in experiments?\nAnswer:", + " Claude-Opus" + ], + [ + "Question:Which strong external reward model failed to improve the LC win rate in experiments?\nAnswer:", + " SPPO" + ], + [ + "Question:Which strong external reward model failed to improve the LC win rate in experiments?\nAnswer:", + " GPT-4-0314" + ] + ], + "resps": [ + [ + [ + -33.60536193847656, + false + ] + ], + [ + [ + -20.415925979614258, + false + ] + ], + [ + [ + -15.506284713745117, + false + ] + ], + [ + [ + -19.35174560546875, + false + ] + ] + ], + "filtered_resps": [ + [ + -33.60536193847656, + false + ], + [ + -20.415925979614258, + false + ], + [ + -15.506284713745117, + false + ], + [ + -19.35174560546875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "01506c30c1ffb5f3d34526184f0caaf6b0ac57ae5d3ef1294a5db9c7ac44b382", + "prompt_hash": "4a6107d70777877591a3f8886e2ef6bbc0d0830fa0869dee0cfdb68d9cf22027", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + } + ], + "arxiv_mc_2408": [ + { + "doc_id": 0, + "doc": { + "question": "What activation function do JumpReLU SAEs use in place of the ReLU function?", + "choices": [ + "Softmax", + "TopK", + "JumpReLU", + "Sigmoid" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What activation function do JumpReLU SAEs use in place of the ReLU function?\nAnswer:", + " Softmax" + ], + [ + "Question:What activation function do JumpReLU SAEs use in place of the ReLU function?\nAnswer:", + " TopK" + ], + [ + "Question:What activation function do JumpReLU SAEs use in place of the ReLU function?\nAnswer:", + " JumpReLU" + ], + [ + "Question:What activation function do JumpReLU SAEs use in place of the ReLU function?\nAnswer:", + " Sigmoid" + ] + ], + "resps": [ + [ + [ + -7.370088577270508, + false + ] + ], + [ + [ + -15.097153663635254, + false + ] + ], + [ + [ + -10.174015998840332, + false + ] + ], + [ + [ + -2.91923189163208, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.370088577270508, + false + ], + [ + -15.097153663635254, + false + ], + [ + -10.174015998840332, + false + ], + [ + -2.91923189163208, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9fc664838b502e0969a0d397b3807dd078d361efa07bb02fce0d3b229d6d70eb", + "prompt_hash": "b541b7d775a03e2f5dffc078cc247a2493ab42c838ac3de8f3625b1bbb5734a5", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 1, + "doc": { + "question": "Which loss function do JumpReLU SAEs use for enforcing sparsity?", + "choices": [ + "L1 norm penalty", + "L2 norm penalty", + "KL divergence", + "L0 sparsity penalty" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which loss function do JumpReLU SAEs use for enforcing sparsity?\nAnswer:", + " L1 norm penalty" + ], + [ + "Question:Which loss function do JumpReLU SAEs use for enforcing sparsity?\nAnswer:", + " L2 norm penalty" + ], + [ + "Question:Which loss function do JumpReLU SAEs use for enforcing sparsity?\nAnswer:", + " KL divergence" + ], + [ + "Question:Which loss function do JumpReLU SAEs use for enforcing sparsity?\nAnswer:", + " L0 sparsity penalty" + ] + ], + "resps": [ + [ + [ + -16.749191284179688, + false + ] + ], + [ + [ + -16.98282241821289, + false + ] + ], + [ + [ + -9.006500244140625, + false + ] + ], + [ + [ + -21.58685302734375, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.749191284179688, + false + ], + [ + -16.98282241821289, + false + ], + [ + -9.006500244140625, + false + ], + [ + -21.58685302734375, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3083d36e417a35b7fa92a09da8aaf3ab53bba600d0297616c7530c16aab83764", + "prompt_hash": "d44b0a6d22d713fabb047ea2f9974b6dd36525eea409ff85bd6052fb23fb085c", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 2, + "doc": { + "question": "What is the main role of the straight-through estimator (STE) in training JumpReLU SAEs?", + "choices": [ + "To estimate the TopK feature indices", + "To avoid dead neurons during initialization", + "To enable training through the threshold parameter", + "To calculate decoder weights directly" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the main role of the straight-through estimator (STE) in training JumpReLU SAEs?\nAnswer:", + " To estimate the TopK feature indices" + ], + [ + "Question:What is the main role of the straight-through estimator (STE) in training JumpReLU SAEs?\nAnswer:", + " To avoid dead neurons during initialization" + ], + [ + "Question:What is the main role of the straight-through estimator (STE) in training JumpReLU SAEs?\nAnswer:", + " To enable training through the threshold parameter" + ], + [ + "Question:What is the main role of the straight-through estimator (STE) in training JumpReLU SAEs?\nAnswer:", + " To calculate decoder weights directly" + ] + ], + "resps": [ + [ + [ + -32.831172943115234, + false + ] + ], + [ + [ + -29.019672393798828, + false + ] + ], + [ + [ + -32.593955993652344, + false + ] + ], + [ + [ + -31.319580078125, + false + ] + ] + ], + "filtered_resps": [ + [ + -32.831172943115234, + false + ], + [ + -29.019672393798828, + false + ], + [ + -32.593955993652344, + false + ], + [ + -31.319580078125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6e057a450980431afda4f7c99a40bae2259ebd40c01007a7d3c605ba1eae3bbb", + "prompt_hash": "3e38d72fc777f347897440eac6491904bfa9a74477d3cd6b0c0ef9488461822f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 3, + "doc": { + "question": "Compared to Gated and TopK SAEs, which of the following is an advantage of JumpReLU SAEs?", + "choices": [ + "They use more auxiliary loss terms", + "They require partial sorting of activations", + "They are more efficient to train", + "They rely on dead feature resampling" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Compared to Gated and TopK SAEs, which of the following is an advantage of JumpReLU SAEs?\nAnswer:", + " They use more auxiliary loss terms" + ], + [ + "Question:Compared to Gated and TopK SAEs, which of the following is an advantage of JumpReLU SAEs?\nAnswer:", + " They require partial sorting of activations" + ], + [ + "Question:Compared to Gated and TopK SAEs, which of the following is an advantage of JumpReLU SAEs?\nAnswer:", + " They are more efficient to train" + ], + [ + "Question:Compared to Gated and TopK SAEs, which of the following is an advantage of JumpReLU SAEs?\nAnswer:", + " They rely on dead feature resampling" + ] + ], + "resps": [ + [ + [ + -32.257041931152344, + false + ] + ], + [ + [ + -35.025230407714844, + false + ] + ], + [ + [ + -19.65345573425293, + false + ] + ], + [ + [ + -45.214271545410156, + false + ] + ] + ], + "filtered_resps": [ + [ + -32.257041931152344, + false + ], + [ + -35.025230407714844, + false + ], + [ + -19.65345573425293, + false + ], + [ + -45.214271545410156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f27eed3312662f0cba14084f910dcc19898dade925b92cb182c54756586dc699", + "prompt_hash": "73c0c75bccfd2179a85934c79c1e4df44fb62f7d04764903f2332e11bdc742c1", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 4, + "doc": { + "question": "What is the function of the Heaviside step function in JumpReLU SAEs?", + "choices": [ + "To smooth gradients for backpropagation", + "To normalize decoder weights", + "To zero out activations below a threshold", + "To increase model capacity" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the function of the Heaviside step function in JumpReLU SAEs?\nAnswer:", + " To smooth gradients for backpropagation" + ], + [ + "Question:What is the function of the Heaviside step function in JumpReLU SAEs?\nAnswer:", + " To normalize decoder weights" + ], + [ + "Question:What is the function of the Heaviside step function in JumpReLU SAEs?\nAnswer:", + " To zero out activations below a threshold" + ], + [ + "Question:What is the function of the Heaviside step function in JumpReLU SAEs?\nAnswer:", + " To increase model capacity" + ] + ], + "resps": [ + [ + [ + -25.059850692749023, + false + ] + ], + [ + [ + -26.03301239013672, + false + ] + ], + [ + [ + -21.755271911621094, + false + ] + ], + [ + [ + -18.161693572998047, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.059850692749023, + false + ], + [ + -26.03301239013672, + false + ], + [ + -21.755271911621094, + false + ], + [ + -18.161693572998047, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "95fba6f112c51a73314c20365df8a022603cc6c4cbc821783a1f2033bdd519ce", + "prompt_hash": "8792b4fc53b183725631029120e0b33a31b36460f7a4876f205e8b0ddef5e9b0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 5, + "doc": { + "question": "What is the main architectural difference between KAN and traditional MLP?", + "choices": [ + "KAN uses fixed activation functions on nodes, while MLP uses learnable splines on edges", + "KAN uses learnable activation functions on edges, while MLP uses fixed activation functions on nodes", + "KAN employs convolutional layers instead of fully connected layers", + "KAN is trained using unsupervised learning while MLP is trained using supervised learning" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main architectural difference between KAN and traditional MLP?\nAnswer:", + " KAN uses fixed activation functions on nodes, while MLP uses learnable splines on edges" + ], + [ + "Question:What is the main architectural difference between KAN and traditional MLP?\nAnswer:", + " KAN uses learnable activation functions on edges, while MLP uses fixed activation functions on nodes" + ], + [ + "Question:What is the main architectural difference between KAN and traditional MLP?\nAnswer:", + " KAN employs convolutional layers instead of fully connected layers" + ], + [ + "Question:What is the main architectural difference between KAN and traditional MLP?\nAnswer:", + " KAN is trained using unsupervised learning while MLP is trained using supervised learning" + ] + ], + "resps": [ + [ + [ + -46.42828369140625, + false + ] + ], + [ + [ + -34.54854202270508, + false + ] + ], + [ + [ + -14.264701843261719, + false + ] + ], + [ + [ + -24.592636108398438, + false + ] + ] + ], + "filtered_resps": [ + [ + -46.42828369140625, + false + ], + [ + -34.54854202270508, + false + ], + [ + -14.264701843261719, + false + ], + [ + -24.592636108398438, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "911fd11e4b4b821da26541f5d056b94e2bba893d69a41294b9bc51ef3ed360aa", + "prompt_hash": "0c8417a74421a7ad9ab39e3bcb35cba73d03b091094d707553204f4773194c3e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 6, + "doc": { + "question": "What is the formula for the number of parameters in one layer of KAN?", + "choices": [ + "(d_in × d_out) + d_out", + "(d_in × d_out) × (G + K + 3) + d_out", + "(d_in × d_out) × (K + 1) + G", + "(d_in × d_out) × G + K + d_out" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the formula for the number of parameters in one layer of KAN?\nAnswer:", + " (d_in × d_out) + d_out" + ], + [ + "Question:What is the formula for the number of parameters in one layer of KAN?\nAnswer:", + " (d_in × d_out) × (G + K + 3) + d_out" + ], + [ + "Question:What is the formula for the number of parameters in one layer of KAN?\nAnswer:", + " (d_in × d_out) × (K + 1) + G" + ], + [ + "Question:What is the formula for the number of parameters in one layer of KAN?\nAnswer:", + " (d_in × d_out) × G + K + d_out" + ] + ], + "resps": [ + [ + [ + -35.33420944213867, + false + ] + ], + [ + [ + -62.38064193725586, + false + ] + ], + [ + [ + -51.30956268310547, + false + ] + ], + [ + [ + -52.87654113769531, + false + ] + ] + ], + "filtered_resps": [ + [ + -35.33420944213867, + false + ], + [ + -62.38064193725586, + false + ], + [ + -51.30956268310547, + false + ], + [ + -52.87654113769531, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6daacd702e0384b76b887d3bd437cd38b9be76e1212ae8929806b70a47db4db6", + "prompt_hash": "6c9621501c25caaf5fe65ccccbf091ede450342dfedd95e3427774408a59a814", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 7, + "doc": { + "question": "In symbolic formula representing tasks, what is the observed performance of KAN when compared to MLP under equal parameter settings?", + "choices": [ + "KAN outperforms MLP on all datasets", + "KAN performs worse than MLP on all datasets", + "KAN outperforms MLP on 7 out of 8 datasets", + "KAN matches MLP's performance on all datasets" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In symbolic formula representing tasks, what is the observed performance of KAN when compared to MLP under equal parameter settings?\nAnswer:", + " KAN outperforms MLP on all datasets" + ], + [ + "Question:In symbolic formula representing tasks, what is the observed performance of KAN when compared to MLP under equal parameter settings?\nAnswer:", + " KAN performs worse than MLP on all datasets" + ], + [ + "Question:In symbolic formula representing tasks, what is the observed performance of KAN when compared to MLP under equal parameter settings?\nAnswer:", + " KAN outperforms MLP on 7 out of 8 datasets" + ], + [ + "Question:In symbolic formula representing tasks, what is the observed performance of KAN when compared to MLP under equal parameter settings?\nAnswer:", + " KAN matches MLP's performance on all datasets" + ] + ], + "resps": [ + [ + [ + -15.381976127624512, + false + ] + ], + [ + [ + -17.349285125732422, + false + ] + ], + [ + [ + -20.085693359375, + false + ] + ], + [ + [ + -25.82134246826172, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.381976127624512, + false + ], + [ + -17.349285125732422, + false + ], + [ + -20.085693359375, + false + ], + [ + -25.82134246826172, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0422e76e7024b4c69b03f0b91afd3ee7ac85822d48f642aec7da11033312ccd2", + "prompt_hash": "27932ccdc73ae1363aa597199437308d54d250a9fe6c9800766248ce37794fe0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 8, + "doc": { + "question": "What is the impact of using B-spline activation functions in MLP for symbolic formula representation?", + "choices": [ + "It degrades performance compared to ReLU and GELU", + "It causes MLP to perform worse than KAN", + "It allows MLP to match or exceed the performance of KAN", + "It leads to unstable training results" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the impact of using B-spline activation functions in MLP for symbolic formula representation?\nAnswer:", + " It degrades performance compared to ReLU and GELU" + ], + [ + "Question:What is the impact of using B-spline activation functions in MLP for symbolic formula representation?\nAnswer:", + " It causes MLP to perform worse than KAN" + ], + [ + "Question:What is the impact of using B-spline activation functions in MLP for symbolic formula representation?\nAnswer:", + " It allows MLP to match or exceed the performance of KAN" + ], + [ + "Question:What is the impact of using B-spline activation functions in MLP for symbolic formula representation?\nAnswer:", + " It leads to unstable training results" + ] + ], + "resps": [ + [ + [ + -26.361316680908203, + false + ] + ], + [ + [ + -32.22442626953125, + false + ] + ], + [ + [ + -33.5744743347168, + false + ] + ], + [ + [ + -22.515552520751953, + false + ] + ] + ], + "filtered_resps": [ + [ + -26.361316680908203, + false + ], + [ + -32.22442626953125, + false + ], + [ + -33.5744743347168, + false + ], + [ + -22.515552520751953, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5819bef38b6c4a8aedafc0779016465827c41c35096c1de0f338950589d288c0", + "prompt_hash": "edb092d62713dc9bf15940bb778ea24bef9a4af3cd8757d7028acfdcf1d11dbf", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 9, + "doc": { + "question": "In continual learning experiments on MNIST, what was the key difference between KAN and MLP?", + "choices": [ + "KAN exhibited better retention on earlier tasks than MLP", + "MLP's accuracy dropped to zero on all tasks", + "KAN's accuracy dropped to zero on the first and second tasks", + "Both models achieved similar performance on all tasks" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In continual learning experiments on MNIST, what was the key difference between KAN and MLP?\nAnswer:", + " KAN exhibited better retention on earlier tasks than MLP" + ], + [ + "Question:In continual learning experiments on MNIST, what was the key difference between KAN and MLP?\nAnswer:", + " MLP's accuracy dropped to zero on all tasks" + ], + [ + "Question:In continual learning experiments on MNIST, what was the key difference between KAN and MLP?\nAnswer:", + " KAN's accuracy dropped to zero on the first and second tasks" + ], + [ + "Question:In continual learning experiments on MNIST, what was the key difference between KAN and MLP?\nAnswer:", + " Both models achieved similar performance on all tasks" + ] + ], + "resps": [ + [ + [ + -31.27310562133789, + false + ] + ], + [ + [ + -35.415992736816406, + false + ] + ], + [ + [ + -39.42005157470703, + false + ] + ], + [ + [ + -25.61410140991211, + false + ] + ] + ], + "filtered_resps": [ + [ + -31.27310562133789, + false + ], + [ + -35.415992736816406, + false + ], + [ + -39.42005157470703, + false + ], + [ + -25.61410140991211, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7e145678f19dd1a4b83d91cb268d1242b9e4f1b1dc845419014afacca59518d4", + "prompt_hash": "16c3f9f67701a26d194a004914d00521cc41f1bc6aea494b53e38fa8293ea555", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 10, + "doc": { + "question": "What percentage of the world population is estimated to carry a location sensor due to smartphone ownership?", + "choices": [ + "65%", + "70%", + "85%", + "90%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What percentage of the world population is estimated to carry a location sensor due to smartphone ownership?\nAnswer:", + " 65%" + ], + [ + "Question:What percentage of the world population is estimated to carry a location sensor due to smartphone ownership?\nAnswer:", + " 70%" + ], + [ + "Question:What percentage of the world population is estimated to carry a location sensor due to smartphone ownership?\nAnswer:", + " 85%" + ], + [ + "Question:What percentage of the world population is estimated to carry a location sensor due to smartphone ownership?\nAnswer:", + " 90%" + ] + ], + "resps": [ + [ + [ + -11.196426391601562, + false + ] + ], + [ + [ + -10.125539779663086, + false + ] + ], + [ + [ + -11.262310028076172, + false + ] + ], + [ + [ + -10.178831100463867, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.196426391601562, + false + ], + [ + -10.125539779663086, + false + ], + [ + -11.262310028076172, + false + ], + [ + -10.178831100463867, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1141f4d522435d88a7e99c842d874368792cc64ef497e76fadc94e22760ea1d1", + "prompt_hash": "af6926ce5c08968af4aef7cc9ae47432a9649e3156aa85fae74a2b1c68a3004c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 11, + "doc": { + "question": "Which type of cells in the brain encode regional boundaries in spatial memory?", + "choices": [ + "Place cells", + "Border cells", + "Grid cells", + "Goal direction cells" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which type of cells in the brain encode regional boundaries in spatial memory?\nAnswer:", + " Place cells" + ], + [ + "Question:Which type of cells in the brain encode regional boundaries in spatial memory?\nAnswer:", + " Border cells" + ], + [ + "Question:Which type of cells in the brain encode regional boundaries in spatial memory?\nAnswer:", + " Grid cells" + ], + [ + "Question:Which type of cells in the brain encode regional boundaries in spatial memory?\nAnswer:", + " Goal direction cells" + ] + ], + "resps": [ + [ + [ + -9.464567184448242, + false + ] + ], + [ + [ + -12.598109245300293, + false + ] + ], + [ + [ + -10.90217399597168, + false + ] + ], + [ + [ + -22.165721893310547, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.464567184448242, + false + ], + [ + -12.598109245300293, + false + ], + [ + -10.90217399597168, + false + ], + [ + -22.165721893310547, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fbcbb23568e5ca42e2399aa75ebcd4d5ac6bad4d1546255ebf5e1d4c249a04ef", + "prompt_hash": "8916e785d3289e19cee658505debe89b3d67ae8e1d4274f3d30de8a64c39cbc7", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 12, + "doc": { + "question": "Which five elements constitute the perceptions of cities in Kevin Lynch’s framework?", + "choices": [ + "Paths, zones, directions, nodes, views", + "Landmarks, borders, areas, paths, cues", + "Edges, nodes, landmarks, regions, paths", + "Streets, plazas, edges, centers, directions" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which five elements constitute the perceptions of cities in Kevin Lynch’s framework?\nAnswer:", + " Paths, zones, directions, nodes, views" + ], + [ + "Question:Which five elements constitute the perceptions of cities in Kevin Lynch’s framework?\nAnswer:", + " Landmarks, borders, areas, paths, cues" + ], + [ + "Question:Which five elements constitute the perceptions of cities in Kevin Lynch’s framework?\nAnswer:", + " Edges, nodes, landmarks, regions, paths" + ], + [ + "Question:Which five elements constitute the perceptions of cities in Kevin Lynch’s framework?\nAnswer:", + " Streets, plazas, edges, centers, directions" + ] + ], + "resps": [ + [ + [ + -44.72914505004883, + false + ] + ], + [ + [ + -50.86505889892578, + false + ] + ], + [ + [ + -42.66518020629883, + false + ] + ], + [ + [ + -45.75241470336914, + false + ] + ] + ], + "filtered_resps": [ + [ + -44.72914505004883, + false + ], + [ + -50.86505889892578, + false + ], + [ + -42.66518020629883, + false + ], + [ + -45.75241470336914, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "79a4226348287d8cfd75ea3ab13ab54f5720f434c1a1ce9a3907d0bea4e1be4e", + "prompt_hash": "a76f1a1be672cf9a876ede967bb11db4dcdfb132b57f0ddb98837b99a000de9e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 13, + "doc": { + "question": "What proportion of mobile phone subscriptions per 100 people was recorded in high-income countries in 2020?", + "choices": [ + "59", + "85", + "100", + "122" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What proportion of mobile phone subscriptions per 100 people was recorded in high-income countries in 2020?\nAnswer:", + " 59" + ], + [ + "Question:What proportion of mobile phone subscriptions per 100 people was recorded in high-income countries in 2020?\nAnswer:", + " 85" + ], + [ + "Question:What proportion of mobile phone subscriptions per 100 people was recorded in high-income countries in 2020?\nAnswer:", + " 100" + ], + [ + "Question:What proportion of mobile phone subscriptions per 100 people was recorded in high-income countries in 2020?\nAnswer:", + " 122" + ] + ], + "resps": [ + [ + [ + -7.109789848327637, + false + ] + ], + [ + [ + -7.487434387207031, + false + ] + ], + [ + [ + -7.320727348327637, + false + ] + ], + [ + [ + -9.849771499633789, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.109789848327637, + false + ], + [ + -7.487434387207031, + false + ], + [ + -7.320727348327637, + false + ], + [ + -9.849771499633789, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "515273aa9ee0792dc3371cff6d2cb2b7c54d4b91ed4d9bba6aab772d426428b0", + "prompt_hash": "769fc9a81750bd9200d69d2df3de1d5c2c66edbbab6a6daac4bcc0e912cd9a1a", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 14, + "doc": { + "question": "Which AI architectures are currently being investigated for modeling individual trajectories?", + "choices": [ + "Transformer networks", + "Decision trees", + "Naive Bayes classifiers", + "Support vector machines" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which AI architectures are currently being investigated for modeling individual trajectories?\nAnswer:", + " Transformer networks" + ], + [ + "Question:Which AI architectures are currently being investigated for modeling individual trajectories?\nAnswer:", + " Decision trees" + ], + [ + "Question:Which AI architectures are currently being investigated for modeling individual trajectories?\nAnswer:", + " Naive Bayes classifiers" + ], + [ + "Question:Which AI architectures are currently being investigated for modeling individual trajectories?\nAnswer:", + " Support vector machines" + ] + ], + "resps": [ + [ + [ + -11.995979309082031, + false + ] + ], + [ + [ + -9.45278549194336, + false + ] + ], + [ + [ + -18.51933479309082, + false + ] + ], + [ + [ + -16.56319808959961, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.995979309082031, + false + ], + [ + -9.45278549194336, + false + ], + [ + -18.51933479309082, + false + ], + [ + -16.56319808959961, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "475ef1d0e01261c4e570972e3e9e62387d7a1ec4d652bc497728c5a0ae2edc72", + "prompt_hash": "9d6e1d74a7034601fd57fc7d31c8595cfd9c915813ecb87f2766abdf316baa48", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 15, + "doc": { + "question": "What model is used as the base language model in the revision model experiments?", + "choices": [ + "GPT-4", + "Gemini", + "PaLM 2-S*", + "DeepSeekMath" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What model is used as the base language model in the revision model experiments?\nAnswer:", + " GPT-4" + ], + [ + "Question:What model is used as the base language model in the revision model experiments?\nAnswer:", + " Gemini" + ], + [ + "Question:What model is used as the base language model in the revision model experiments?\nAnswer:", + " PaLM 2-S*" + ], + [ + "Question:What model is used as the base language model in the revision model experiments?\nAnswer:", + " DeepSeekMath" + ] + ], + "resps": [ + [ + [ + -4.3542022705078125, + false + ] + ], + [ + [ + -7.10027551651001, + false + ] + ], + [ + [ + -22.711105346679688, + false + ] + ], + [ + [ + -8.928313255310059, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.3542022705078125, + false + ], + [ + -7.10027551651001, + false + ], + [ + -22.711105346679688, + false + ], + [ + -8.928313255310059, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8337c7e84c1d80f0359656a5652ee2eccec31324c1a601d8f001f21c3d6eb339", + "prompt_hash": "8b451711454ba8e6bdddaeb09477f1dfab62d3a1213412ba92e032097ccd54eb", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 16, + "doc": { + "question": "What optimizer is used to train the PRM model?", + "choices": [ + "SGD", + "Adam", + "RMSProp", + "AdamW" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What optimizer is used to train the PRM model?\nAnswer:", + " SGD" + ], + [ + "Question:What optimizer is used to train the PRM model?\nAnswer:", + " Adam" + ], + [ + "Question:What optimizer is used to train the PRM model?\nAnswer:", + " RMSProp" + ], + [ + "Question:What optimizer is used to train the PRM model?\nAnswer:", + " AdamW" + ] + ], + "resps": [ + [ + [ + -6.562948226928711, + false + ] + ], + [ + [ + -3.0958411693573, + false + ] + ], + [ + [ + -7.73266077041626, + false + ] + ], + [ + [ + -6.134220123291016, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.562948226928711, + false + ], + [ + -3.0958411693573, + false + ], + [ + -7.73266077041626, + false + ], + [ + -6.134220123291016, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "51df039bbc085148d23b1cdb96c6980b12dc391a8c701d0a0f90203d7bdaf97d", + "prompt_hash": "d3ceb3491c7b106b1ac9831e2ac891743f9610880eaf5e54e5640aedae1c7beb", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 17, + "doc": { + "question": "What aggregation strategy is found to perform best when aggregating PRM per-step scores?", + "choices": [ + "Min", + "Product", + "Last", + "Average" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What aggregation strategy is found to perform best when aggregating PRM per-step scores?\nAnswer:", + " Min" + ], + [ + "Question:What aggregation strategy is found to perform best when aggregating PRM per-step scores?\nAnswer:", + " Product" + ], + [ + "Question:What aggregation strategy is found to perform best when aggregating PRM per-step scores?\nAnswer:", + " Last" + ], + [ + "Question:What aggregation strategy is found to perform best when aggregating PRM per-step scores?\nAnswer:", + " Average" + ] + ], + "resps": [ + [ + [ + -7.221947193145752, + false + ] + ], + [ + [ + -12.822408676147461, + false + ] + ], + [ + [ + -9.58028793334961, + false + ] + ], + [ + [ + -5.567903518676758, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.221947193145752, + false + ], + [ + -12.822408676147461, + false + ], + [ + -9.58028793334961, + false + ], + [ + -5.567903518676758, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a2cca9fe75f490b9d2c160579dcc2cbc475bd78c0212f3a1262c0b35e1d734e4", + "prompt_hash": "a26d398514ea5f1ad9c41d45ab6110045960417a6404f3537fd8ad8e0be9af26", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 18, + "doc": { + "question": "How many revision trajectories are generated per question in the ReST^EM experiment?", + "choices": [ + "16", + "32", + "64", + "128" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many revision trajectories are generated per question in the ReST^EM experiment?\nAnswer:", + " 16" + ], + [ + "Question:How many revision trajectories are generated per question in the ReST^EM experiment?\nAnswer:", + " 32" + ], + [ + "Question:How many revision trajectories are generated per question in the ReST^EM experiment?\nAnswer:", + " 64" + ], + [ + "Question:How many revision trajectories are generated per question in the ReST^EM experiment?\nAnswer:", + " 128" + ] + ], + "resps": [ + [ + [ + -5.620156288146973, + false + ] + ], + [ + [ + -5.936687469482422, + false + ] + ], + [ + [ + -7.073380470275879, + false + ] + ], + [ + [ + -8.123558044433594, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.620156288146973, + false + ], + [ + -5.936687469482422, + false + ], + [ + -7.073380470275879, + false + ], + [ + -8.123558044433594, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "25a48f5278cf30831b8fcc9459e68c0629abecb992d2ad92897f60dbb567c294", + "prompt_hash": "76fced7a6d9d02e3bc0d5d3a95f27d27dff4ed7aa83814c3d31253d98376feba", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 19, + "doc": { + "question": "What is used as the final answer selection method for majority voting in revision experiments?", + "choices": [ + "Best-of-N weighted within each chain only", + "Majority voting across all trajectories at once", + "Greedy selection from last revision", + "Averaging PRM scores over revisions" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is used as the final answer selection method for majority voting in revision experiments?\nAnswer:", + " Best-of-N weighted within each chain only" + ], + [ + "Question:What is used as the final answer selection method for majority voting in revision experiments?\nAnswer:", + " Majority voting across all trajectories at once" + ], + [ + "Question:What is used as the final answer selection method for majority voting in revision experiments?\nAnswer:", + " Greedy selection from last revision" + ], + [ + "Question:What is used as the final answer selection method for majority voting in revision experiments?\nAnswer:", + " Averaging PRM scores over revisions" + ] + ], + "resps": [ + [ + [ + -50.39833068847656, + false + ] + ], + [ + [ + -38.51416778564453, + false + ] + ], + [ + [ + -28.31541633605957, + false + ] + ], + [ + [ + -37.91179275512695, + false + ] + ] + ], + "filtered_resps": [ + [ + -50.39833068847656, + false + ], + [ + -38.51416778564453, + false + ], + [ + -28.31541633605957, + false + ], + [ + -37.91179275512695, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "32cec6c9dac45cdbd7b4ae2ff91380b2948758777d6d00398cf5fec3dc69ca7a", + "prompt_hash": "c252151d86183b34e325f3f2a07d726cb0f9c50eae68c43ccc1c5f589dd497f8", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 20, + "doc": { + "question": "What is the image resolution supported by MiniCPM-V 2.0 and MiniCPM-Llama3-V 2.5?", + "choices": [ + "448 × 448", + "1024 × 1024", + "1344 × 1344", + "2048 × 2048" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the image resolution supported by MiniCPM-V 2.0 and MiniCPM-Llama3-V 2.5?\nAnswer:", + " 448 × 448" + ], + [ + "Question:What is the image resolution supported by MiniCPM-V 2.0 and MiniCPM-Llama3-V 2.5?\nAnswer:", + " 1024 × 1024" + ], + [ + "Question:What is the image resolution supported by MiniCPM-V 2.0 and MiniCPM-Llama3-V 2.5?\nAnswer:", + " 1344 × 1344" + ], + [ + "Question:What is the image resolution supported by MiniCPM-V 2.0 and MiniCPM-Llama3-V 2.5?\nAnswer:", + " 2048 × 2048" + ] + ], + "resps": [ + [ + [ + -8.775908470153809, + false + ] + ], + [ + [ + -8.181621551513672, + false + ] + ], + [ + [ + -16.271177291870117, + false + ] + ], + [ + [ + -9.781907081604004, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.775908470153809, + false + ], + [ + -8.181621551513672, + false + ], + [ + -16.271177291870117, + false + ], + [ + -9.781907081604004, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "82090db10931345cca27913e7f38d4bfff7672f3fce02e9017bce5f71b0cdecd", + "prompt_hash": "330f53849fa0a8ddf62d3e07304633ea021f4b8ae15e19c127f3ccce591f8140", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 21, + "doc": { + "question": "Which model in the MiniCPM-V series uses Llama3-Instruct 8B as its base LLM?", + "choices": [ + "MiniCPM-V 1.0", + "MiniCPM-V 2.0", + "MiniCPM-Llama3-V 2.5", + "MiniCPM-Gemma 2.0" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model in the MiniCPM-V series uses Llama3-Instruct 8B as its base LLM?\nAnswer:", + " MiniCPM-V 1.0" + ], + [ + "Question:Which model in the MiniCPM-V series uses Llama3-Instruct 8B as its base LLM?\nAnswer:", + " MiniCPM-V 2.0" + ], + [ + "Question:Which model in the MiniCPM-V series uses Llama3-Instruct 8B as its base LLM?\nAnswer:", + " MiniCPM-Llama3-V 2.5" + ], + [ + "Question:Which model in the MiniCPM-V series uses Llama3-Instruct 8B as its base LLM?\nAnswer:", + " MiniCPM-Gemma 2.0" + ] + ], + "resps": [ + [ + [ + -9.62120246887207, + false + ] + ], + [ + [ + -9.776691436767578, + false + ] + ], + [ + [ + -20.5820255279541, + false + ] + ], + [ + [ + -21.101959228515625, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.62120246887207, + false + ], + [ + -9.776691436767578, + false + ], + [ + -20.5820255279541, + false + ], + [ + -21.101959228515625, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "85bd1e157ef3f4bd174f5747f899fc2ec436766a522f5474ac8656ebc11a1d46", + "prompt_hash": "aa8c134e336cda1667e37d843d645a0159908280757421028f42bf05f788bc7f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 22, + "doc": { + "question": "Which technique is used in MiniCPM-Llama3-V 2.5 to reduce hallucination rates?", + "choices": [ + "Multi-head Attention Dropout", + "Gradient Checkpointing", + "RLAIF-V", + "Contextual Retrieval Augmentation" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which technique is used in MiniCPM-Llama3-V 2.5 to reduce hallucination rates?\nAnswer:", + " Multi-head Attention Dropout" + ], + [ + "Question:Which technique is used in MiniCPM-Llama3-V 2.5 to reduce hallucination rates?\nAnswer:", + " Gradient Checkpointing" + ], + [ + "Question:Which technique is used in MiniCPM-Llama3-V 2.5 to reduce hallucination rates?\nAnswer:", + " RLAIF-V" + ], + [ + "Question:Which technique is used in MiniCPM-Llama3-V 2.5 to reduce hallucination rates?\nAnswer:", + " Contextual Retrieval Augmentation" + ] + ], + "resps": [ + [ + [ + -20.53322410583496, + false + ] + ], + [ + [ + -13.587026596069336, + false + ] + ], + [ + [ + -17.670570373535156, + false + ] + ], + [ + [ + -19.511798858642578, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.53322410583496, + false + ], + [ + -13.587026596069336, + false + ], + [ + -17.670570373535156, + false + ], + [ + -19.511798858642578, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9cc7b1431eecfbefaab0cad4d1bfc115fbee148efffefcfe3d47f7606be3bfa1", + "prompt_hash": "fd737561175617926bfa8f244f46d3099c33f30f5b671c138f38204c596bfcc6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 23, + "doc": { + "question": "What visual encoder is employed in MiniCPM-V?", + "choices": [ + "ViT-B/16", + "ResNet-101", + "SigLIP SoViT-400m/14", + "ConvNeXt-Base" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What visual encoder is employed in MiniCPM-V?\nAnswer:", + " ViT-B/16" + ], + [ + "Question:What visual encoder is employed in MiniCPM-V?\nAnswer:", + " ResNet-101" + ], + [ + "Question:What visual encoder is employed in MiniCPM-V?\nAnswer:", + " SigLIP SoViT-400m/14" + ], + [ + "Question:What visual encoder is employed in MiniCPM-V?\nAnswer:", + " ConvNeXt-Base" + ] + ], + "resps": [ + [ + [ + -8.633031845092773, + false + ] + ], + [ + [ + -7.921908378601074, + false + ] + ], + [ + [ + -60.48279571533203, + false + ] + ], + [ + [ + -12.991559028625488, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.633031845092773, + false + ], + [ + -7.921908378601074, + false + ], + [ + -60.48279571533203, + false + ], + [ + -12.991559028625488, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0bcf0b692a7f5d9b6dac4b2405e70338d6c5e404c7bf8402cd5cf3bcd39a08a5", + "prompt_hash": "a19a91d025aa9bcc97165320583e88300005349ba251f66af54c7b65b72dba7a", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 24, + "doc": { + "question": "What type of quantization is used to reduce the memory requirement of MiniCPM-Llama3-V 2.5 to around 5GB?", + "choices": [ + "Q8_0 mode", + "Q4_K_M mode 4-bit", + "INT8 symmetric quantization", + "Post-training dynamic quantization" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What type of quantization is used to reduce the memory requirement of MiniCPM-Llama3-V 2.5 to around 5GB?\nAnswer:", + " Q8_0 mode" + ], + [ + "Question:What type of quantization is used to reduce the memory requirement of MiniCPM-Llama3-V 2.5 to around 5GB?\nAnswer:", + " Q4_K_M mode 4-bit" + ], + [ + "Question:What type of quantization is used to reduce the memory requirement of MiniCPM-Llama3-V 2.5 to around 5GB?\nAnswer:", + " INT8 symmetric quantization" + ], + [ + "Question:What type of quantization is used to reduce the memory requirement of MiniCPM-Llama3-V 2.5 to around 5GB?\nAnswer:", + " Post-training dynamic quantization" + ] + ], + "resps": [ + [ + [ + -27.991249084472656, + false + ] + ], + [ + [ + -52.5985107421875, + false + ] + ], + [ + [ + -18.92511749267578, + false + ] + ], + [ + [ + -19.703536987304688, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.991249084472656, + false + ], + [ + -52.5985107421875, + false + ], + [ + -18.92511749267578, + false + ], + [ + -19.703536987304688, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "171a9a1a3420e7c88940f30480db74ec63f19e986bc54eb2b008f0d0f6dab4fe", + "prompt_hash": "9371558e2e30178b96278ed6261f1d3f2c9360cb841f0503c558ab264ac0153b", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 25, + "doc": { + "question": "Which visual encoder is used by default in mPLUG-Owl3?", + "choices": [ + "Siglip-400m", + "CLIP-ViT", + "BLIP-2", + "ResNet-101" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which visual encoder is used by default in mPLUG-Owl3?\nAnswer:", + " Siglip-400m" + ], + [ + "Question:Which visual encoder is used by default in mPLUG-Owl3?\nAnswer:", + " CLIP-ViT" + ], + [ + "Question:Which visual encoder is used by default in mPLUG-Owl3?\nAnswer:", + " BLIP-2" + ], + [ + "Question:Which visual encoder is used by default in mPLUG-Owl3?\nAnswer:", + " ResNet-101" + ] + ], + "resps": [ + [ + [ + -32.85588836669922, + false + ] + ], + [ + [ + -10.125411987304688, + false + ] + ], + [ + [ + -7.34080696105957, + false + ] + ], + [ + [ + -9.431852340698242, + false + ] + ] + ], + "filtered_resps": [ + [ + -32.85588836669922, + false + ], + [ + -10.125411987304688, + false + ], + [ + -7.34080696105957, + false + ], + [ + -9.431852340698242, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "45f0233992f0526f9e2ddd64462932a6d005f950483384bbfbca4e03e739a662", + "prompt_hash": "3720b56a27492cba6007f1d309029ab312d06cd133740ae26813ef842ac9ad04", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 26, + "doc": { + "question": "Which benchmark evaluates a model's ability to understand science diagrams?", + "choices": [ + "POPE", + "AI2D", + "MM-Vet", + "MMBench" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which benchmark evaluates a model's ability to understand science diagrams?\nAnswer:", + " POPE" + ], + [ + "Question:Which benchmark evaluates a model's ability to understand science diagrams?\nAnswer:", + " AI2D" + ], + [ + "Question:Which benchmark evaluates a model's ability to understand science diagrams?\nAnswer:", + " MM-Vet" + ], + [ + "Question:Which benchmark evaluates a model's ability to understand science diagrams?\nAnswer:", + " MMBench" + ] + ], + "resps": [ + [ + [ + -10.768207550048828, + false + ] + ], + [ + [ + -10.82089900970459, + false + ] + ], + [ + [ + -18.138641357421875, + false + ] + ], + [ + [ + -13.92343521118164, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.768207550048828, + false + ], + [ + -10.82089900970459, + false + ], + [ + -18.138641357421875, + false + ], + [ + -13.92343521118164, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a0c39f7cbd5af613b6c4a6a7fe47428530770d0eef8a26d5baaf7d7f040ee0a4", + "prompt_hash": "76dc790ed1dd7a04d27b4229b74a54ac654ebe1a670bdc39f84965a75d3fd8b0", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 27, + "doc": { + "question": "Which component is reused as the Query vector in the Hyper Attention cross-attention mechanism?", + "choices": [ + "Key from visual features", + "Value from textual features", + "Query from the self-attention mechanism", + "Projection from the vision encoder" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which component is reused as the Query vector in the Hyper Attention cross-attention mechanism?\nAnswer:", + " Key from visual features" + ], + [ + "Question:Which component is reused as the Query vector in the Hyper Attention cross-attention mechanism?\nAnswer:", + " Value from textual features" + ], + [ + "Question:Which component is reused as the Query vector in the Hyper Attention cross-attention mechanism?\nAnswer:", + " Query from the self-attention mechanism" + ], + [ + "Question:Which component is reused as the Query vector in the Hyper Attention cross-attention mechanism?\nAnswer:", + " Projection from the vision encoder" + ] + ], + "resps": [ + [ + [ + -24.740676879882812, + false + ] + ], + [ + [ + -28.112730026245117, + false + ] + ], + [ + [ + -18.406658172607422, + false + ] + ], + [ + [ + -28.374183654785156, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.740676879882812, + false + ], + [ + -28.112730026245117, + false + ], + [ + -18.406658172607422, + false + ], + [ + -28.374183654785156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d6e1aeb928e60b21d65c1bac533238b17407b50ce0764e7918e781c5cf0ee053", + "prompt_hash": "11bffdf541567d8f57e5e5a7070dfad452865ebcb490b23d144e8a14c5068895", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 28, + "doc": { + "question": "How many image-text pairs are used for pre-training mPLUG-Owl3?", + "choices": [ + "10 million", + "25 million", + "41 million", + "65 million" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many image-text pairs are used for pre-training mPLUG-Owl3?\nAnswer:", + " 10 million" + ], + [ + "Question:How many image-text pairs are used for pre-training mPLUG-Owl3?\nAnswer:", + " 25 million" + ], + [ + "Question:How many image-text pairs are used for pre-training mPLUG-Owl3?\nAnswer:", + " 41 million" + ], + [ + "Question:How many image-text pairs are used for pre-training mPLUG-Owl3?\nAnswer:", + " 65 million" + ] + ], + "resps": [ + [ + [ + -8.595640182495117, + false + ] + ], + [ + [ + -10.362495422363281, + false + ] + ], + [ + [ + -12.96182918548584, + false + ] + ], + [ + [ + -11.332928657531738, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.595640182495117, + false + ], + [ + -10.362495422363281, + false + ], + [ + -12.96182918548584, + false + ], + [ + -11.332928657531738, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6223644cc4cf68c322f7b5014ae6ff464cef6f65987f256ad3709d4167105ed3", + "prompt_hash": "8dc3d0f2ed68203b2a262b65e63a7bd7cfc7354608ce01694ec99ac2d9ff13af", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 29, + "doc": { + "question": "Which technique is introduced in mPLUG-Owl3 to preserve image positional information in interleaved sequences?", + "choices": [ + "Rotary Positional Embedding", + "Vision Transformer", + "Multimodal-Interleaved Rotary Position Embedding (MI-Rope)", + "Temporal Fusion Encoder" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which technique is introduced in mPLUG-Owl3 to preserve image positional information in interleaved sequences?\nAnswer:", + " Rotary Positional Embedding" + ], + [ + "Question:Which technique is introduced in mPLUG-Owl3 to preserve image positional information in interleaved sequences?\nAnswer:", + " Vision Transformer" + ], + [ + "Question:Which technique is introduced in mPLUG-Owl3 to preserve image positional information in interleaved sequences?\nAnswer:", + " Multimodal-Interleaved Rotary Position Embedding (MI-Rope)" + ], + [ + "Question:Which technique is introduced in mPLUG-Owl3 to preserve image positional information in interleaved sequences?\nAnswer:", + " Temporal Fusion Encoder" + ] + ], + "resps": [ + [ + [ + -20.903343200683594, + false + ] + ], + [ + [ + -11.205512046813965, + false + ] + ], + [ + [ + -62.90228271484375, + false + ] + ], + [ + [ + -19.450517654418945, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.903343200683594, + false + ], + [ + -11.205512046813965, + false + ], + [ + -62.90228271484375, + false + ], + [ + -19.450517654418945, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "cc614c25a588bd9452a11be0ad9545544972a67e547ef9d3087f295344e8ef11", + "prompt_hash": "81b34377e5e16772b0ab5f0bf087b92e357f2585521d1c53857c44370f4f33e3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 30, + "doc": { + "question": "What does the HybridRAG system combine to enhance information retrieval?", + "choices": [ + "VectorRAG and WebRAG", + "GraphRAG and Symbolic Logic", + "VectorRAG and GraphRAG", + "Knowledge Graph and Decision Trees" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does the HybridRAG system combine to enhance information retrieval?\nAnswer:", + " VectorRAG and WebRAG" + ], + [ + "Question:What does the HybridRAG system combine to enhance information retrieval?\nAnswer:", + " GraphRAG and Symbolic Logic" + ], + [ + "Question:What does the HybridRAG system combine to enhance information retrieval?\nAnswer:", + " VectorRAG and GraphRAG" + ], + [ + "Question:What does the HybridRAG system combine to enhance information retrieval?\nAnswer:", + " Knowledge Graph and Decision Trees" + ] + ], + "resps": [ + [ + [ + -34.7384147644043, + false + ] + ], + [ + [ + -39.80352020263672, + false + ] + ], + [ + [ + -29.742319107055664, + false + ] + ], + [ + [ + -24.235549926757812, + false + ] + ] + ], + "filtered_resps": [ + [ + -34.7384147644043, + false + ], + [ + -39.80352020263672, + false + ], + [ + -29.742319107055664, + false + ], + [ + -24.235549926757812, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4e4bddf8d1d7a09a1e25a76b3b334f4773dd3a665f508a46dcf705157f7e844f", + "prompt_hash": "0fdedbbb63fe8d41251f7883aa0a4a161425d586290781222f71de0ac51d00e7", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 31, + "doc": { + "question": "What is the first step in the knowledge extraction process for constructing a knowledge graph?", + "choices": [ + "Entity Disambiguation", + "Knowledge Fusion", + "Knowledge Extraction", + "Temporal KG Construction" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the first step in the knowledge extraction process for constructing a knowledge graph?\nAnswer:", + " Entity Disambiguation" + ], + [ + "Question:What is the first step in the knowledge extraction process for constructing a knowledge graph?\nAnswer:", + " Knowledge Fusion" + ], + [ + "Question:What is the first step in the knowledge extraction process for constructing a knowledge graph?\nAnswer:", + " Knowledge Extraction" + ], + [ + "Question:What is the first step in the knowledge extraction process for constructing a knowledge graph?\nAnswer:", + " Temporal KG Construction" + ] + ], + "resps": [ + [ + [ + -13.80317497253418, + false + ] + ], + [ + [ + -13.312759399414062, + false + ] + ], + [ + [ + -6.910901069641113, + false + ] + ], + [ + [ + -18.86545181274414, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.80317497253418, + false + ], + [ + -13.312759399414062, + false + ], + [ + -6.910901069641113, + false + ], + [ + -18.86545181274414, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "121e6b0d2999afd4b068adada4169a6657a154acf6504ec312f11dbd1f473325", + "prompt_hash": "bd6cd21014f88e0e46adf28cece79c686071b520ee6f8b416105fa27e93b2cc7", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 32, + "doc": { + "question": "Which LLM embedding model is used in the VectorRAG setup?", + "choices": [ + "text-embedding-babbage-001", + "text-embedding-ada-002", + "sentence-transformers/all-MiniLM-L6-v2", + "bert-base-nli-mean-tokens" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which LLM embedding model is used in the VectorRAG setup?\nAnswer:", + " text-embedding-babbage-001" + ], + [ + "Question:Which LLM embedding model is used in the VectorRAG setup?\nAnswer:", + " text-embedding-ada-002" + ], + [ + "Question:Which LLM embedding model is used in the VectorRAG setup?\nAnswer:", + " sentence-transformers/all-MiniLM-L6-v2" + ], + [ + "Question:Which LLM embedding model is used in the VectorRAG setup?\nAnswer:", + " bert-base-nli-mean-tokens" + ] + ], + "resps": [ + [ + [ + -35.047874450683594, + false + ] + ], + [ + [ + -8.769493103027344, + false + ] + ], + [ + [ + -9.794758796691895, + false + ] + ], + [ + [ + -12.21292781829834, + false + ] + ] + ], + "filtered_resps": [ + [ + -35.047874450683594, + false + ], + [ + -8.769493103027344, + false + ], + [ + -9.794758796691895, + false + ], + [ + -12.21292781829834, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e5d0a848899461bf1ccf87ebe38f1531bbd1129582b1096b830f7c05a1fc2068", + "prompt_hash": "286cc4d87a6f8915033be3fe3283998246025225bbdb787a003cace0368e9b03", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 33, + "doc": { + "question": "Which sector is NOT listed among those included in the Nifty 50 earnings call dataset?", + "choices": [ + "Consumer Goods", + "Aerospace", + "Healthcare", + "Energy - Power" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which sector is NOT listed among those included in the Nifty 50 earnings call dataset?\nAnswer:", + " Consumer Goods" + ], + [ + "Question:Which sector is NOT listed among those included in the Nifty 50 earnings call dataset?\nAnswer:", + " Aerospace" + ], + [ + "Question:Which sector is NOT listed among those included in the Nifty 50 earnings call dataset?\nAnswer:", + " Healthcare" + ], + [ + "Question:Which sector is NOT listed among those included in the Nifty 50 earnings call dataset?\nAnswer:", + " Energy - Power" + ] + ], + "resps": [ + [ + [ + -12.80567741394043, + false + ] + ], + [ + [ + -8.0639066696167, + false + ] + ], + [ + [ + -6.295193195343018, + false + ] + ], + [ + [ + -19.14309310913086, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.80567741394043, + false + ], + [ + -8.0639066696167, + false + ], + [ + -6.295193195343018, + false + ], + [ + -19.14309310913086, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c4f0abe77cfd8a8682c3bd77a15030f0ad96cd9ccdaee1b95042dad325560bbd", + "prompt_hash": "7f81198daf1670620d9f0b27b723a619ea15b2c7005f9d523ac2b38ac90e1a67", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 34, + "doc": { + "question": "What metric measures how much of the generated answer can be inferred from the provided context?", + "choices": [ + "Context Precision", + "Faithfulness", + "Answer Relevance", + "Context Recall" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What metric measures how much of the generated answer can be inferred from the provided context?\nAnswer:", + " Context Precision" + ], + [ + "Question:What metric measures how much of the generated answer can be inferred from the provided context?\nAnswer:", + " Faithfulness" + ], + [ + "Question:What metric measures how much of the generated answer can be inferred from the provided context?\nAnswer:", + " Answer Relevance" + ], + [ + "Question:What metric measures how much of the generated answer can be inferred from the provided context?\nAnswer:", + " Context Recall" + ] + ], + "resps": [ + [ + [ + -13.284663200378418, + false + ] + ], + [ + [ + -9.042125701904297, + false + ] + ], + [ + [ + -9.404643058776855, + false + ] + ], + [ + [ + -12.565004348754883, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.284663200378418, + false + ], + [ + -9.042125701904297, + false + ], + [ + -9.404643058776855, + false + ], + [ + -12.565004348754883, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0a7992bd5405e2455bb7ae703195734ee1820ba790899d28e13c6ebe89a23403", + "prompt_hash": "c9b81d11c3bf2b77ca20585ce25ccdcb42b99e3832f02d5ec511fea14e837b3b", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 35, + "doc": { + "question": "What is the activation function used in JumpReLU sparse autoencoders?", + "choices": [ + "Sigmoid function", + "Shifted Heaviside step function combined with ReLU", + "Standard ReLU without thresholding", + "Tanh with adaptive scaling" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the activation function used in JumpReLU sparse autoencoders?\nAnswer:", + " Sigmoid function" + ], + [ + "Question:What is the activation function used in JumpReLU sparse autoencoders?\nAnswer:", + " Shifted Heaviside step function combined with ReLU" + ], + [ + "Question:What is the activation function used in JumpReLU sparse autoencoders?\nAnswer:", + " Standard ReLU without thresholding" + ], + [ + "Question:What is the activation function used in JumpReLU sparse autoencoders?\nAnswer:", + " Tanh with adaptive scaling" + ] + ], + "resps": [ + [ + [ + -6.715402603149414, + false + ] + ], + [ + [ + -36.528255462646484, + false + ] + ], + [ + [ + -22.49020004272461, + false + ] + ], + [ + [ + -20.276412963867188, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.715402603149414, + false + ], + [ + -36.528255462646484, + false + ], + [ + -22.49020004272461, + false + ], + [ + -20.276412963867188, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "cd9fc825aa138a5de9abaf4b87783158bd71ee7a58cc07a955ccee016045d4b4", + "prompt_hash": "dd3b8472617fc9214c3074ab154d01f4b93ce8dc5cc3ce9cad8f5a02e1ca9c65", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 36, + "doc": { + "question": "What optimizer is used to train the sparse autoencoders in Gemma Scope?", + "choices": [ + "SGD with momentum", + "RMSprop", + "Adam with (β₁, β₂) = (0, 0.999)", + "LAMB optimizer" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What optimizer is used to train the sparse autoencoders in Gemma Scope?\nAnswer:", + " SGD with momentum" + ], + [ + "Question:What optimizer is used to train the sparse autoencoders in Gemma Scope?\nAnswer:", + " RMSprop" + ], + [ + "Question:What optimizer is used to train the sparse autoencoders in Gemma Scope?\nAnswer:", + " Adam with (β₁, β₂) = (0, 0.999)" + ], + [ + "Question:What optimizer is used to train the sparse autoencoders in Gemma Scope?\nAnswer:", + " LAMB optimizer" + ] + ], + "resps": [ + [ + [ + -9.294715881347656, + false + ] + ], + [ + [ + -6.094078540802002, + false + ] + ], + [ + [ + -42.684532165527344, + false + ] + ], + [ + [ + -18.675086975097656, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.294715881347656, + false + ], + [ + -6.094078540802002, + false + ], + [ + -42.684532165527344, + false + ], + [ + -18.675086975097656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "de2f0308b302e5574f464cfe4d280ec12ddce5ae7374d0329b8ba50f15149a2b", + "prompt_hash": "bfc1545c194ee3ba20c4127cc60c14269bd7c1acea9a6895bc2d69d521fd7eee", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 37, + "doc": { + "question": "How many sparse autoencoders were released in the main Gemma Scope suite?", + "choices": [ + "Over 500", + "More than 400", + "Exactly 1000", + "About 100" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many sparse autoencoders were released in the main Gemma Scope suite?\nAnswer:", + " Over 500" + ], + [ + "Question:How many sparse autoencoders were released in the main Gemma Scope suite?\nAnswer:", + " More than 400" + ], + [ + "Question:How many sparse autoencoders were released in the main Gemma Scope suite?\nAnswer:", + " Exactly 1000" + ], + [ + "Question:How many sparse autoencoders were released in the main Gemma Scope suite?\nAnswer:", + " About 100" + ] + ], + "resps": [ + [ + [ + -14.328933715820312, + false + ] + ], + [ + [ + -15.292580604553223, + false + ] + ], + [ + [ + -18.465763092041016, + false + ] + ], + [ + [ + -13.992822647094727, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.328933715820312, + false + ], + [ + -15.292580604553223, + false + ], + [ + -18.465763092041016, + false + ], + [ + -13.992822647094727, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7850c75efcc76953fb620cc4a3af670b82c3e762747707a436a0957e4d0b0305", + "prompt_hash": "48ee84c6b1656b477831163a4cc8b71f25196393ec4831517a8282d90481b03f", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 38, + "doc": { + "question": "Which subset of The Pile dataset did the SAEs perform worst on?", + "choices": [ + "PubMed Central", + "Books3", + "Europarl", + "DeepMind Mathematics" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which subset of The Pile dataset did the SAEs perform worst on?\nAnswer:", + " PubMed Central" + ], + [ + "Question:Which subset of The Pile dataset did the SAEs perform worst on?\nAnswer:", + " Books3" + ], + [ + "Question:Which subset of The Pile dataset did the SAEs perform worst on?\nAnswer:", + " Europarl" + ], + [ + "Question:Which subset of The Pile dataset did the SAEs perform worst on?\nAnswer:", + " DeepMind Mathematics" + ] + ], + "resps": [ + [ + [ + -16.985855102539062, + false + ] + ], + [ + [ + -16.546205520629883, + false + ] + ], + [ + [ + -13.369335174560547, + false + ] + ], + [ + [ + -20.967653274536133, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.985855102539062, + false + ], + [ + -16.546205520629883, + false + ], + [ + -13.369335174560547, + false + ], + [ + -20.967653274536133, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a4c6cc58667fe8997ffb8b3b6267ff80e8ed45b7165caa316b931df0e4d133d0", + "prompt_hash": "076909a7af06d7da4669d0113738cb750045745399d1ab5485f09fa67193a98e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 39, + "doc": { + "question": "Where are the released weights of Gemma Scope hosted?", + "choices": [ + "OpenAI API", + "Google Cloud Storage", + "HuggingFace", + "Kaggle Datasets" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Where are the released weights of Gemma Scope hosted?\nAnswer:", + " OpenAI API" + ], + [ + "Question:Where are the released weights of Gemma Scope hosted?\nAnswer:", + " Google Cloud Storage" + ], + [ + "Question:Where are the released weights of Gemma Scope hosted?\nAnswer:", + " HuggingFace" + ], + [ + "Question:Where are the released weights of Gemma Scope hosted?\nAnswer:", + " Kaggle Datasets" + ] + ], + "resps": [ + [ + [ + -10.245248794555664, + false + ] + ], + [ + [ + -8.508049011230469, + false + ] + ], + [ + [ + -8.190152168273926, + false + ] + ], + [ + [ + -14.849760055541992, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.245248794555664, + false + ], + [ + -8.508049011230469, + false + ], + [ + -8.190152168273926, + false + ], + [ + -14.849760055541992, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e6a6066ec0f1781c89108fb4c80a0f7fecc28035e65e670b6c243bb3407759fa", + "prompt_hash": "e21190f5c3f44f85ac12be0cfc218a9571772a27f53432e7b90a7b2bce991577", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 40, + "doc": { + "question": "What is the accuracy of LLaMA2-7B on GSM8K when using few-shot CoT prompting?", + "choices": [ + "12.51%", + "24.34%", + "36.46%", + "63.91%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the accuracy of LLaMA2-7B on GSM8K when using few-shot CoT prompting?\nAnswer:", + " 12.51%" + ], + [ + "Question:What is the accuracy of LLaMA2-7B on GSM8K when using few-shot CoT prompting?\nAnswer:", + " 24.34%" + ], + [ + "Question:What is the accuracy of LLaMA2-7B on GSM8K when using few-shot CoT prompting?\nAnswer:", + " 36.46%" + ], + [ + "Question:What is the accuracy of LLaMA2-7B on GSM8K when using few-shot CoT prompting?\nAnswer:", + " 63.91%" + ] + ], + "resps": [ + [ + [ + -18.55419921875, + false + ] + ], + [ + [ + -17.668134689331055, + false + ] + ], + [ + [ + -17.610607147216797, + false + ] + ], + [ + [ + -16.299419403076172, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.55419921875, + false + ], + [ + -17.668134689331055, + false + ], + [ + -17.610607147216797, + false + ], + [ + -16.299419403076172, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "740e8672e7d357a835b0e5f48081bf47de1508d71d7ddbacef30b6ccfe7647b6", + "prompt_hash": "86183804e555821736b88e17d55e2e58e0f2ba608af3b603957c175c4bc474d1", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 41, + "doc": { + "question": "Which action in the reasoning process is inspired by least-to-most prompting and involves generating sub-questions and their answers?", + "choices": [ + "A1: Propose an one-step thought", + "A2: Propose the remaining thought steps", + "A3: Propose next sub-question along with its answer", + "A5: Rephrase the question/sub-question" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which action in the reasoning process is inspired by least-to-most prompting and involves generating sub-questions and their answers?\nAnswer:", + " A1: Propose an one-step thought" + ], + [ + "Question:Which action in the reasoning process is inspired by least-to-most prompting and involves generating sub-questions and their answers?\nAnswer:", + " A2: Propose the remaining thought steps" + ], + [ + "Question:Which action in the reasoning process is inspired by least-to-most prompting and involves generating sub-questions and their answers?\nAnswer:", + " A3: Propose next sub-question along with its answer" + ], + [ + "Question:Which action in the reasoning process is inspired by least-to-most prompting and involves generating sub-questions and their answers?\nAnswer:", + " A5: Rephrase the question/sub-question" + ] + ], + "resps": [ + [ + [ + -41.947181701660156, + false + ] + ], + [ + [ + -44.858089447021484, + false + ] + ], + [ + [ + -43.70464324951172, + false + ] + ], + [ + [ + -39.109134674072266, + false + ] + ] + ], + "filtered_resps": [ + [ + -41.947181701660156, + false + ], + [ + -44.858089447021484, + false + ], + [ + -43.70464324951172, + false + ], + [ + -39.109134674072266, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3ea6706835ff4cef39765d4fbc51408146c9d806c0b39ea78d8336ca66d2f0c0", + "prompt_hash": "4473cd7385465f7f8305ee6cbb2a236cd6ae73493ae23b5dec906975d167b2f7", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 42, + "doc": { + "question": "According to the ablation study, replacing \\( r_2 \\) with random values in RAP causes a significant drop in accuracy on which dataset?", + "choices": [ + "StrategyQA", + "GSM8K", + "MATH-500", + "SVAMP" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:According to the ablation study, replacing \\( r_2 \\) with random values in RAP causes a significant drop in accuracy on which dataset?\nAnswer:", + " StrategyQA" + ], + [ + "Question:According to the ablation study, replacing \\( r_2 \\) with random values in RAP causes a significant drop in accuracy on which dataset?\nAnswer:", + " GSM8K" + ], + [ + "Question:According to the ablation study, replacing \\( r_2 \\) with random values in RAP causes a significant drop in accuracy on which dataset?\nAnswer:", + " MATH-500" + ], + [ + "Question:According to the ablation study, replacing \\( r_2 \\) with random values in RAP causes a significant drop in accuracy on which dataset?\nAnswer:", + " SVAMP" + ] + ], + "resps": [ + [ + [ + -18.641420364379883, + false + ] + ], + [ + [ + -9.388848304748535, + false + ] + ], + [ + [ + -26.169233322143555, + false + ] + ], + [ + [ + -20.590858459472656, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.641420364379883, + false + ], + [ + -9.388848304748535, + false + ], + [ + -26.169233322143555, + false + ], + [ + -20.590858459472656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d54783b122fe68a4da778a99bf5b317c3d85719946a207ee26874b0ebdf376b7", + "prompt_hash": "d3f2d537a02b19810dfd28784b8c1fb5d5a0db3ef651fe26c655f2a64292a1ad", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 43, + "doc": { + "question": "What average number of inferences is required to solve a GSM8K question using the MCTS self-generator on LLaMA2-7B?", + "choices": [ + "148.90", + "166.81", + "348.6k", + "367.1k" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What average number of inferences is required to solve a GSM8K question using the MCTS self-generator on LLaMA2-7B?\nAnswer:", + " 148.90" + ], + [ + "Question:What average number of inferences is required to solve a GSM8K question using the MCTS self-generator on LLaMA2-7B?\nAnswer:", + " 166.81" + ], + [ + "Question:What average number of inferences is required to solve a GSM8K question using the MCTS self-generator on LLaMA2-7B?\nAnswer:", + " 348.6k" + ], + [ + "Question:What average number of inferences is required to solve a GSM8K question using the MCTS self-generator on LLaMA2-7B?\nAnswer:", + " 367.1k" + ] + ], + "resps": [ + [ + [ + -18.378009796142578, + false + ] + ], + [ + [ + -17.80447006225586, + false + ] + ], + [ + [ + -22.77227020263672, + false + ] + ], + [ + [ + -22.721303939819336, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.378009796142578, + false + ], + [ + -17.80447006225586, + false + ], + [ + -22.77227020263672, + false + ], + [ + -22.721303939819336, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3376851ef9784dbe4a9c6462e7e4f78293c0bbdef2ccf6833768c5eb0ba07529", + "prompt_hash": "2266edfa9c35484137149027d1a51807b862ff7d81c9c1a8e6f890426a2a021e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 44, + "doc": { + "question": "Which discriminator model achieved the highest accuracy when verifying LLaMA3-8B-Instruct reasoning trajectories?", + "choices": [ + "LLaMA3.1-8B-Instruct", + "GPT-4 (2024-05-01)", + "Phi3-Mini-Instruct", + "LLaMA3-8B-Instruct" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which discriminator model achieved the highest accuracy when verifying LLaMA3-8B-Instruct reasoning trajectories?\nAnswer:", + " LLaMA3.1-8B-Instruct" + ], + [ + "Question:Which discriminator model achieved the highest accuracy when verifying LLaMA3-8B-Instruct reasoning trajectories?\nAnswer:", + " GPT-4 (2024-05-01)" + ], + [ + "Question:Which discriminator model achieved the highest accuracy when verifying LLaMA3-8B-Instruct reasoning trajectories?\nAnswer:", + " Phi3-Mini-Instruct" + ], + [ + "Question:Which discriminator model achieved the highest accuracy when verifying LLaMA3-8B-Instruct reasoning trajectories?\nAnswer:", + " LLaMA3-8B-Instruct" + ] + ], + "resps": [ + [ + [ + -19.940017700195312, + false + ] + ], + [ + [ + -19.82349395751953, + false + ] + ], + [ + [ + -19.991985321044922, + false + ] + ], + [ + [ + -6.56914758682251, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.940017700195312, + false + ], + [ + -19.82349395751953, + false + ], + [ + -19.991985321044922, + false + ], + [ + -6.56914758682251, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a5fa2750dda4b0a422394d424529a33f963cd403d2e052df741d8830b813cecd", + "prompt_hash": "8de90c8da0c1c343da54a74d5862f70ef62ae9ebc2eb448c91b7095ec982465b", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 45, + "doc": { + "question": "What is the name of the benchmark designed to evaluate long-form writing with outputs over 4,000 words?", + "choices": [ + "LongBench-Write", + "WriteEval-Bench", + "UltraLong-Test", + "GenLength-Benchmark" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the name of the benchmark designed to evaluate long-form writing with outputs over 4,000 words?\nAnswer:", + " LongBench-Write" + ], + [ + "Question:What is the name of the benchmark designed to evaluate long-form writing with outputs over 4,000 words?\nAnswer:", + " WriteEval-Bench" + ], + [ + "Question:What is the name of the benchmark designed to evaluate long-form writing with outputs over 4,000 words?\nAnswer:", + " UltraLong-Test" + ], + [ + "Question:What is the name of the benchmark designed to evaluate long-form writing with outputs over 4,000 words?\nAnswer:", + " GenLength-Benchmark" + ] + ], + "resps": [ + [ + [ + -17.639150619506836, + false + ] + ], + [ + [ + -22.761621475219727, + false + ] + ], + [ + [ + -22.010944366455078, + false + ] + ], + [ + [ + -26.39200210571289, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.639150619506836, + false + ], + [ + -22.761621475219727, + false + ], + [ + -22.010944366455078, + false + ], + [ + -26.39200210571289, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "54fb257c443ac9f24ce1b11083739858fd3f2cdf765fd1910193190b209fade9", + "prompt_hash": "aa259a28a254ba6c18a005278ba04aed52a7d3bfb7946d5685b0f9eabcea9c25", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 46, + "doc": { + "question": "What dataset was constructed using GPT-4o to produce long outputs exceeding 10,000 words?", + "choices": [ + "WriteMaster-10k", + "LongForm-Builder", + "AgentData-6k", + "LongWriter-6k" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What dataset was constructed using GPT-4o to produce long outputs exceeding 10,000 words?\nAnswer:", + " WriteMaster-10k" + ], + [ + "Question:What dataset was constructed using GPT-4o to produce long outputs exceeding 10,000 words?\nAnswer:", + " LongForm-Builder" + ], + [ + "Question:What dataset was constructed using GPT-4o to produce long outputs exceeding 10,000 words?\nAnswer:", + " AgentData-6k" + ], + [ + "Question:What dataset was constructed using GPT-4o to produce long outputs exceeding 10,000 words?\nAnswer:", + " LongWriter-6k" + ] + ], + "resps": [ + [ + [ + -33.696617126464844, + false + ] + ], + [ + [ + -27.476381301879883, + false + ] + ], + [ + [ + -29.436206817626953, + false + ] + ], + [ + [ + -28.958377838134766, + false + ] + ] + ], + "filtered_resps": [ + [ + -33.696617126464844, + false + ], + [ + -27.476381301879883, + false + ], + [ + -29.436206817626953, + false + ], + [ + -28.958377838134766, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "27d094e756d8f9ebc7fb50418d9b28707cfe3d244a12dc577e75c18f11d89e18", + "prompt_hash": "517bc5fb77a00c66e82a92783f2a4c11c8247ed5245f6b23c04fb4d3124ce0dc", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 47, + "doc": { + "question": "What approach does AgentWrite use to generate long-form content?", + "choices": [ + "Retrieval-augmented prompting", + "Divide-and-conquer planning", + "Backtranslation-based expansion", + "Reinforcement learning tuning" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What approach does AgentWrite use to generate long-form content?\nAnswer:", + " Retrieval-augmented prompting" + ], + [ + "Question:What approach does AgentWrite use to generate long-form content?\nAnswer:", + " Divide-and-conquer planning" + ], + [ + "Question:What approach does AgentWrite use to generate long-form content?\nAnswer:", + " Backtranslation-based expansion" + ], + [ + "Question:What approach does AgentWrite use to generate long-form content?\nAnswer:", + " Reinforcement learning tuning" + ] + ], + "resps": [ + [ + [ + -15.73776626586914, + false + ] + ], + [ + [ + -23.26085662841797, + false + ] + ], + [ + [ + -28.372581481933594, + false + ] + ], + [ + [ + -19.8125, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.73776626586914, + false + ], + [ + -23.26085662841797, + false + ], + [ + -28.372581481933594, + false + ], + [ + -19.8125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3c94c38725119f23b11c90a449ac7187d854ac1e0b4570f5f12610bd3bf40f2a", + "prompt_hash": "bdefd51ded190912aa2c257c03204c8892b280aec8dc0411bdb4075b87ebe1cf", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 48, + "doc": { + "question": "What is the typical maximum output length achieved by current long-context models before improvement?", + "choices": [ + "Around 2,000 words", + "Approximately 5,000 words", + "Less than 1,000 words", + "Over 10,000 words" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the typical maximum output length achieved by current long-context models before improvement?\nAnswer:", + " Around 2,000 words" + ], + [ + "Question:What is the typical maximum output length achieved by current long-context models before improvement?\nAnswer:", + " Approximately 5,000 words" + ], + [ + "Question:What is the typical maximum output length achieved by current long-context models before improvement?\nAnswer:", + " Less than 1,000 words" + ], + [ + "Question:What is the typical maximum output length achieved by current long-context models before improvement?\nAnswer:", + " Over 10,000 words" + ] + ], + "resps": [ + [ + [ + -16.012025833129883, + false + ] + ], + [ + [ + -17.129650115966797, + false + ] + ], + [ + [ + -13.902952194213867, + false + ] + ], + [ + [ + -14.598623275756836, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.012025833129883, + false + ], + [ + -17.129650115966797, + false + ], + [ + -13.902952194213867, + false + ], + [ + -14.598623275756836, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "46daa63f9f10a6d83f6a7fa2ea4265bd4a87bc34680ce77a6fcfee59cc290a17", + "prompt_hash": "98cf74434b5cc943c6188659f011b859a46162e99a69d36fd75e3768051df5c8", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 49, + "doc": { + "question": "Which fine-tuning method is used to help the model better follow long writing instructions?", + "choices": [ + "Contrastive Alignment Learning", + "DPO (Direct Preference Optimization)", + "Reinforcement Reward Ranking", + "Instruction Injection Strategy" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which fine-tuning method is used to help the model better follow long writing instructions?\nAnswer:", + " Contrastive Alignment Learning" + ], + [ + "Question:Which fine-tuning method is used to help the model better follow long writing instructions?\nAnswer:", + " DPO (Direct Preference Optimization)" + ], + [ + "Question:Which fine-tuning method is used to help the model better follow long writing instructions?\nAnswer:", + " Reinforcement Reward Ranking" + ], + [ + "Question:Which fine-tuning method is used to help the model better follow long writing instructions?\nAnswer:", + " Instruction Injection Strategy" + ] + ], + "resps": [ + [ + [ + -19.225650787353516, + false + ] + ], + [ + [ + -16.75514793395996, + false + ] + ], + [ + [ + -30.011341094970703, + false + ] + ], + [ + [ + -21.480850219726562, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.225650787353516, + false + ], + [ + -16.75514793395996, + false + ], + [ + -30.011341094970703, + false + ], + [ + -21.480850219726562, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2defad948ec9d397fb72a91e28c7e97a7601e4e89f3e19b32498452725bc6940", + "prompt_hash": "8797c5e2080c8bef6869ed8990e4a53cc5e23ffd3c224c7b33dc177f608eae0c", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 50, + "doc": { + "question": "Which models are cited as examples of frontier Large Language Models?", + "choices": [ + "ChatGPT, Gemini, Opus, LLaMA-3", + "GPT-3.5, Claude, Gemini, Mistral", + "ChatGPT, Palm, BERT, LLaMA-2", + "GPT-4, PaLM-2, Claude, Falcon" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which models are cited as examples of frontier Large Language Models?\nAnswer:", + " ChatGPT, Gemini, Opus, LLaMA-3" + ], + [ + "Question:Which models are cited as examples of frontier Large Language Models?\nAnswer:", + " GPT-3.5, Claude, Gemini, Mistral" + ], + [ + "Question:Which models are cited as examples of frontier Large Language Models?\nAnswer:", + " ChatGPT, Palm, BERT, LLaMA-2" + ], + [ + "Question:Which models are cited as examples of frontier Large Language Models?\nAnswer:", + " GPT-4, PaLM-2, Claude, Falcon" + ] + ], + "resps": [ + [ + [ + -23.581823348999023, + false + ] + ], + [ + [ + -16.29320526123047, + false + ] + ], + [ + [ + -29.092092514038086, + false + ] + ], + [ + [ + -26.421634674072266, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.581823348999023, + false + ], + [ + -16.29320526123047, + false + ], + [ + -29.092092514038086, + false + ], + [ + -26.421634674072266, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f6ada91318d37cb7f383e38593a0e69095e66499aa66b673bcdc84d3a89c40e9", + "prompt_hash": "59bc8556ec178cc8d50dc761afb81b367a36dc4f12592b3bb9608d2bc6942926", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 51, + "doc": { + "question": "Which benchmark is used to evaluate performance in a simulated e-commerce environment?", + "choices": [ + "WebShop", + "WebArena", + "OpenAgent", + "SimWeb" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which benchmark is used to evaluate performance in a simulated e-commerce environment?\nAnswer:", + " WebShop" + ], + [ + "Question:Which benchmark is used to evaluate performance in a simulated e-commerce environment?\nAnswer:", + " WebArena" + ], + [ + "Question:Which benchmark is used to evaluate performance in a simulated e-commerce environment?\nAnswer:", + " OpenAgent" + ], + [ + "Question:Which benchmark is used to evaluate performance in a simulated e-commerce environment?\nAnswer:", + " SimWeb" + ] + ], + "resps": [ + [ + [ + -10.959060668945312, + false + ] + ], + [ + [ + -12.799799919128418, + false + ] + ], + [ + [ + -15.167540550231934, + false + ] + ], + [ + [ + -18.200868606567383, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.959060668945312, + false + ], + [ + -12.799799919128418, + false + ], + [ + -15.167540550231934, + false + ], + [ + -18.200868606567383, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "bcd557127a7ef8d24395fff0b170b5c6f0002615a0db26a1fac8c3f77df7782a", + "prompt_hash": "5923b738e042d2f50011132393f1b705029ab444f06fba66639ea6f14d902829", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 52, + "doc": { + "question": "What was the zero-shot success rate of Agent Q in the real-world booking experiment?", + "choices": [ + "81.7%", + "62.6%", + "67.2%", + "95.4%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What was the zero-shot success rate of Agent Q in the real-world booking experiment?\nAnswer:", + " 81.7%" + ], + [ + "Question:What was the zero-shot success rate of Agent Q in the real-world booking experiment?\nAnswer:", + " 62.6%" + ], + [ + "Question:What was the zero-shot success rate of Agent Q in the real-world booking experiment?\nAnswer:", + " 67.2%" + ], + [ + "Question:What was the zero-shot success rate of Agent Q in the real-world booking experiment?\nAnswer:", + " 95.4%" + ] + ], + "resps": [ + [ + [ + -14.445587158203125, + false + ] + ], + [ + [ + -14.898151397705078, + false + ] + ], + [ + [ + -14.475791931152344, + false + ] + ], + [ + [ + -14.124256134033203, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.445587158203125, + false + ], + [ + -14.898151397705078, + false + ], + [ + -14.475791931152344, + false + ], + [ + -14.124256134033203, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9b124c03b2602c4e92e8adcbfdb16a4ad8c364b3c126fb59a84f65b6cef4a1fc", + "prompt_hash": "0a95943469f8ba1096d26ad4f4a42c731bf57e1a3ca93c33dff45a879ee03016", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 53, + "doc": { + "question": "Which algorithm is used to optimize the policy in the proposed training method?", + "choices": [ + "Direct Preference Optimization (DPO)", + "Proximal Policy Optimization (PPO)", + "Q-learning", + "Actor-Critic" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which algorithm is used to optimize the policy in the proposed training method?\nAnswer:", + " Direct Preference Optimization (DPO)" + ], + [ + "Question:Which algorithm is used to optimize the policy in the proposed training method?\nAnswer:", + " Proximal Policy Optimization (PPO)" + ], + [ + "Question:Which algorithm is used to optimize the policy in the proposed training method?\nAnswer:", + " Q-learning" + ], + [ + "Question:Which algorithm is used to optimize the policy in the proposed training method?\nAnswer:", + " Actor-Critic" + ] + ], + "resps": [ + [ + [ + -10.699943542480469, + false + ] + ], + [ + [ + -8.01321792602539, + false + ] + ], + [ + [ + -4.749229907989502, + false + ] + ], + [ + [ + -8.794071197509766, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.699943542480469, + false + ], + [ + -8.01321792602539, + false + ], + [ + -4.749229907989502, + false + ], + [ + -8.794071197509766, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7a8ffac9727722acc7e16ea4e31c950d04a2414d310d30d10782400f58c19526", + "prompt_hash": "929eea14da123173493f83ba8cafdaed60dab174680214e80fedcdf45aef33b5", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 54, + "doc": { + "question": "What is the average number of steps required to complete a task in the OpenTable environment?", + "choices": [ + "13.9", + "6.8", + "20.3", + "10.5" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the average number of steps required to complete a task in the OpenTable environment?\nAnswer:", + " 13.9" + ], + [ + "Question:What is the average number of steps required to complete a task in the OpenTable environment?\nAnswer:", + " 6.8" + ], + [ + "Question:What is the average number of steps required to complete a task in the OpenTable environment?\nAnswer:", + " 20.3" + ], + [ + "Question:What is the average number of steps required to complete a task in the OpenTable environment?\nAnswer:", + " 10.5" + ] + ], + "resps": [ + [ + [ + -9.529898643493652, + false + ] + ], + [ + [ + -8.548165321350098, + false + ] + ], + [ + [ + -9.393954277038574, + false + ] + ], + [ + [ + -8.277305603027344, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.529898643493652, + false + ], + [ + -8.548165321350098, + false + ], + [ + -9.393954277038574, + false + ], + [ + -8.277305603027344, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "661302f971fbae21fe155e9a5cf88b752f5af1e05fbb2c9bf77ac54642151bdd", + "prompt_hash": "1d55c170c55f70a6d6dbff3674955aab12ce4afd95c6b5036e2d432fe92c0d73", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 55, + "doc": { + "question": "What is the pass@128 accuracy of the RL model using CoT prompting on the miniF2F-test dataset?", + "choices": [ + "51.6%", + "49.8%", + "50.4%", + "50.5%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the pass@128 accuracy of the RL model using CoT prompting on the miniF2F-test dataset?\nAnswer:", + " 51.6%" + ], + [ + "Question:What is the pass@128 accuracy of the RL model using CoT prompting on the miniF2F-test dataset?\nAnswer:", + " 49.8%" + ], + [ + "Question:What is the pass@128 accuracy of the RL model using CoT prompting on the miniF2F-test dataset?\nAnswer:", + " 50.4%" + ], + [ + "Question:What is the pass@128 accuracy of the RL model using CoT prompting on the miniF2F-test dataset?\nAnswer:", + " 50.5%" + ] + ], + "resps": [ + [ + [ + -14.946520805358887, + false + ] + ], + [ + [ + -15.035083770751953, + false + ] + ], + [ + [ + -14.563060760498047, + false + ] + ], + [ + [ + -14.588592529296875, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.946520805358887, + false + ], + [ + -15.035083770751953, + false + ], + [ + -14.563060760498047, + false + ], + [ + -14.588592529296875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "bea78ff6e30c23805612ce89fc7e2cefaea9962704630da8fc21a1edb3e75a45", + "prompt_hash": "31568cdaed8248fb9347dd0f50d2e6a40ffe00ebf9f77248d7a9fe570922cde2", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 56, + "doc": { + "question": "What reward is assigned to a generated proof if it is verified as correct by the Lean prover?", + "choices": [ + "1", + "0", + "0.5", + "2" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What reward is assigned to a generated proof if it is verified as correct by the Lean prover?\nAnswer:", + " 1" + ], + [ + "Question:What reward is assigned to a generated proof if it is verified as correct by the Lean prover?\nAnswer:", + " 0" + ], + [ + "Question:What reward is assigned to a generated proof if it is verified as correct by the Lean prover?\nAnswer:", + " 0.5" + ], + [ + "Question:What reward is assigned to a generated proof if it is verified as correct by the Lean prover?\nAnswer:", + " 2" + ] + ], + "resps": [ + [ + [ + -4.328453540802002, + false + ] + ], + [ + [ + -4.019128322601318, + false + ] + ], + [ + [ + -6.707319259643555, + false + ] + ], + [ + [ + -5.582779407501221, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.328453540802002, + false + ], + [ + -4.019128322601318, + false + ], + [ + -6.707319259643555, + false + ], + [ + -5.582779407501221, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "81c2eecd521ab20a30312484726d6d4385d46d1dd5c6e3b646d62c95390c4d1e", + "prompt_hash": "db6dff08fbd10cab9630fb8848105787301cae9945b45dfc78f32ac6c943c813", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 57, + "doc": { + "question": "What guiding information is added to the prompt during supervised fine-tuning to aid proof completion?", + "choices": [ + "Intermediate tactic state as a comment block", + "Problem difficulty level", + "Proof hints from humans", + "External symbolic solvers" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What guiding information is added to the prompt during supervised fine-tuning to aid proof completion?\nAnswer:", + " Intermediate tactic state as a comment block" + ], + [ + "Question:What guiding information is added to the prompt during supervised fine-tuning to aid proof completion?\nAnswer:", + " Problem difficulty level" + ], + [ + "Question:What guiding information is added to the prompt during supervised fine-tuning to aid proof completion?\nAnswer:", + " Proof hints from humans" + ], + [ + "Question:What guiding information is added to the prompt during supervised fine-tuning to aid proof completion?\nAnswer:", + " External symbolic solvers" + ] + ], + "resps": [ + [ + [ + -48.464988708496094, + false + ] + ], + [ + [ + -23.567293167114258, + false + ] + ], + [ + [ + -25.195629119873047, + false + ] + ], + [ + [ + -31.9509220123291, + false + ] + ] + ], + "filtered_resps": [ + [ + -48.464988708496094, + false + ], + [ + -23.567293167114258, + false + ], + [ + -25.195629119873047, + false + ], + [ + -31.9509220123291, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3bd462c4b12de939d731deec2d9c2b24baecef0a3bc3cad2463612293aad9bbd", + "prompt_hash": "7ab890996c46750808a93880844c85ebc90272d8217bc68c896abac71c3d1c7b", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 58, + "doc": { + "question": "What dataset is used to evaluate high-school level theorem-proving performance?", + "choices": [ + "miniF2F", + "ProofNet", + "LeanDojo", + "MATH" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What dataset is used to evaluate high-school level theorem-proving performance?\nAnswer:", + " miniF2F" + ], + [ + "Question:What dataset is used to evaluate high-school level theorem-proving performance?\nAnswer:", + " ProofNet" + ], + [ + "Question:What dataset is used to evaluate high-school level theorem-proving performance?\nAnswer:", + " LeanDojo" + ], + [ + "Question:What dataset is used to evaluate high-school level theorem-proving performance?\nAnswer:", + " MATH" + ] + ], + "resps": [ + [ + [ + -24.148351669311523, + false + ] + ], + [ + [ + -10.6226167678833, + false + ] + ], + [ + [ + -20.867656707763672, + false + ] + ], + [ + [ + -7.023367881774902, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.148351669311523, + false + ], + [ + -10.6226167678833, + false + ], + [ + -20.867656707763672, + false + ], + [ + -7.023367881774902, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "66a9544783bca91ed0271a51b7911425a7e817ac41cfc9bbf1b22f97967d7f97", + "prompt_hash": "f12755e00fb283a488fd026d2cdba0e31b3fb40c906e349ed7a9b4dd052e4296", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 59, + "doc": { + "question": "What is the purpose of the 'truncate-and-resume' mechanism in the tree search method?", + "choices": [ + "To continue proof generation from the last verified step", + "To minimize inference time", + "To switch to a different model mid-proof", + "To discard all failed proof attempts" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the purpose of the 'truncate-and-resume' mechanism in the tree search method?\nAnswer:", + " To continue proof generation from the last verified step" + ], + [ + "Question:What is the purpose of the 'truncate-and-resume' mechanism in the tree search method?\nAnswer:", + " To minimize inference time" + ], + [ + "Question:What is the purpose of the 'truncate-and-resume' mechanism in the tree search method?\nAnswer:", + " To switch to a different model mid-proof" + ], + [ + "Question:What is the purpose of the 'truncate-and-resume' mechanism in the tree search method?\nAnswer:", + " To discard all failed proof attempts" + ] + ], + "resps": [ + [ + [ + -39.913795471191406, + false + ] + ], + [ + [ + -18.19488525390625, + false + ] + ], + [ + [ + -40.979583740234375, + false + ] + ], + [ + [ + -31.746326446533203, + false + ] + ] + ], + "filtered_resps": [ + [ + -39.913795471191406, + false + ], + [ + -18.19488525390625, + false + ], + [ + -40.979583740234375, + false + ], + [ + -31.746326446533203, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a5a46607c4afedd35a65857b77845b884203f0330d97aed61bcf885dd9c7fe0b", + "prompt_hash": "315a3809a2d4007dfeed2413ac84811418552960e915f01cb771ab3031e6fcd1", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 60, + "doc": { + "question": "What is the primary evaluation metric for Track 1 in ASVspoof~5?", + "choices": [ + "Equal Error Rate (EER)", + "Tandem Detection Cost Function (t-DCF)", + "Minimum Detection Cost Function (minDCF)", + "Architecture-Agnostic Detection Cost Function (a-DCF)" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the primary evaluation metric for Track 1 in ASVspoof~5?\nAnswer:", + " Equal Error Rate (EER)" + ], + [ + "Question:What is the primary evaluation metric for Track 1 in ASVspoof~5?\nAnswer:", + " Tandem Detection Cost Function (t-DCF)" + ], + [ + "Question:What is the primary evaluation metric for Track 1 in ASVspoof~5?\nAnswer:", + " Minimum Detection Cost Function (minDCF)" + ], + [ + "Question:What is the primary evaluation metric for Track 1 in ASVspoof~5?\nAnswer:", + " Architecture-Agnostic Detection Cost Function (a-DCF)" + ] + ], + "resps": [ + [ + [ + -8.293449401855469, + false + ] + ], + [ + [ + -31.978302001953125, + false + ] + ], + [ + [ + -17.47964096069336, + false + ] + ], + [ + [ + -41.507606506347656, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.293449401855469, + false + ], + [ + -31.978302001953125, + false + ], + [ + -17.47964096069336, + false + ], + [ + -41.507606506347656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "95d3ba17b14751019214fa4ae9518138be8ff4a6b03f3e88ff5cbe369c6bb78d", + "prompt_hash": "2990d2f55d6dc5185802979087c79d0697bcbbe40d9deeb2f34554bb1ef0c0d8", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 61, + "doc": { + "question": "Which dataset is used as the source for ASVspoof~5 database?", + "choices": [ + "VCTK Corpus", + "Librispeech Clean", + "CommonVoice", + "Multilingual Librispeech (MLS) English partition" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which dataset is used as the source for ASVspoof~5 database?\nAnswer:", + " VCTK Corpus" + ], + [ + "Question:Which dataset is used as the source for ASVspoof~5 database?\nAnswer:", + " Librispeech Clean" + ], + [ + "Question:Which dataset is used as the source for ASVspoof~5 database?\nAnswer:", + " CommonVoice" + ], + [ + "Question:Which dataset is used as the source for ASVspoof~5 database?\nAnswer:", + " Multilingual Librispeech (MLS) English partition" + ] + ], + "resps": [ + [ + [ + -14.687994003295898, + false + ] + ], + [ + [ + -21.001323699951172, + false + ] + ], + [ + [ + -10.582550048828125, + false + ] + ], + [ + [ + -39.72848129272461, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.687994003295898, + false + ], + [ + -21.001323699951172, + false + ], + [ + -10.582550048828125, + false + ], + [ + -39.72848129272461, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5bf324f6272f99e64abdb10484dbd372127010afacd962840086cabdbc778e16", + "prompt_hash": "2330b3711bd77be2ac52843201f12f5d6d53f885329c0cf9fae3e7365cb83525", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 62, + "doc": { + "question": "Which two pre-trained systems are used in ASVspoof~5 to perform zero-shot voice cloning?", + "choices": [ + "Tacotron2 and WaveGlow", + "YourTTS and XTTS", + "FastSpeech and DeepVoice", + "Glow-TTS and MelGAN" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which two pre-trained systems are used in ASVspoof~5 to perform zero-shot voice cloning?\nAnswer:", + " Tacotron2 and WaveGlow" + ], + [ + "Question:Which two pre-trained systems are used in ASVspoof~5 to perform zero-shot voice cloning?\nAnswer:", + " YourTTS and XTTS" + ], + [ + "Question:Which two pre-trained systems are used in ASVspoof~5 to perform zero-shot voice cloning?\nAnswer:", + " FastSpeech and DeepVoice" + ], + [ + "Question:Which two pre-trained systems are used in ASVspoof~5 to perform zero-shot voice cloning?\nAnswer:", + " Glow-TTS and MelGAN" + ] + ], + "resps": [ + [ + [ + -23.865249633789062, + false + ] + ], + [ + [ + -27.683504104614258, + false + ] + ], + [ + [ + -19.734195709228516, + false + ] + ], + [ + [ + -25.660432815551758, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.865249633789062, + false + ], + [ + -27.683504104614258, + false + ], + [ + -19.734195709228516, + false + ], + [ + -25.660432815551758, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2db558857d21d05a1c8a6fa58aeff22e54d375701abcb3140367b8bee946e24a", + "prompt_hash": "6389dcbde532b910b8790d57751c0cdcf9247572d9f7ca74529c8c6a47fb7287", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 63, + "doc": { + "question": "What is the speaker encoder used in the common ASV system provided by ASVspoof~5 organisers?", + "choices": [ + "TDNN-F", + "ECAPA-TDNN", + "Conformer-XL", + "X-Vector" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the speaker encoder used in the common ASV system provided by ASVspoof~5 organisers?\nAnswer:", + " TDNN-F" + ], + [ + "Question:What is the speaker encoder used in the common ASV system provided by ASVspoof~5 organisers?\nAnswer:", + " ECAPA-TDNN" + ], + [ + "Question:What is the speaker encoder used in the common ASV system provided by ASVspoof~5 organisers?\nAnswer:", + " Conformer-XL" + ], + [ + "Question:What is the speaker encoder used in the common ASV system provided by ASVspoof~5 organisers?\nAnswer:", + " X-Vector" + ] + ], + "resps": [ + [ + [ + -15.269363403320312, + false + ] + ], + [ + [ + -7.209596157073975, + false + ] + ], + [ + [ + -12.833272933959961, + false + ] + ], + [ + [ + -11.507760047912598, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.269363403320312, + false + ], + [ + -7.209596157073975, + false + ], + [ + -12.833272933959961, + false + ], + [ + -11.507760047912598, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4617e752ea45a441df277dfe97cdfbe02b4ce00b0138647f2860fe561600d97e", + "prompt_hash": "ccf9aa7d197663dcc62b2793d0c6bae5b70d0fda8efc656d19f5f3a46002e3a5", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 64, + "doc": { + "question": "Which attack filter introduced in ASVspoof~5 escalates the threat of spoofed data to ASV systems?", + "choices": [ + "Malafide", + "Malacopula", + "MFA-Conformer", + "ZMM-TTS" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which attack filter introduced in ASVspoof~5 escalates the threat of spoofed data to ASV systems?\nAnswer:", + " Malafide" + ], + [ + "Question:Which attack filter introduced in ASVspoof~5 escalates the threat of spoofed data to ASV systems?\nAnswer:", + " Malacopula" + ], + [ + "Question:Which attack filter introduced in ASVspoof~5 escalates the threat of spoofed data to ASV systems?\nAnswer:", + " MFA-Conformer" + ], + [ + "Question:Which attack filter introduced in ASVspoof~5 escalates the threat of spoofed data to ASV systems?\nAnswer:", + " ZMM-TTS" + ] + ], + "resps": [ + [ + [ + -19.35630226135254, + false + ] + ], + [ + [ + -30.244190216064453, + false + ] + ], + [ + [ + -26.727115631103516, + false + ] + ], + [ + [ + -24.757612228393555, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.35630226135254, + false + ], + [ + -30.244190216064453, + false + ], + [ + -26.727115631103516, + false + ], + [ + -24.757612228393555, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0fb2458d508f2d9147675e4363f0942d600d225d663c5fc2ee26073611a48489", + "prompt_hash": "328fce83d3ace1b2ab7bf92fb975d199a432f841f1ee570bc90ade3a388e2066", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 65, + "doc": { + "question": "What is the loss function used to train text tokens in Transfusion?", + "choices": [ + "Contrastive loss", + "Perplexity loss", + "Next-token prediction (LM loss)", + "Reconstruction loss" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the loss function used to train text tokens in Transfusion?\nAnswer:", + " Contrastive loss" + ], + [ + "Question:What is the loss function used to train text tokens in Transfusion?\nAnswer:", + " Perplexity loss" + ], + [ + "Question:What is the loss function used to train text tokens in Transfusion?\nAnswer:", + " Next-token prediction (LM loss)" + ], + [ + "Question:What is the loss function used to train text tokens in Transfusion?\nAnswer:", + " Reconstruction loss" + ] + ], + "resps": [ + [ + [ + -7.832241058349609, + false + ] + ], + [ + [ + -10.432533264160156, + false + ] + ], + [ + [ + -28.241981506347656, + false + ] + ], + [ + [ + -10.100059509277344, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.832241058349609, + false + ], + [ + -10.432533264160156, + false + ], + [ + -28.241981506347656, + false + ], + [ + -10.100059509277344, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6b0ace319fe15fa9f5d0bf46932d99e5bd2793db908e2c5b2f2bb258b28f6072", + "prompt_hash": "570aa441db2093a1b3421527b81e83562db6be29dde0a696b673b704491a7ddb", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 66, + "doc": { + "question": "Which patch encoding architecture improves Transfusion’s image generation performance the most?", + "choices": [ + "Linear layer with attention", + "U-Net with AdaLayerNorm", + "U-Net with regular layer norm", + "Convolutional GRU blocks" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which patch encoding architecture improves Transfusion’s image generation performance the most?\nAnswer:", + " Linear layer with attention" + ], + [ + "Question:Which patch encoding architecture improves Transfusion’s image generation performance the most?\nAnswer:", + " U-Net with AdaLayerNorm" + ], + [ + "Question:Which patch encoding architecture improves Transfusion’s image generation performance the most?\nAnswer:", + " U-Net with regular layer norm" + ], + [ + "Question:Which patch encoding architecture improves Transfusion’s image generation performance the most?\nAnswer:", + " Convolutional GRU blocks" + ] + ], + "resps": [ + [ + [ + -18.62102508544922, + false + ] + ], + [ + [ + -32.28581237792969, + false + ] + ], + [ + [ + -31.470565795898438, + false + ] + ], + [ + [ + -27.30034637451172, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.62102508544922, + false + ], + [ + -32.28581237792969, + false + ], + [ + -31.470565795898438, + false + ], + [ + -27.30034637451172, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "29cae45755b78d5aedaa01332ccf6118f4c9b014417563694179a1960d6fc7e9", + "prompt_hash": "a1c19527a1165a8e5d2900fbc716ef9b0a394aaa0a23aef8dc8bcaca1f488f21", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 67, + "doc": { + "question": "What is the purpose of the BOI and EOI tokens in Transfusion?", + "choices": [ + "To indicate the boundaries of text sequences", + "To mark the beginning and end of the entire multi-modal document", + "To separate images from each other during training", + "To surround image sequences in mixed-modal examples" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the purpose of the BOI and EOI tokens in Transfusion?\nAnswer:", + " To indicate the boundaries of text sequences" + ], + [ + "Question:What is the purpose of the BOI and EOI tokens in Transfusion?\nAnswer:", + " To mark the beginning and end of the entire multi-modal document" + ], + [ + "Question:What is the purpose of the BOI and EOI tokens in Transfusion?\nAnswer:", + " To separate images from each other during training" + ], + [ + "Question:What is the purpose of the BOI and EOI tokens in Transfusion?\nAnswer:", + " To surround image sequences in mixed-modal examples" + ] + ], + "resps": [ + [ + [ + -29.421863555908203, + false + ] + ], + [ + [ + -33.25529861450195, + false + ] + ], + [ + [ + -27.936172485351562, + false + ] + ], + [ + [ + -54.490135192871094, + false + ] + ] + ], + "filtered_resps": [ + [ + -29.421863555908203, + false + ], + [ + -33.25529861450195, + false + ], + [ + -27.936172485351562, + false + ], + [ + -54.490135192871094, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2caec057216248e225a486a5981b79b5095d183fd431a4d2e6425a38484e4039", + "prompt_hash": "91cb037498b2189102728567418847c093f76af545eaaed3fd616858c3e8827f", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 68, + "doc": { + "question": "What metric is used to evaluate text-to-image alignment in Transfusion?", + "choices": [ + "BLEU score", + "FID score", + "CLIP score", + "ROUGE-L score" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What metric is used to evaluate text-to-image alignment in Transfusion?\nAnswer:", + " BLEU score" + ], + [ + "Question:What metric is used to evaluate text-to-image alignment in Transfusion?\nAnswer:", + " FID score" + ], + [ + "Question:What metric is used to evaluate text-to-image alignment in Transfusion?\nAnswer:", + " CLIP score" + ], + [ + "Question:What metric is used to evaluate text-to-image alignment in Transfusion?\nAnswer:", + " ROUGE-L score" + ] + ], + "resps": [ + [ + [ + -9.753241539001465, + false + ] + ], + [ + [ + -8.78629207611084, + false + ] + ], + [ + [ + -6.556405067443848, + false + ] + ], + [ + [ + -14.04121208190918, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.753241539001465, + false + ], + [ + -8.78629207611084, + false + ], + [ + -6.556405067443848, + false + ], + [ + -14.04121208190918, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b521b9c06830e960b02ae0ea71ffdfa87e7339a23c2d6484253ad1161c0301cb", + "prompt_hash": "ce159dbcd9e01068e8c6c93f0f5b4e4a654718f7d8ad80f9ba8381063ac2c2bd", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 69, + "doc": { + "question": "Which dataset provides the 2T tokens used for Transfusion's large-scale training?", + "choices": [ + "LAION-5B and C4", + "Wikipedia and ImageNet", + "Llama 2 corpus and Shutterstock + Conceptual 12M", + "OpenWebText and MS-COCO" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which dataset provides the 2T tokens used for Transfusion's large-scale training?\nAnswer:", + " LAION-5B and C4" + ], + [ + "Question:Which dataset provides the 2T tokens used for Transfusion's large-scale training?\nAnswer:", + " Wikipedia and ImageNet" + ], + [ + "Question:Which dataset provides the 2T tokens used for Transfusion's large-scale training?\nAnswer:", + " Llama 2 corpus and Shutterstock + Conceptual 12M" + ], + [ + "Question:Which dataset provides the 2T tokens used for Transfusion's large-scale training?\nAnswer:", + " OpenWebText and MS-COCO" + ] + ], + "resps": [ + [ + [ + -18.460792541503906, + false + ] + ], + [ + [ + -15.87649917602539, + false + ] + ], + [ + [ + -65.93722534179688, + false + ] + ], + [ + [ + -22.97542953491211, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.460792541503906, + false + ], + [ + -15.87649917602539, + false + ], + [ + -65.93722534179688, + false + ], + [ + -22.97542953491211, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ce183691522901f523c3138a5c649a02dbdbd4008d6c70d0596b60660ff8963f", + "prompt_hash": "7a61699516049970699cc38aeb4d62adb1611a28704277f83b5ad9d41bdf14f9", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 70, + "doc": { + "question": "What key feature distinguishes Kolmogorov-Arnold Networks (KANs) from multilayer perceptrons (MLPs)?", + "choices": [ + "KANs use learnable activation functions on edges", + "KANs have fixed activation functions on nodes", + "KANs use convolutional layers instead of fully connected layers", + "KANs rely on recurrent architectures for interpretability" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What key feature distinguishes Kolmogorov-Arnold Networks (KANs) from multilayer perceptrons (MLPs)?\nAnswer:", + " KANs use learnable activation functions on edges" + ], + [ + "Question:What key feature distinguishes Kolmogorov-Arnold Networks (KANs) from multilayer perceptrons (MLPs)?\nAnswer:", + " KANs have fixed activation functions on nodes" + ], + [ + "Question:What key feature distinguishes Kolmogorov-Arnold Networks (KANs) from multilayer perceptrons (MLPs)?\nAnswer:", + " KANs use convolutional layers instead of fully connected layers" + ], + [ + "Question:What key feature distinguishes Kolmogorov-Arnold Networks (KANs) from multilayer perceptrons (MLPs)?\nAnswer:", + " KANs rely on recurrent architectures for interpretability" + ] + ], + "resps": [ + [ + [ + -24.835132598876953, + false + ] + ], + [ + [ + -25.078359603881836, + false + ] + ], + [ + [ + -14.701416015625, + false + ] + ], + [ + [ + -32.338748931884766, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.835132598876953, + false + ], + [ + -25.078359603881836, + false + ], + [ + -14.701416015625, + false + ], + [ + -32.338748931884766, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2d5acfa87e3ab29480130d3e939c998ab2e76f35cb68b702fb4f7e3bf7909d32", + "prompt_hash": "aed15bb52a7728777f97d79fb23a7e5e41c96fedefd5d574c8fa3267850a078b", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 71, + "doc": { + "question": "What is the primary enhancement that MultKAN introduces over the original KAN?", + "choices": [ + "It introduces explicit multiplication operations between subnodes", + "It replaces all activation functions with ReLU", + "It allows recurrent feedback connections", + "It removes all addition nodes to reduce complexity" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the primary enhancement that MultKAN introduces over the original KAN?\nAnswer:", + " It introduces explicit multiplication operations between subnodes" + ], + [ + "Question:What is the primary enhancement that MultKAN introduces over the original KAN?\nAnswer:", + " It replaces all activation functions with ReLU" + ], + [ + "Question:What is the primary enhancement that MultKAN introduces over the original KAN?\nAnswer:", + " It allows recurrent feedback connections" + ], + [ + "Question:What is the primary enhancement that MultKAN introduces over the original KAN?\nAnswer:", + " It removes all addition nodes to reduce complexity" + ] + ], + "resps": [ + [ + [ + -34.7784538269043, + false + ] + ], + [ + [ + -20.881629943847656, + false + ] + ], + [ + [ + -27.06278419494629, + false + ] + ], + [ + [ + -40.277320861816406, + false + ] + ] + ], + "filtered_resps": [ + [ + -34.7784538269043, + false + ], + [ + -20.881629943847656, + false + ], + [ + -27.06278419494629, + false + ], + [ + -40.277320861816406, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fb27a0f9ec84e9f94e9b94583d80b07de729f25497b2c72e4719f5ff0483d5be", + "prompt_hash": "f0c4cbb2d6ce4e80ffa1c1860fc18807ddf434f9b562da0fe5cd3e73656c49fa", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 72, + "doc": { + "question": "What is the purpose of the 'kanpiler' tool introduced in the paper?", + "choices": [ + "To convert symbolic formulas into KAN architectures", + "To extract symbolic expressions from trained KANs", + "To optimize KAN performance through weight pruning", + "To compress large KANs into smaller versions" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the purpose of the 'kanpiler' tool introduced in the paper?\nAnswer:", + " To convert symbolic formulas into KAN architectures" + ], + [ + "Question:What is the purpose of the 'kanpiler' tool introduced in the paper?\nAnswer:", + " To extract symbolic expressions from trained KANs" + ], + [ + "Question:What is the purpose of the 'kanpiler' tool introduced in the paper?\nAnswer:", + " To optimize KAN performance through weight pruning" + ], + [ + "Question:What is the purpose of the 'kanpiler' tool introduced in the paper?\nAnswer:", + " To compress large KANs into smaller versions" + ] + ], + "resps": [ + [ + [ + -37.53588104248047, + false + ] + ], + [ + [ + -38.3877067565918, + false + ] + ], + [ + [ + -34.88467025756836, + false + ] + ], + [ + [ + -30.9049072265625, + false + ] + ] + ], + "filtered_resps": [ + [ + -37.53588104248047, + false + ], + [ + -38.3877067565918, + false + ], + [ + -34.88467025756836, + false + ], + [ + -30.9049072265625, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "aaeac0339e72fc575d316c42e2c96a6e83d98979c63d77b1374b1e94e07143a5", + "prompt_hash": "ca2994149774fcb47bcb0098296ea6473f8ad6dbe711695fccf9a6c6044922bc", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 73, + "doc": { + "question": "Which method is introduced to assign scores to input variables in KANs to reflect their importance?", + "choices": [ + "Attribution score", + "Gradient clipping", + "Dropout rate", + "Batch normalization" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which method is introduced to assign scores to input variables in KANs to reflect their importance?\nAnswer:", + " Attribution score" + ], + [ + "Question:Which method is introduced to assign scores to input variables in KANs to reflect their importance?\nAnswer:", + " Gradient clipping" + ], + [ + "Question:Which method is introduced to assign scores to input variables in KANs to reflect their importance?\nAnswer:", + " Dropout rate" + ], + [ + "Question:Which method is introduced to assign scores to input variables in KANs to reflect their importance?\nAnswer:", + " Batch normalization" + ] + ], + "resps": [ + [ + [ + -14.220232963562012, + false + ] + ], + [ + [ + -13.326740264892578, + false + ] + ], + [ + [ + -16.06420135498047, + false + ] + ], + [ + [ + -12.452866554260254, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.220232963562012, + false + ], + [ + -13.326740264892578, + false + ], + [ + -16.06420135498047, + false + ], + [ + -12.452866554260254, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f2198d6262fe8ca5c6bec0aa07d9085fe88f229976cde0456857cb3a842d1e58", + "prompt_hash": "2f05faff3fb35d1886cbf3237ae1313cbf1f731bd8403d103c36c10ca9cd99e0", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 74, + "doc": { + "question": "What property must a function satisfy to be considered additively separable?", + "choices": [ + "It can be expressed as the sum of functions over non-overlapping variable groups", + "It is invariant under any permutation of its variables", + "It only involves multiplicative combinations of inputs", + "It satisfies the condition of zero second derivatives across all input variables" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What property must a function satisfy to be considered additively separable?\nAnswer:", + " It can be expressed as the sum of functions over non-overlapping variable groups" + ], + [ + "Question:What property must a function satisfy to be considered additively separable?\nAnswer:", + " It is invariant under any permutation of its variables" + ], + [ + "Question:What property must a function satisfy to be considered additively separable?\nAnswer:", + " It only involves multiplicative combinations of inputs" + ], + [ + "Question:What property must a function satisfy to be considered additively separable?\nAnswer:", + " It satisfies the condition of zero second derivatives across all input variables" + ] + ], + "resps": [ + [ + [ + -51.171424865722656, + false + ] + ], + [ + [ + -22.00145721435547, + false + ] + ], + [ + [ + -32.46794128417969, + false + ] + ], + [ + [ + -46.33948516845703, + false + ] + ] + ], + "filtered_resps": [ + [ + -51.171424865722656, + false + ], + [ + -22.00145721435547, + false + ], + [ + -32.46794128417969, + false + ], + [ + -46.33948516845703, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ca1ded857af8cbbce2e7be004360e511f4a4b7dab07917fcbbec93bc80290b7c", + "prompt_hash": "43d2ceabec8b8b54d9da9df9ae1d9e73050475a31ca8789daf1e6e1f3a5074a5", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 75, + "doc": { + "question": "What is the total number of A100 GPUs used to pretrain the Sapiens-2B model?", + "choices": [ + "512", + "768", + "1024", + "2048" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the total number of A100 GPUs used to pretrain the Sapiens-2B model?\nAnswer:", + " 512" + ], + [ + "Question:What is the total number of A100 GPUs used to pretrain the Sapiens-2B model?\nAnswer:", + " 768" + ], + [ + "Question:What is the total number of A100 GPUs used to pretrain the Sapiens-2B model?\nAnswer:", + " 1024" + ], + [ + "Question:What is the total number of A100 GPUs used to pretrain the Sapiens-2B model?\nAnswer:", + " 2048" + ] + ], + "resps": [ + [ + [ + -7.625455379486084, + false + ] + ], + [ + [ + -7.575810432434082, + false + ] + ], + [ + [ + -7.642750263214111, + false + ] + ], + [ + [ + -8.01429557800293, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.625455379486084, + false + ], + [ + -7.575810432434082, + false + ], + [ + -7.642750263214111, + false + ], + [ + -8.01429557800293, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "528dd299433b3d911f39a72b35f638ff4c8a8cedde10d8c3ac216018203cac9e", + "prompt_hash": "0789a74cd22126b5cbae6c1be7e15e03778db286e3bf042ec9dd5b0bd6b34b50", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 76, + "doc": { + "question": "What is the masking ratio used during pretraining in the masked autoencoder setup?", + "choices": [ + "75%", + "80%", + "90%", + "95%" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the masking ratio used during pretraining in the masked autoencoder setup?\nAnswer:", + " 75%" + ], + [ + "Question:What is the masking ratio used during pretraining in the masked autoencoder setup?\nAnswer:", + " 80%" + ], + [ + "Question:What is the masking ratio used during pretraining in the masked autoencoder setup?\nAnswer:", + " 90%" + ], + [ + "Question:What is the masking ratio used during pretraining in the masked autoencoder setup?\nAnswer:", + " 95%" + ] + ], + "resps": [ + [ + [ + -10.97347354888916, + false + ] + ], + [ + [ + -9.670283317565918, + false + ] + ], + [ + [ + -10.75654125213623, + false + ] + ], + [ + [ + -11.948522567749023, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.97347354888916, + false + ], + [ + -9.670283317565918, + false + ], + [ + -10.75654125213623, + false + ], + [ + -11.948522567749023, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e8382292752998f36fe4788c6504c9081c1414f2d0968a303ced787a1a38c768", + "prompt_hash": "27d9f7078490b4956caea33002cad3f1b07ad508db4129edf312f78eb525cf18", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 77, + "doc": { + "question": "How many facial keypoints are included in the new detailed skeleton used for 2D pose estimation?", + "choices": [ + "68", + "133", + "243", + "308" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many facial keypoints are included in the new detailed skeleton used for 2D pose estimation?\nAnswer:", + " 68" + ], + [ + "Question:How many facial keypoints are included in the new detailed skeleton used for 2D pose estimation?\nAnswer:", + " 133" + ], + [ + "Question:How many facial keypoints are included in the new detailed skeleton used for 2D pose estimation?\nAnswer:", + " 243" + ], + [ + "Question:How many facial keypoints are included in the new detailed skeleton used for 2D pose estimation?\nAnswer:", + " 308" + ] + ], + "resps": [ + [ + [ + -6.453620910644531, + false + ] + ], + [ + [ + -8.351236343383789, + false + ] + ], + [ + [ + -9.181026458740234, + false + ] + ], + [ + [ + -8.84392261505127, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.453620910644531, + false + ], + [ + -8.351236343383789, + false + ], + [ + -9.181026458740234, + false + ], + [ + -8.84392261505127, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "dfd9b997c52aa078ec356428d1c4511a2bad04a2dfd3edfccbf0978417918deb", + "prompt_hash": "108d87a3f6ded0b18c149bed498682e004bbe4a51516edc01fd171d7b90b72df", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 78, + "doc": { + "question": "How many synthetic images are rendered for training the depth estimation model?", + "choices": [ + "100,000", + "250,000", + "500,000", + "1,000,000" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many synthetic images are rendered for training the depth estimation model?\nAnswer:", + " 100,000" + ], + [ + "Question:How many synthetic images are rendered for training the depth estimation model?\nAnswer:", + " 250,000" + ], + [ + "Question:How many synthetic images are rendered for training the depth estimation model?\nAnswer:", + " 500,000" + ], + [ + "Question:How many synthetic images are rendered for training the depth estimation model?\nAnswer:", + " 1,000,000" + ] + ], + "resps": [ + [ + [ + -7.2633056640625, + false + ] + ], + [ + [ + -9.151877403259277, + false + ] + ], + [ + [ + -8.701854705810547, + false + ] + ], + [ + [ + -8.134246826171875, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.2633056640625, + false + ], + [ + -9.151877403259277, + false + ], + [ + -8.701854705810547, + false + ], + [ + -8.134246826171875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a530c8a96b5dcf04ebb711438555ef1844421b25d1bd65e8a7d1179ec529fb08", + "prompt_hash": "de28faf4590b46ded02d75a918fc983f41e86d217d29fcb34fcc6aac506ff354", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 79, + "doc": { + "question": "What is the number of classes in the new vocabulary used for body-part segmentation?", + "choices": [ + "20", + "25", + "28", + "30" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the number of classes in the new vocabulary used for body-part segmentation?\nAnswer:", + " 20" + ], + [ + "Question:What is the number of classes in the new vocabulary used for body-part segmentation?\nAnswer:", + " 25" + ], + [ + "Question:What is the number of classes in the new vocabulary used for body-part segmentation?\nAnswer:", + " 28" + ], + [ + "Question:What is the number of classes in the new vocabulary used for body-part segmentation?\nAnswer:", + " 30" + ] + ], + "resps": [ + [ + [ + -4.645571708679199, + false + ] + ], + [ + [ + -5.531121253967285, + false + ] + ], + [ + [ + -6.0805463790893555, + false + ] + ], + [ + [ + -4.977658271789551, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.645571708679199, + false + ], + [ + -5.531121253967285, + false + ], + [ + -6.0805463790893555, + false + ], + [ + -4.977658271789551, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2155b83e33c84634108fee0adac58d9233fc4ae33e1af8a63ee713084ec1c697", + "prompt_hash": "daf55490b1941e597310bfb2fbc17a7bbef7e53bbeb445d5de70be9b1a9a66dd", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 80, + "doc": { + "question": "What is the total number of parameters in the self-attention architecture model using Mistral-7B and SigLIP-SO400M?", + "choices": [ + "A. 10B", + "B. 8.3B", + "C. 7.4B", + "D. 12.1B" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the total number of parameters in the self-attention architecture model using Mistral-7B and SigLIP-SO400M?\nAnswer:", + " A. 10B" + ], + [ + "Question:What is the total number of parameters in the self-attention architecture model using Mistral-7B and SigLIP-SO400M?\nAnswer:", + " B. 8.3B" + ], + [ + "Question:What is the total number of parameters in the self-attention architecture model using Mistral-7B and SigLIP-SO400M?\nAnswer:", + " C. 7.4B" + ], + [ + "Question:What is the total number of parameters in the self-attention architecture model using Mistral-7B and SigLIP-SO400M?\nAnswer:", + " D. 12.1B" + ] + ], + "resps": [ + [ + [ + -19.912111282348633, + false + ] + ], + [ + [ + -26.02424430847168, + false + ] + ], + [ + [ + -25.916641235351562, + false + ] + ], + [ + [ + -29.070547103881836, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.912111282348633, + false + ], + [ + -26.02424430847168, + false + ], + [ + -25.916641235351562, + false + ], + [ + -29.070547103881836, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3186417562470569683dcde8ba032394572cd147a9accfdf3a8415b0dddc5574", + "prompt_hash": "f1275803e5fe3ae87a49bb0ba73b02f83e368b95c61df7c05fda292ac06ff3f4", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 81, + "doc": { + "question": "Which model feeds image patches directly into the language model using a simple linear projection?", + "choices": [ + "A. Idefics3", + "B. LLaVA", + "C. Fuyu", + "D. BLIP2" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model feeds image patches directly into the language model using a simple linear projection?\nAnswer:", + " A. Idefics3" + ], + [ + "Question:Which model feeds image patches directly into the language model using a simple linear projection?\nAnswer:", + " B. LLaVA" + ], + [ + "Question:Which model feeds image patches directly into the language model using a simple linear projection?\nAnswer:", + " C. Fuyu" + ], + [ + "Question:Which model feeds image patches directly into the language model using a simple linear projection?\nAnswer:", + " D. BLIP2" + ] + ], + "resps": [ + [ + [ + -29.24394416809082, + false + ] + ], + [ + [ + -21.120861053466797, + false + ] + ], + [ + [ + -27.182903289794922, + false + ] + ], + [ + [ + -22.341655731201172, + false + ] + ] + ], + "filtered_resps": [ + [ + -29.24394416809082, + false + ], + [ + -21.120861053466797, + false + ], + [ + -27.182903289794922, + false + ], + [ + -22.341655731201172, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4e044d2aaa20375be6cd8c444dca3b26a8dab26ce04dc4bc9ca080f73e9b67d4", + "prompt_hash": "724be7e74e772a2e41442b4f177b04d04d82390c745bc0803b5e18d09232a6d7", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 82, + "doc": { + "question": "What token is inserted after each row of tiles in Idefics3 to preserve spatial layout?", + "choices": [ + "A. ", + "B. ", + "C. ", + "D. \\n" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What token is inserted after each row of tiles in Idefics3 to preserve spatial layout?\nAnswer:", + " A. " + ], + [ + "Question:What token is inserted after each row of tiles in Idefics3 to preserve spatial layout?\nAnswer:", + " B. " + ], + [ + "Question:What token is inserted after each row of tiles in Idefics3 to preserve spatial layout?\nAnswer:", + " C. " + ], + [ + "Question:What token is inserted after each row of tiles in Idefics3 to preserve spatial layout?\nAnswer:", + " D. \\n" + ] + ], + "resps": [ + [ + [ + -19.776901245117188, + false + ] + ], + [ + [ + -26.64897918701172, + false + ] + ], + [ + [ + -30.480342864990234, + false + ] + ], + [ + [ + -20.79037094116211, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.776901245117188, + false + ], + [ + -26.64897918701172, + false + ], + [ + -30.480342864990234, + false + ], + [ + -20.79037094116211, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7252e08a4477e278086b5719b659420ecbbe5c6b03d49ff0430c5d1b75f442c7", + "prompt_hash": "6f9e27249956f97a187ee548ff99115a8575ea2f5fa94228895a4fc701f1938d", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 83, + "doc": { + "question": "How many QA pairs does the Docmatix dataset contain?", + "choices": [ + "A. 2.4 million", + "B. 1.3 million", + "C. 9.5 million", + "D. 40 thousand" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many QA pairs does the Docmatix dataset contain?\nAnswer:", + " A. 2.4 million" + ], + [ + "Question:How many QA pairs does the Docmatix dataset contain?\nAnswer:", + " B. 1.3 million" + ], + [ + "Question:How many QA pairs does the Docmatix dataset contain?\nAnswer:", + " C. 9.5 million" + ], + [ + "Question:How many QA pairs does the Docmatix dataset contain?\nAnswer:", + " D. 40 thousand" + ] + ], + "resps": [ + [ + [ + -20.991960525512695, + false + ] + ], + [ + [ + -25.32790756225586, + false + ] + ], + [ + [ + -27.947509765625, + false + ] + ], + [ + [ + -26.229846954345703, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.991960525512695, + false + ], + [ + -25.32790756225586, + false + ], + [ + -27.947509765625, + false + ], + [ + -26.229846954345703, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0ac12468abd789b0f9037786e07f66561aa6aff5b692663db49d0947cbc58283", + "prompt_hash": "be81629cf550f614486bee6f4e41327c8b3d3a1ec1029e2750ce0c8b6f2214ba", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 84, + "doc": { + "question": "Which benchmark score is used to evaluate performance on the DocVQA task?", + "choices": [ + "A. Accuracy", + "B. ANLS", + "C. BLEU", + "D. VQA Score" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which benchmark score is used to evaluate performance on the DocVQA task?\nAnswer:", + " A. Accuracy" + ], + [ + "Question:Which benchmark score is used to evaluate performance on the DocVQA task?\nAnswer:", + " B. ANLS" + ], + [ + "Question:Which benchmark score is used to evaluate performance on the DocVQA task?\nAnswer:", + " C. BLEU" + ], + [ + "Question:Which benchmark score is used to evaluate performance on the DocVQA task?\nAnswer:", + " D. VQA Score" + ] + ], + "resps": [ + [ + [ + -11.713038444519043, + false + ] + ], + [ + [ + -27.411283493041992, + false + ] + ], + [ + [ + -15.24494457244873, + false + ] + ], + [ + [ + -23.070180892944336, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.713038444519043, + false + ], + [ + -27.411283493041992, + false + ], + [ + -15.24494457244873, + false + ], + [ + -23.070180892944336, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ed66f74f3c578709f069bfccd45cde901335880474927d904012d9eb1b0f641f", + "prompt_hash": "1dcf27ccd8b63f609b9e414a8d4155fa598dec6391547a212abd20758440d679", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 85, + "doc": { + "question": "What is the approximate logical error suppression factor (Λ) observed using the neural network decoder for the surface code?", + "choices": [ + "1.71", + "2.14", + "2.04", + "1.43" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the approximate logical error suppression factor (Λ) observed using the neural network decoder for the surface code?\nAnswer:", + " 1.71" + ], + [ + "Question:What is the approximate logical error suppression factor (Λ) observed using the neural network decoder for the surface code?\nAnswer:", + " 2.14" + ], + [ + "Question:What is the approximate logical error suppression factor (Λ) observed using the neural network decoder for the surface code?\nAnswer:", + " 2.04" + ], + [ + "Question:What is the approximate logical error suppression factor (Λ) observed using the neural network decoder for the surface code?\nAnswer:", + " 1.43" + ] + ], + "resps": [ + [ + [ + -9.656071662902832, + false + ] + ], + [ + [ + -10.014352798461914, + false + ] + ], + [ + [ + -10.03097915649414, + false + ] + ], + [ + [ + -9.16510009765625, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.656071662902832, + false + ], + [ + -10.014352798461914, + false + ], + [ + -10.03097915649414, + false + ], + [ + -9.16510009765625, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f83dcb8fa32e705ac558c88af25d83b05b9235f1ba3ae7c15ad1b6a626edc67f", + "prompt_hash": "2855d1e6072b4821bc28e368c7ca76e92054e3f7dae601b26cc65c12ae23bb94", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 86, + "doc": { + "question": "How many physical qubits are used in the distance-7 surface code implementation on the 105-qubit processor?", + "choices": [ + "49", + "48", + "101", + "105" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many physical qubits are used in the distance-7 surface code implementation on the 105-qubit processor?\nAnswer:", + " 49" + ], + [ + "Question:How many physical qubits are used in the distance-7 surface code implementation on the 105-qubit processor?\nAnswer:", + " 48" + ], + [ + "Question:How many physical qubits are used in the distance-7 surface code implementation on the 105-qubit processor?\nAnswer:", + " 101" + ], + [ + "Question:How many physical qubits are used in the distance-7 surface code implementation on the 105-qubit processor?\nAnswer:", + " 105" + ] + ], + "resps": [ + [ + [ + -6.783773422241211, + false + ] + ], + [ + [ + -6.077127456665039, + false + ] + ], + [ + [ + -7.488647937774658, + false + ] + ], + [ + [ + -5.374570369720459, + true + ] + ] + ], + "filtered_resps": [ + [ + -6.783773422241211, + false + ], + [ + -6.077127456665039, + false + ], + [ + -7.488647937774658, + false + ], + [ + -5.374570369720459, + true + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3ac616f94d005707134ede2e1c37cad01760706de6162fd22fdddef31fa7a996", + "prompt_hash": "e8041853efe4132459bdbeafc15bd95282ad9772ff068dfb2c03880e3ae2f4a0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 87, + "doc": { + "question": "What is the reported cycle time for the superconducting processor requiring real-time decoding?", + "choices": [ + "1.1 ms", + "24 µs", + "1.1 µs", + "10 µs" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the reported cycle time for the superconducting processor requiring real-time decoding?\nAnswer:", + " 1.1 ms" + ], + [ + "Question:What is the reported cycle time for the superconducting processor requiring real-time decoding?\nAnswer:", + " 24 µs" + ], + [ + "Question:What is the reported cycle time for the superconducting processor requiring real-time decoding?\nAnswer:", + " 1.1 µs" + ], + [ + "Question:What is the reported cycle time for the superconducting processor requiring real-time decoding?\nAnswer:", + " 10 µs" + ] + ], + "resps": [ + [ + [ + -10.168475151062012, + false + ] + ], + [ + [ + -14.862546920776367, + false + ] + ], + [ + [ + -14.548297882080078, + false + ] + ], + [ + [ + -11.576759338378906, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.168475151062012, + false + ], + [ + -14.862546920776367, + false + ], + [ + -14.548297882080078, + false + ], + [ + -11.576759338378906, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e5a76c2d4a27925f448844ee634c1336c9eb420f57be71bf6bde21e0c69f8477", + "prompt_hash": "690ad43c485740a8a906fe16420a2ef6163d4e4876bf79d429256dcca0f3dfb1", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 88, + "doc": { + "question": "Which component contributes the most to the error budget in the 72-qubit processor according to simulations?", + "choices": [ + "Measurement and reset error", + "Leakage during CZs and due to heating", + "CZ gate errors including stray interactions", + "Single-qubit gate error" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which component contributes the most to the error budget in the 72-qubit processor according to simulations?\nAnswer:", + " Measurement and reset error" + ], + [ + "Question:Which component contributes the most to the error budget in the 72-qubit processor according to simulations?\nAnswer:", + " Leakage during CZs and due to heating" + ], + [ + "Question:Which component contributes the most to the error budget in the 72-qubit processor according to simulations?\nAnswer:", + " CZ gate errors including stray interactions" + ], + [ + "Question:Which component contributes the most to the error budget in the 72-qubit processor according to simulations?\nAnswer:", + " Single-qubit gate error" + ] + ], + "resps": [ + [ + [ + -20.62749481201172, + false + ] + ], + [ + [ + -50.088890075683594, + false + ] + ], + [ + [ + -41.10272216796875, + false + ] + ], + [ + [ + -10.830703735351562, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.62749481201172, + false + ], + [ + -50.088890075683594, + false + ], + [ + -41.10272216796875, + false + ], + [ + -10.830703735351562, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e33cdc4c62885ee64dd285f973173693dcae0da6c8297e11d45f259b5f7c89ad", + "prompt_hash": "26648c0d5fdbfedcc1b4c10426c3760df47ac02e344db4805edb96fd47a50733", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 89, + "doc": { + "question": "What method is used to mitigate data qubit leakage during surface code operation?", + "choices": [ + "Passive cooling", + "Dynamical decoupling", + "Data qubit leakage removal (DQLR)", + "Multi-level reset" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What method is used to mitigate data qubit leakage during surface code operation?\nAnswer:", + " Passive cooling" + ], + [ + "Question:What method is used to mitigate data qubit leakage during surface code operation?\nAnswer:", + " Dynamical decoupling" + ], + [ + "Question:What method is used to mitigate data qubit leakage during surface code operation?\nAnswer:", + " Data qubit leakage removal (DQLR)" + ], + [ + "Question:What method is used to mitigate data qubit leakage during surface code operation?\nAnswer:", + " Multi-level reset" + ] + ], + "resps": [ + [ + [ + -14.593650817871094, + false + ] + ], + [ + [ + -9.150117874145508, + false + ] + ], + [ + [ + -26.144542694091797, + false + ] + ], + [ + [ + -20.310792922973633, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.593650817871094, + false + ], + [ + -9.150117874145508, + false + ], + [ + -26.144542694091797, + false + ], + [ + -20.310792922973633, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2c5843b32d9ee6229b54175779e7c1d53fa45c7089f073735fe900bc778a6641", + "prompt_hash": "ad8aeed8ec0324509fadc57ee2c062485ed9a7411672ea9a453fb225ef02e5c3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 90, + "doc": { + "question": "Which model is based on LLaMA3-8B and adopts visual expert modules?", + "choices": [ + "CogAgent", + "GLM-4V-9B", + "CogVLM2", + "VisualGLM" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model is based on LLaMA3-8B and adopts visual expert modules?\nAnswer:", + " CogAgent" + ], + [ + "Question:Which model is based on LLaMA3-8B and adopts visual expert modules?\nAnswer:", + " GLM-4V-9B" + ], + [ + "Question:Which model is based on LLaMA3-8B and adopts visual expert modules?\nAnswer:", + " CogVLM2" + ], + [ + "Question:Which model is based on LLaMA3-8B and adopts visual expert modules?\nAnswer:", + " VisualGLM" + ] + ], + "resps": [ + [ + [ + -15.072737693786621, + false + ] + ], + [ + [ + -24.121051788330078, + false + ] + ], + [ + [ + -16.509078979492188, + false + ] + ], + [ + [ + -12.67910385131836, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.072737693786621, + false + ], + [ + -24.121051788330078, + false + ], + [ + -16.509078979492188, + false + ], + [ + -12.67910385131836, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "17140906c051bae2097e001a7d8abdf39877f5558dc393130c5a3adb52fb259c", + "prompt_hash": "e8cecc717ed96ab7822f3e7bea68d62464c74e7c1f0a1f19854dfaadf291243f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 91, + "doc": { + "question": "Which dataset contains 7 million English and 5 million Chinese entries created from browser-captured screenshots?", + "choices": [ + "CLAY-1B", + "Digital World Grounding Dataset", + "Synthetic OCR Dataset", + "LAION-40M-grounding" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset contains 7 million English and 5 million Chinese entries created from browser-captured screenshots?\nAnswer:", + " CLAY-1B" + ], + [ + "Question:Which dataset contains 7 million English and 5 million Chinese entries created from browser-captured screenshots?\nAnswer:", + " Digital World Grounding Dataset" + ], + [ + "Question:Which dataset contains 7 million English and 5 million Chinese entries created from browser-captured screenshots?\nAnswer:", + " Synthetic OCR Dataset" + ], + [ + "Question:Which dataset contains 7 million English and 5 million Chinese entries created from browser-captured screenshots?\nAnswer:", + " LAION-40M-grounding" + ] + ], + "resps": [ + [ + [ + -30.70033073425293, + false + ] + ], + [ + [ + -35.19849395751953, + false + ] + ], + [ + [ + -22.89411163330078, + false + ] + ], + [ + [ + -34.586299896240234, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.70033073425293, + false + ], + [ + -35.19849395751953, + false + ], + [ + -22.89411163330078, + false + ], + [ + -34.586299896240234, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e18c9e2c9f7e6b3038aef222e26ffcdce6251b96c24cc379159162b29e79729b", + "prompt_hash": "9753bf05d93a5e528813540a56d21ba96eb749574ff9d408a6650516cb96d993", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 92, + "doc": { + "question": "Which architecture component reduces image sequence length to one-quarter its original size?", + "choices": [ + "Q-former", + "Linear layer", + "2×2 convolutional layer", + "ViT encoder" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which architecture component reduces image sequence length to one-quarter its original size?\nAnswer:", + " Q-former" + ], + [ + "Question:Which architecture component reduces image sequence length to one-quarter its original size?\nAnswer:", + " Linear layer" + ], + [ + "Question:Which architecture component reduces image sequence length to one-quarter its original size?\nAnswer:", + " 2×2 convolutional layer" + ], + [ + "Question:Which architecture component reduces image sequence length to one-quarter its original size?\nAnswer:", + " ViT encoder" + ] + ], + "resps": [ + [ + [ + -10.258474349975586, + false + ] + ], + [ + [ + -10.730835914611816, + false + ] + ], + [ + [ + -18.697784423828125, + false + ] + ], + [ + [ + -14.105308532714844, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.258474349975586, + false + ], + [ + -10.730835914611816, + false + ], + [ + -18.697784423828125, + false + ], + [ + -14.105308532714844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "af39862375d46684cfdb2dc96930bf4914e05937e23f1ab07d1c6a9490ae4a92", + "prompt_hash": "e3c3ebaf6ae5c48799dd038be07a5c66a8af5874ad2c98fa58f76dedac17beca", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 93, + "doc": { + "question": "What was the learning rate used in the second stage of image supervised fine-tuning for CogVLM2?", + "choices": [ + "1e-4", + "1e-6", + "1e-5", + "Not specified" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What was the learning rate used in the second stage of image supervised fine-tuning for CogVLM2?\nAnswer:", + " 1e-4" + ], + [ + "Question:What was the learning rate used in the second stage of image supervised fine-tuning for CogVLM2?\nAnswer:", + " 1e-6" + ], + [ + "Question:What was the learning rate used in the second stage of image supervised fine-tuning for CogVLM2?\nAnswer:", + " 1e-5" + ], + [ + "Question:What was the learning rate used in the second stage of image supervised fine-tuning for CogVLM2?\nAnswer:", + " Not specified" + ] + ], + "resps": [ + [ + [ + -5.152227401733398, + false + ] + ], + [ + [ + -6.627537727355957, + false + ] + ], + [ + [ + -5.376708984375, + false + ] + ], + [ + [ + -10.316917419433594, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.152227401733398, + false + ], + [ + -6.627537727355957, + false + ], + [ + -5.376708984375, + false + ], + [ + -10.316917419433594, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4399dd4afd8a3359d187f5d62e0738fc49b754310b74e1483e2faca1f0496c18", + "prompt_hash": "b3d42e6622a06ce8d5e9fdad5e5637a516c8b11334e8eae2fc1c90ba5de91ff3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 94, + "doc": { + "question": "Which post-training dataset includes 150 million Chinese entries focused on OCR scenarios?", + "choices": [ + "CLAY-1B", + "Digital World Grounding Dataset", + "Synthetic OCR Dataset", + "LAION-2B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which post-training dataset includes 150 million Chinese entries focused on OCR scenarios?\nAnswer:", + " CLAY-1B" + ], + [ + "Question:Which post-training dataset includes 150 million Chinese entries focused on OCR scenarios?\nAnswer:", + " Digital World Grounding Dataset" + ], + [ + "Question:Which post-training dataset includes 150 million Chinese entries focused on OCR scenarios?\nAnswer:", + " Synthetic OCR Dataset" + ], + [ + "Question:Which post-training dataset includes 150 million Chinese entries focused on OCR scenarios?\nAnswer:", + " LAION-2B" + ] + ], + "resps": [ + [ + [ + -28.669137954711914, + false + ] + ], + [ + [ + -34.62318420410156, + false + ] + ], + [ + [ + -17.01984977722168, + false + ] + ], + [ + [ + -14.952802658081055, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.669137954711914, + false + ], + [ + -34.62318420410156, + false + ], + [ + -17.01984977722168, + false + ], + [ + -14.952802658081055, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "03e6f501becf101b8451db2075dd9cbefc9c5273162555dd24142c94a997bfc3", + "prompt_hash": "d0789a0d2362a5a4facaa5ded8fc2cea59864ee3bad34cfcffbb419b4f55151c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 95, + "doc": { + "question": "Which dataset is used to train the model for novel view synthesis?", + "choices": [ + "KITTI", + "ScanNet++", + "Cityscapes", + "MegaDepth" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset is used to train the model for novel view synthesis?\nAnswer:", + " KITTI" + ], + [ + "Question:Which dataset is used to train the model for novel view synthesis?\nAnswer:", + " ScanNet++" + ], + [ + "Question:Which dataset is used to train the model for novel view synthesis?\nAnswer:", + " Cityscapes" + ], + [ + "Question:Which dataset is used to train the model for novel view synthesis?\nAnswer:", + " MegaDepth" + ] + ], + "resps": [ + [ + [ + -6.542630672454834, + false + ] + ], + [ + [ + -11.459537506103516, + false + ] + ], + [ + [ + -7.253916263580322, + false + ] + ], + [ + [ + -11.059212684631348, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.542630672454834, + false + ], + [ + -11.459537506103516, + false + ], + [ + -7.253916263580322, + false + ], + [ + -11.059212684631348, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "329f9d464b42c81c455e7a95a0a531e934b2d871c240ea8a3cd0f6296ce9c045", + "prompt_hash": "435e87cff1bc9dc01a42c14bc019d2595c9b9f66304252673b9ee8888b9f4904", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 96, + "doc": { + "question": "What is the primary input to the proposed model for 3D reconstruction?", + "choices": [ + "Calibrated image pair", + "Depth map and RGB image", + "Two uncalibrated images", + "Monocular video stream" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the primary input to the proposed model for 3D reconstruction?\nAnswer:", + " Calibrated image pair" + ], + [ + "Question:What is the primary input to the proposed model for 3D reconstruction?\nAnswer:", + " Depth map and RGB image" + ], + [ + "Question:What is the primary input to the proposed model for 3D reconstruction?\nAnswer:", + " Two uncalibrated images" + ], + [ + "Question:What is the primary input to the proposed model for 3D reconstruction?\nAnswer:", + " Monocular video stream" + ] + ], + "resps": [ + [ + [ + -21.52069664001465, + false + ] + ], + [ + [ + -17.651397705078125, + false + ] + ], + [ + [ + -22.90048599243164, + false + ] + ], + [ + [ + -18.692649841308594, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.52069664001465, + false + ], + [ + -17.651397705078125, + false + ], + [ + -22.90048599243164, + false + ], + [ + -18.692649841308594, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "054f7f6b7928cc3c3df3a3a2677f1bb85093ecf64f9fc86fbf18428192a6c62f", + "prompt_hash": "e984d345e214624538cd6076bc0cdbe27590bb6da385151f3d3f141f61859d57", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 97, + "doc": { + "question": "What does the model use to parameterize the covariance matrix of each 3D Gaussian?", + "choices": [ + "Matrix multiplication and bias", + "Euler angles and scale", + "Rotation quaternion and scale", + "Pose vectors and tensors" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does the model use to parameterize the covariance matrix of each 3D Gaussian?\nAnswer:", + " Matrix multiplication and bias" + ], + [ + "Question:What does the model use to parameterize the covariance matrix of each 3D Gaussian?\nAnswer:", + " Euler angles and scale" + ], + [ + "Question:What does the model use to parameterize the covariance matrix of each 3D Gaussian?\nAnswer:", + " Rotation quaternion and scale" + ], + [ + "Question:What does the model use to parameterize the covariance matrix of each 3D Gaussian?\nAnswer:", + " Pose vectors and tensors" + ] + ], + "resps": [ + [ + [ + -19.31600570678711, + false + ] + ], + [ + [ + -16.103641510009766, + false + ] + ], + [ + [ + -19.320552825927734, + false + ] + ], + [ + [ + -25.82880401611328, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.31600570678711, + false + ], + [ + -16.103641510009766, + false + ], + [ + -19.320552825927734, + false + ], + [ + -25.82880401611328, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6a81dcfab0959a77b46d84e940ed374e7c1ae77922ce9c436ac83baa1cb49086", + "prompt_hash": "d562f05e3bd10c0e334befc0db9db03a35259e4277c131ddb0da2c44c6f74cc5", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 98, + "doc": { + "question": "What rendering losses are applied to train the Gaussian prediction head?", + "choices": [ + "L1 loss and SSIM", + "MSE and LPIPS", + "Cross-entropy and Focal loss", + "KL divergence and Huber loss" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What rendering losses are applied to train the Gaussian prediction head?\nAnswer:", + " L1 loss and SSIM" + ], + [ + "Question:What rendering losses are applied to train the Gaussian prediction head?\nAnswer:", + " MSE and LPIPS" + ], + [ + "Question:What rendering losses are applied to train the Gaussian prediction head?\nAnswer:", + " Cross-entropy and Focal loss" + ], + [ + "Question:What rendering losses are applied to train the Gaussian prediction head?\nAnswer:", + " KL divergence and Huber loss" + ] + ], + "resps": [ + [ + [ + -15.82857894897461, + false + ] + ], + [ + [ + -13.206483840942383, + false + ] + ], + [ + [ + -20.15416145324707, + false + ] + ], + [ + [ + -18.332340240478516, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.82857894897461, + false + ], + [ + -13.206483840942383, + false + ], + [ + -20.15416145324707, + false + ], + [ + -18.332340240478516, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7d12159581cafe65f033926ab36dfe7446ac692929a70477c621c3d96b5b6f6c", + "prompt_hash": "6f111e0ff63c4d6c53ca139655096ade67eda570b7417bade09f30f7236276fa", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 99, + "doc": { + "question": "What purpose does the loss masking strategy serve during training?", + "choices": [ + "To reduce GPU memory usage", + "To encourage smooth transitions between views", + "To avoid supervising pixels not visible in context images", + "To enhance the visibility of depth maps" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What purpose does the loss masking strategy serve during training?\nAnswer:", + " To reduce GPU memory usage" + ], + [ + "Question:What purpose does the loss masking strategy serve during training?\nAnswer:", + " To encourage smooth transitions between views" + ], + [ + "Question:What purpose does the loss masking strategy serve during training?\nAnswer:", + " To avoid supervising pixels not visible in context images" + ], + [ + "Question:What purpose does the loss masking strategy serve during training?\nAnswer:", + " To enhance the visibility of depth maps" + ] + ], + "resps": [ + [ + [ + -14.565971374511719, + false + ] + ], + [ + [ + -22.251873016357422, + false + ] + ], + [ + [ + -47.93913269042969, + false + ] + ], + [ + [ + -28.24319076538086, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.565971374511719, + false + ], + [ + -22.251873016357422, + false + ], + [ + -47.93913269042969, + false + ], + [ + -28.24319076538086, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0734bd5e11f0e733942ec8180ee434e9a3c295a28e77baf791843224e58b1d37", + "prompt_hash": "3ca9c635bfa6bb022e05cc5a47c2a7a230bd6ee0c46b649a6219726d4f4c5862", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + } + ], + "arxiv_mc_2409": [ + { + "doc_id": 0, + "doc": { + "question": "What is the visual encoder used in the architecture of \\u005cmodelname?", + "choices": [ + "Video Swin Transformer", + "TubeViT", + "CLIP-ViT", + "SigLIP" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the visual encoder used in the architecture of \\u005cmodelname?\nAnswer:", + " Video Swin Transformer" + ], + [ + "Question:What is the visual encoder used in the architecture of \\u005cmodelname?\nAnswer:", + " TubeViT" + ], + [ + "Question:What is the visual encoder used in the architecture of \\u005cmodelname?\nAnswer:", + " CLIP-ViT" + ], + [ + "Question:What is the visual encoder used in the architecture of \\u005cmodelname?\nAnswer:", + " SigLIP" + ] + ], + "resps": [ + [ + [ + -15.495767593383789, + false + ] + ], + [ + [ + -17.449329376220703, + false + ] + ], + [ + [ + -12.375112533569336, + false + ] + ], + [ + [ + -13.049660682678223, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.495767593383789, + false + ], + [ + -17.449329376220703, + false + ], + [ + -12.375112533569336, + false + ], + [ + -13.049660682678223, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8f43676dfa8e99d64af0868235cfc9038da15b626ad93e21c4b65905122e0d60", + "prompt_hash": "1b419424cdc31dcd1715f04450f85f4d9581dcae81486bf8952af146acbdc655", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 1, + "doc": { + "question": "How many video clips are included in the \\u005cdatasetname benchmark?", + "choices": [ + "500", + "1,000", + "2,000", + "10,000" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many video clips are included in the \\u005cdatasetname benchmark?\nAnswer:", + " 500" + ], + [ + "Question:How many video clips are included in the \\u005cdatasetname benchmark?\nAnswer:", + " 1,000" + ], + [ + "Question:How many video clips are included in the \\u005cdatasetname benchmark?\nAnswer:", + " 2,000" + ], + [ + "Question:How many video clips are included in the \\u005cdatasetname benchmark?\nAnswer:", + " 10,000" + ] + ], + "resps": [ + [ + [ + -6.426371097564697, + false + ] + ], + [ + [ + -8.035820960998535, + false + ] + ], + [ + [ + -8.831319808959961, + false + ] + ], + [ + [ + -8.405014038085938, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.426371097564697, + false + ], + [ + -8.035820960998535, + false + ], + [ + -8.831319808959961, + false + ], + [ + -8.405014038085938, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "aa7b82cc13c10a1cfd45dcefe86437a90471b75fdc8e8173718d949889adecbb", + "prompt_hash": "68377350cdd14f3bdf1dd8b868e22f0683d7519a7b67d812761480c27d8c051d", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 2, + "doc": { + "question": "What two components does \\u005cevalname use to evaluate descriptions?", + "choices": [ + "Precision and recall metrics", + "BLEU and METEOR", + "Event extraction and entailment models", + "Human ratings and CIDEr scores" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What two components does \\u005cevalname use to evaluate descriptions?\nAnswer:", + " Precision and recall metrics" + ], + [ + "Question:What two components does \\u005cevalname use to evaluate descriptions?\nAnswer:", + " BLEU and METEOR" + ], + [ + "Question:What two components does \\u005cevalname use to evaluate descriptions?\nAnswer:", + " Event extraction and entailment models" + ], + [ + "Question:What two components does \\u005cevalname use to evaluate descriptions?\nAnswer:", + " Human ratings and CIDEr scores" + ] + ], + "resps": [ + [ + [ + -19.11516571044922, + false + ] + ], + [ + [ + -12.382637977600098, + false + ] + ], + [ + [ + -30.587970733642578, + false + ] + ], + [ + [ + -26.8033390045166, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.11516571044922, + false + ], + [ + -12.382637977600098, + false + ], + [ + -30.587970733642578, + false + ], + [ + -26.8033390045166, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6225b0f32d9f853fce5a77063c84800ff8eedcbe76a686813ca0e4146ee019aa", + "prompt_hash": "f3e0850221596a87411a4319e2cf07cddc5cb851555a6894f0861579adf68da4", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 3, + "doc": { + "question": "What is the total number of video-text pairs used during \\u005cmodelname's pre-training?", + "choices": [ + "2.8 million", + "6.5 million", + "13.6 million", + "38 million" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the total number of video-text pairs used during \\u005cmodelname's pre-training?\nAnswer:", + " 2.8 million" + ], + [ + "Question:What is the total number of video-text pairs used during \\u005cmodelname's pre-training?\nAnswer:", + " 6.5 million" + ], + [ + "Question:What is the total number of video-text pairs used during \\u005cmodelname's pre-training?\nAnswer:", + " 13.6 million" + ], + [ + "Question:What is the total number of video-text pairs used during \\u005cmodelname's pre-training?\nAnswer:", + " 38 million" + ] + ], + "resps": [ + [ + [ + -10.427526473999023, + false + ] + ], + [ + [ + -10.509381294250488, + false + ] + ], + [ + [ + -11.587591171264648, + false + ] + ], + [ + [ + -10.692402839660645, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.427526473999023, + false + ], + [ + -10.509381294250488, + false + ], + [ + -11.587591171264648, + false + ], + [ + -10.692402839660645, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ec9ad2089dbc4bdfa6bd2478e3dfe84d4fb2e0aa51458b552776d9fcd0f039ce", + "prompt_hash": "decd19be57941d960406e3809c930cc8043d44b9bfd94d681e7ec4a5b369859f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 4, + "doc": { + "question": "Which proprietary model is slightly preferred over \\u005cmodelname-34B in human side-by-side evaluation?", + "choices": [ + "GPT-4V", + "Gemini 1.5 Pro", + "GPT-4o", + "VideoChat2" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which proprietary model is slightly preferred over \\u005cmodelname-34B in human side-by-side evaluation?\nAnswer:", + " GPT-4V" + ], + [ + "Question:Which proprietary model is slightly preferred over \\u005cmodelname-34B in human side-by-side evaluation?\nAnswer:", + " Gemini 1.5 Pro" + ], + [ + "Question:Which proprietary model is slightly preferred over \\u005cmodelname-34B in human side-by-side evaluation?\nAnswer:", + " GPT-4o" + ], + [ + "Question:Which proprietary model is slightly preferred over \\u005cmodelname-34B in human side-by-side evaluation?\nAnswer:", + " VideoChat2" + ] + ], + "resps": [ + [ + [ + -12.168880462646484, + false + ] + ], + [ + [ + -11.913797378540039, + false + ] + ], + [ + [ + -9.950221061706543, + false + ] + ], + [ + [ + -19.21074867248535, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.168880462646484, + false + ], + [ + -11.913797378540039, + false + ], + [ + -9.950221061706543, + false + ], + [ + -19.21074867248535, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3aeea6b1458836b9cfd8e1d80e468cc01769beb5d9c845810764ad974c715b99", + "prompt_hash": "1a27e107b5dcb6af186b7105913c3d136087e2b0c73041bb22c25768fb0b0b6e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 5, + "doc": { + "question": "Which theorem forms the basis of Kolmogorov-Arnold Networks (KANs)?", + "choices": [ + "Universal Approximation Theorem", + "Kolmogorov-Arnold Representation Theorem", + "Central Limit Theorem", + "Banach Fixed-Point Theorem" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which theorem forms the basis of Kolmogorov-Arnold Networks (KANs)?\nAnswer:", + " Universal Approximation Theorem" + ], + [ + "Question:Which theorem forms the basis of Kolmogorov-Arnold Networks (KANs)?\nAnswer:", + " Kolmogorov-Arnold Representation Theorem" + ], + [ + "Question:Which theorem forms the basis of Kolmogorov-Arnold Networks (KANs)?\nAnswer:", + " Central Limit Theorem" + ], + [ + "Question:Which theorem forms the basis of Kolmogorov-Arnold Networks (KANs)?\nAnswer:", + " Banach Fixed-Point Theorem" + ] + ], + "resps": [ + [ + [ + -11.135324478149414, + false + ] + ], + [ + [ + -10.827104568481445, + false + ] + ], + [ + [ + -10.046956062316895, + false + ] + ], + [ + [ + -17.358673095703125, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.135324478149414, + false + ], + [ + -10.827104568481445, + false + ], + [ + -10.046956062316895, + false + ], + [ + -17.358673095703125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "564ec16c40c0c8d379bae9ed11a470dd4f4b978724e42191d260862063c2d06b", + "prompt_hash": "7ea86d8f7700ff5117a06df5224c06b7c949b2625a74137d37d9b73b523a5bec", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 6, + "doc": { + "question": "What activation function was originally proposed by Liu et al. for KANs?", + "choices": [ + "Gaussian radial basis functions", + "Swish", + "ReLU", + "B-spline" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What activation function was originally proposed by Liu et al. for KANs?\nAnswer:", + " Gaussian radial basis functions" + ], + [ + "Question:What activation function was originally proposed by Liu et al. for KANs?\nAnswer:", + " Swish" + ], + [ + "Question:What activation function was originally proposed by Liu et al. for KANs?\nAnswer:", + " ReLU" + ], + [ + "Question:What activation function was originally proposed by Liu et al. for KANs?\nAnswer:", + " B-spline" + ] + ], + "resps": [ + [ + [ + -17.572092056274414, + false + ] + ], + [ + [ + -6.097117900848389, + false + ] + ], + [ + [ + -2.6757876873016357, + false + ] + ], + [ + [ + -12.105659484863281, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.572092056274414, + false + ], + [ + -6.097117900848389, + false + ], + [ + -2.6757876873016357, + false + ], + [ + -12.105659484863281, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "dcbfbd43678c7eaf0b55b660693da39aff35639414c5df78ab209de375693c0f", + "prompt_hash": "e9ffeb039d445a5fccb6de2caa808f2544a0fc5812d3375b0151d9902141da92", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 7, + "doc": { + "question": "Which ODE solver was employed by default in the implementation of KAN-ODEs?", + "choices": [ + "Euler method", + "Tsit5 Runge-Kutta method", + "Rodas5 Rosenbrock method", + "RK4 method" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which ODE solver was employed by default in the implementation of KAN-ODEs?\nAnswer:", + " Euler method" + ], + [ + "Question:Which ODE solver was employed by default in the implementation of KAN-ODEs?\nAnswer:", + " Tsit5 Runge-Kutta method" + ], + [ + "Question:Which ODE solver was employed by default in the implementation of KAN-ODEs?\nAnswer:", + " Rodas5 Rosenbrock method" + ], + [ + "Question:Which ODE solver was employed by default in the implementation of KAN-ODEs?\nAnswer:", + " RK4 method" + ] + ], + "resps": [ + [ + [ + -6.391721725463867, + false + ] + ], + [ + [ + -29.77779769897461, + false + ] + ], + [ + [ + -39.73161315917969, + false + ] + ], + [ + [ + -11.668317794799805, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.391721725463867, + false + ], + [ + -29.77779769897461, + false + ], + [ + -39.73161315917969, + false + ], + [ + -11.668317794799805, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "747de9bc42d4924efdd61e9028dd07faec0e61e0db10a3b467fcf3c01d3a6b18", + "prompt_hash": "ae852e95920e9e1a207f05a15800b95046c3fcf68978f4b13614a111b0dddc28", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 8, + "doc": { + "question": "In the Lotka-Volterra predator-prey example, what was the initial condition used?", + "choices": [ + "[0, 0]", + "[2, 2]", + "[1, 1]", + "[3, 3]" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In the Lotka-Volterra predator-prey example, what was the initial condition used?\nAnswer:", + " [0, 0]" + ], + [ + "Question:In the Lotka-Volterra predator-prey example, what was the initial condition used?\nAnswer:", + " [2, 2]" + ], + [ + "Question:In the Lotka-Volterra predator-prey example, what was the initial condition used?\nAnswer:", + " [1, 1]" + ], + [ + "Question:In the Lotka-Volterra predator-prey example, what was the initial condition used?\nAnswer:", + " [3, 3]" + ] + ], + "resps": [ + [ + [ + -16.677711486816406, + false + ] + ], + [ + [ + -16.641386032104492, + false + ] + ], + [ + [ + -15.760937690734863, + false + ] + ], + [ + [ + -18.1982479095459, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.677711486816406, + false + ], + [ + -16.641386032104492, + false + ], + [ + -15.760937690734863, + false + ], + [ + -18.1982479095459, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "93021bd825ff8280fcea636dfde8e164be3d2313e95d71cf8c3a9a6b8914e9d6", + "prompt_hash": "2329c2216a6718d4bebf58ca3594b68e7b3dde2251b5f6b547995d755e2a37d0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 9, + "doc": { + "question": "Which equation was used as an example of modeling hidden physics in PDEs?", + "choices": [ + "Schrödinger equation", + "Burgers' equation", + "Fisher-KPP equation", + "Allen-Cahn equation" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which equation was used as an example of modeling hidden physics in PDEs?\nAnswer:", + " Schrödinger equation" + ], + [ + "Question:Which equation was used as an example of modeling hidden physics in PDEs?\nAnswer:", + " Burgers' equation" + ], + [ + "Question:Which equation was used as an example of modeling hidden physics in PDEs?\nAnswer:", + " Fisher-KPP equation" + ], + [ + "Question:Which equation was used as an example of modeling hidden physics in PDEs?\nAnswer:", + " Allen-Cahn equation" + ] + ], + "resps": [ + [ + [ + -8.50059986114502, + false + ] + ], + [ + [ + -10.451626777648926, + false + ] + ], + [ + [ + -15.836602210998535, + false + ] + ], + [ + [ + -11.835906028747559, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.50059986114502, + false + ], + [ + -10.451626777648926, + false + ], + [ + -15.836602210998535, + false + ], + [ + -11.835906028747559, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "37fc0f6caf5b17531830efba868ae0162dcf3e8e5a3a619bd6b9c917b96a321b", + "prompt_hash": "b8d3f719e0aa2b23ede74e5f04b31d155a25bf5d6fc7eade785b66a3af24ea0f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 10, + "doc": { + "question": "What is the number of activated parameters in Qwen2-57B-A14B?", + "choices": [ + "12B", + "14B", + "30B", + "57B" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the number of activated parameters in Qwen2-57B-A14B?\nAnswer:", + " 12B" + ], + [ + "Question:What is the number of activated parameters in Qwen2-57B-A14B?\nAnswer:", + " 14B" + ], + [ + "Question:What is the number of activated parameters in Qwen2-57B-A14B?\nAnswer:", + " 30B" + ], + [ + "Question:What is the number of activated parameters in Qwen2-57B-A14B?\nAnswer:", + " 57B" + ] + ], + "resps": [ + [ + [ + -8.936101913452148, + false + ] + ], + [ + [ + -7.763552188873291, + false + ] + ], + [ + [ + -10.072871208190918, + false + ] + ], + [ + [ + -7.646878242492676, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.936101913452148, + false + ], + [ + -7.763552188873291, + false + ], + [ + -10.072871208190918, + false + ], + [ + -7.646878242492676, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "291e995ab3bccb0fa73fd6bb6aaa4aa51ad07b89c69502bf7bba0ac9e0cda457", + "prompt_hash": "d9bf57b047928f08c5addb57d1272fd20ecd564c0618409f6f817f3de20824c6", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 11, + "doc": { + "question": "Which benchmark dataset uses 25-shot prompting for evaluation?", + "choices": [ + "ARC-C", + "GPQA", + "MMLU", + "TruthfulQA" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which benchmark dataset uses 25-shot prompting for evaluation?\nAnswer:", + " ARC-C" + ], + [ + "Question:Which benchmark dataset uses 25-shot prompting for evaluation?\nAnswer:", + " GPQA" + ], + [ + "Question:Which benchmark dataset uses 25-shot prompting for evaluation?\nAnswer:", + " MMLU" + ], + [ + "Question:Which benchmark dataset uses 25-shot prompting for evaluation?\nAnswer:", + " TruthfulQA" + ] + ], + "resps": [ + [ + [ + -9.260011672973633, + false + ] + ], + [ + [ + -12.34814167022705, + false + ] + ], + [ + [ + -5.928160667419434, + false + ] + ], + [ + [ + -9.763457298278809, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.260011672973633, + false + ], + [ + -12.34814167022705, + false + ], + [ + -5.928160667419434, + false + ], + [ + -9.763457298278809, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "410e02e0a3e49a49831d1bc18cf1089e2c13a77e008734557ccc741fc5f69a62", + "prompt_hash": "d2c298b8489b04c8e91b01c98dc9a3b1e72049354d6e818adba81be34aa6290e", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 12, + "doc": { + "question": "Which dataset did Qwen2-72B achieve the highest score on among the coding benchmarks?", + "choices": [ + "HumanEval", + "MBPP", + "EvalPlus", + "MultiPL-E" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which dataset did Qwen2-72B achieve the highest score on among the coding benchmarks?\nAnswer:", + " HumanEval" + ], + [ + "Question:Which dataset did Qwen2-72B achieve the highest score on among the coding benchmarks?\nAnswer:", + " MBPP" + ], + [ + "Question:Which dataset did Qwen2-72B achieve the highest score on among the coding benchmarks?\nAnswer:", + " EvalPlus" + ], + [ + "Question:Which dataset did Qwen2-72B achieve the highest score on among the coding benchmarks?\nAnswer:", + " MultiPL-E" + ] + ], + "resps": [ + [ + [ + -7.070164203643799, + false + ] + ], + [ + [ + -9.641318321228027, + false + ] + ], + [ + [ + -12.466957092285156, + false + ] + ], + [ + [ + -24.767515182495117, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.070164203643799, + false + ], + [ + -9.641318321228027, + false + ], + [ + -12.466957092285156, + false + ], + [ + -24.767515182495117, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9856e8c7bc348dff04f4b4a132918f0f1eb2affb840d3ca01d81a61ea651d092", + "prompt_hash": "99510d8e3c0f104f96e5d57839543c71b371b81be7e5f884f54c54078d7920f7", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 13, + "doc": { + "question": "What is the MMLU score achieved by Qwen2-7B?", + "choices": [ + "64.2", + "66.6", + "61.0", + "70.3" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the MMLU score achieved by Qwen2-7B?\nAnswer:", + " 64.2" + ], + [ + "Question:What is the MMLU score achieved by Qwen2-7B?\nAnswer:", + " 66.6" + ], + [ + "Question:What is the MMLU score achieved by Qwen2-7B?\nAnswer:", + " 61.0" + ], + [ + "Question:What is the MMLU score achieved by Qwen2-7B?\nAnswer:", + " 70.3" + ] + ], + "resps": [ + [ + [ + -8.043088912963867, + false + ] + ], + [ + [ + -8.122020721435547, + false + ] + ], + [ + [ + -8.118026733398438, + false + ] + ], + [ + [ + -8.125846862792969, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.043088912963867, + false + ], + [ + -8.122020721435547, + false + ], + [ + -8.118026733398438, + false + ], + [ + -8.125846862792969, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2851a4005490cadd120a7d3628ea0f7f021360798079b32af678d54dbdea6110", + "prompt_hash": "62f178a26a549057120966177c6af7cd055f261ae620fa49ac14ff14f36ed083", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 14, + "doc": { + "question": "In the multilingual human evaluation, which language did Qwen2-72B-Instruct score the lowest in?", + "choices": [ + "Arabic", + "Korean", + "Japanese", + "Thai" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:In the multilingual human evaluation, which language did Qwen2-72B-Instruct score the lowest in?\nAnswer:", + " Arabic" + ], + [ + "Question:In the multilingual human evaluation, which language did Qwen2-72B-Instruct score the lowest in?\nAnswer:", + " Korean" + ], + [ + "Question:In the multilingual human evaluation, which language did Qwen2-72B-Instruct score the lowest in?\nAnswer:", + " Japanese" + ], + [ + "Question:In the multilingual human evaluation, which language did Qwen2-72B-Instruct score the lowest in?\nAnswer:", + " Thai" + ] + ], + "resps": [ + [ + [ + -6.313841342926025, + false + ] + ], + [ + [ + -8.17893123626709, + false + ] + ], + [ + [ + -5.562231063842773, + false + ] + ], + [ + [ + -8.58788776397705, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.313841342926025, + false + ], + [ + -8.17893123626709, + false + ], + [ + -5.562231063842773, + false + ], + [ + -8.58788776397705, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3ad3b7e606e9801b2a95c0d03c502ff29f5e16cd6d50667c6adf7073e402d078", + "prompt_hash": "5aa70bedcf6224988ae2ce4fef1f93fe7a13492bd675ae91110604a0143d3017", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 15, + "doc": { + "question": "What is the maximum number of visual tokens used by SF-LLaVA to avoid exceeding 80G GPU memory?", + "choices": [ + "4096", + "3200", + "3680", + "4608" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the maximum number of visual tokens used by SF-LLaVA to avoid exceeding 80G GPU memory?\nAnswer:", + " 4096" + ], + [ + "Question:What is the maximum number of visual tokens used by SF-LLaVA to avoid exceeding 80G GPU memory?\nAnswer:", + " 3200" + ], + [ + "Question:What is the maximum number of visual tokens used by SF-LLaVA to avoid exceeding 80G GPU memory?\nAnswer:", + " 3680" + ], + [ + "Question:What is the maximum number of visual tokens used by SF-LLaVA to avoid exceeding 80G GPU memory?\nAnswer:", + " 4608" + ] + ], + "resps": [ + [ + [ + -5.425139427185059, + false + ] + ], + [ + [ + -9.016348838806152, + false + ] + ], + [ + [ + -13.149456024169922, + false + ] + ], + [ + [ + -10.514032363891602, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.425139427185059, + false + ], + [ + -9.016348838806152, + false + ], + [ + -13.149456024169922, + false + ], + [ + -10.514032363891602, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ad13fe9a163f19221c6bf35b82ed2c73b4d4ad49359f2624d8530674e1407ac8", + "prompt_hash": "86ba060cafa27e42e47324515c56ecde20b126a762287ed01eb0fd5c0d91d0f4", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 16, + "doc": { + "question": "Which dataset is used to evaluate the Temporal Understanding (TU) of Video LLMs?", + "choices": [ + "VCGBench", + "EgoSchema", + "MSRVTT-QA", + "ActivityNet-QA" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which dataset is used to evaluate the Temporal Understanding (TU) of Video LLMs?\nAnswer:", + " VCGBench" + ], + [ + "Question:Which dataset is used to evaluate the Temporal Understanding (TU) of Video LLMs?\nAnswer:", + " EgoSchema" + ], + [ + "Question:Which dataset is used to evaluate the Temporal Understanding (TU) of Video LLMs?\nAnswer:", + " MSRVTT-QA" + ], + [ + "Question:Which dataset is used to evaluate the Temporal Understanding (TU) of Video LLMs?\nAnswer:", + " ActivityNet-QA" + ] + ], + "resps": [ + [ + [ + -27.394710540771484, + false + ] + ], + [ + [ + -17.418739318847656, + false + ] + ], + [ + [ + -16.56558609008789, + false + ] + ], + [ + [ + -15.635619163513184, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.394710540771484, + false + ], + [ + -17.418739318847656, + false + ], + [ + -16.56558609008789, + false + ], + [ + -15.635619163513184, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4b72d0430c20d5fc5c3b275e253d36baee93c4a33960828036bb124617144e65", + "prompt_hash": "6f7a3738219247c007d3fa09ddd5f4dc026be5d9ad5f5684ae8fe63f33097793", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 17, + "doc": { + "question": "What is the primary visual encoder used in SF-LLaVA?", + "choices": [ + "ViT-L/14", + "CLIP-L-14", + "BLIP-2", + "AnyRes" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the primary visual encoder used in SF-LLaVA?\nAnswer:", + " ViT-L/14" + ], + [ + "Question:What is the primary visual encoder used in SF-LLaVA?\nAnswer:", + " CLIP-L-14" + ], + [ + "Question:What is the primary visual encoder used in SF-LLaVA?\nAnswer:", + " BLIP-2" + ], + [ + "Question:What is the primary visual encoder used in SF-LLaVA?\nAnswer:", + " AnyRes" + ] + ], + "resps": [ + [ + [ + -10.990835189819336, + false + ] + ], + [ + [ + -18.62881851196289, + false + ] + ], + [ + [ + -6.553082466125488, + false + ] + ], + [ + [ + -15.750223159790039, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.990835189819336, + false + ], + [ + -18.62881851196289, + false + ], + [ + -6.553082466125488, + false + ], + [ + -15.750223159790039, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ddafe5811d17653c4db88355421d924675dc6e564970045ec933199f7d1fadd2", + "prompt_hash": "5c6a92c4e5fb08ca37a06d1e9e9abe48117262d031e354edfeca6b89b5de773d", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 18, + "doc": { + "question": "How many frames are used in the Fast pathway by default in SF-LLaVA?", + "choices": [ + "12", + "32", + "50", + "64" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many frames are used in the Fast pathway by default in SF-LLaVA?\nAnswer:", + " 12" + ], + [ + "Question:How many frames are used in the Fast pathway by default in SF-LLaVA?\nAnswer:", + " 32" + ], + [ + "Question:How many frames are used in the Fast pathway by default in SF-LLaVA?\nAnswer:", + " 50" + ], + [ + "Question:How many frames are used in the Fast pathway by default in SF-LLaVA?\nAnswer:", + " 64" + ] + ], + "resps": [ + [ + [ + -5.225262641906738, + false + ] + ], + [ + [ + -6.255336761474609, + false + ] + ], + [ + [ + -6.196992874145508, + false + ] + ], + [ + [ + -6.987529754638672, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.225262641906738, + false + ], + [ + -6.255336761474609, + false + ], + [ + -6.196992874145508, + false + ], + [ + -6.987529754638672, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1814f8766dbb483c1dc51b998d04b17576a13be811648b5484767fafac01a560", + "prompt_hash": "7eaee63b4adef0215d837ca19de0c801fd35277d6c27ba03361aa161438858f6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 19, + "doc": { + "question": "What kind of prompt does SF-LLaVA use for Multiple Choice VideoQA tasks?", + "choices": [ + "Answer based on image content", + "In this video, the answer is:", + "Best Option:(", + "Choose the correct label" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What kind of prompt does SF-LLaVA use for Multiple Choice VideoQA tasks?\nAnswer:", + " Answer based on image content" + ], + [ + "Question:What kind of prompt does SF-LLaVA use for Multiple Choice VideoQA tasks?\nAnswer:", + " In this video, the answer is:" + ], + [ + "Question:What kind of prompt does SF-LLaVA use for Multiple Choice VideoQA tasks?\nAnswer:", + " Best Option:(" + ], + [ + "Question:What kind of prompt does SF-LLaVA use for Multiple Choice VideoQA tasks?\nAnswer:", + " Choose the correct label" + ] + ], + "resps": [ + [ + [ + -21.13357162475586, + false + ] + ], + [ + [ + -18.91696548461914, + false + ] + ], + [ + [ + -25.731494903564453, + false + ] + ], + [ + [ + -16.9730281829834, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.13357162475586, + false + ], + [ + -18.91696548461914, + false + ], + [ + -25.731494903564453, + false + ], + [ + -16.9730281829834, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "67f0766718f330067f520f7a6b1b7f8182af7a2ac16cc7c42325778a9f1e77ac", + "prompt_hash": "764ec196e50ebad35d033eb2c74554ffdc888ebdc44bcb6debabfac2c87e97fe", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 20, + "doc": { + "question": "What condition must the dark energy equation of state (EoS) satisfy to cause cosmic acceleration?", + "choices": [ + "w > 0", + "w < -1/3", + "w = 0", + "w > 1" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What condition must the dark energy equation of state (EoS) satisfy to cause cosmic acceleration?\nAnswer:", + " w > 0" + ], + [ + "Question:What condition must the dark energy equation of state (EoS) satisfy to cause cosmic acceleration?\nAnswer:", + " w < -1/3" + ], + [ + "Question:What condition must the dark energy equation of state (EoS) satisfy to cause cosmic acceleration?\nAnswer:", + " w = 0" + ], + [ + "Question:What condition must the dark energy equation of state (EoS) satisfy to cause cosmic acceleration?\nAnswer:", + " w > 1" + ] + ], + "resps": [ + [ + [ + -14.41776180267334, + false + ] + ], + [ + [ + -19.103534698486328, + false + ] + ], + [ + [ + -13.807682037353516, + false + ] + ], + [ + [ + -13.923848152160645, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.41776180267334, + false + ], + [ + -19.103534698486328, + false + ], + [ + -13.807682037353516, + false + ], + [ + -13.923848152160645, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "82323cd2bfdb0979387995bd708d5e4373cea800896b67b4c7fbcbd77a25ba6c", + "prompt_hash": "85096bbeff721214820f6ef5e78686e35fa5052e6205e981155da67c7fb588a3", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 21, + "doc": { + "question": "What is the main conceptual problem associated with inserting a positive cosmological constant into Einstein's equations?", + "choices": [ + "It predicts an open universe.", + "It leads to a cyclic universe model.", + "It implies an asymptotically de Sitter universe, conflicting with quantum gravity theories.", + "It eliminates the need for dark matter." + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the main conceptual problem associated with inserting a positive cosmological constant into Einstein's equations?\nAnswer:", + " It predicts an open universe." + ], + [ + "Question:What is the main conceptual problem associated with inserting a positive cosmological constant into Einstein's equations?\nAnswer:", + " It leads to a cyclic universe model." + ], + [ + "Question:What is the main conceptual problem associated with inserting a positive cosmological constant into Einstein's equations?\nAnswer:", + " It implies an asymptotically de Sitter universe, conflicting with quantum gravity theories." + ], + [ + "Question:What is the main conceptual problem associated with inserting a positive cosmological constant into Einstein's equations?\nAnswer:", + " It eliminates the need for dark matter." + ] + ], + "resps": [ + [ + [ + -24.33345603942871, + false + ] + ], + [ + [ + -26.65534210205078, + false + ] + ], + [ + [ + -47.05223846435547, + false + ] + ], + [ + [ + -16.610246658325195, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.33345603942871, + false + ], + [ + -26.65534210205078, + false + ], + [ + -47.05223846435547, + false + ], + [ + -16.610246658325195, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "86148b6f729db3fd0ac1d7971cc469ddad875ba18641ffa5a971809b9250a4ac", + "prompt_hash": "5ab838d236da42281a56567224f23b0bbb4b61eb83a09089fafbd94761d5565f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 22, + "doc": { + "question": "Which experiment provided measurements of the CMB lensing potential using the PR4 NPIPE data release?", + "choices": [ + "ACT-DR6", + "Planck", + "DESI", + "DES" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which experiment provided measurements of the CMB lensing potential using the PR4 NPIPE data release?\nAnswer:", + " ACT-DR6" + ], + [ + "Question:Which experiment provided measurements of the CMB lensing potential using the PR4 NPIPE data release?\nAnswer:", + " Planck" + ], + [ + "Question:Which experiment provided measurements of the CMB lensing potential using the PR4 NPIPE data release?\nAnswer:", + " DESI" + ], + [ + "Question:Which experiment provided measurements of the CMB lensing potential using the PR4 NPIPE data release?\nAnswer:", + " DES" + ] + ], + "resps": [ + [ + [ + -23.941354751586914, + false + ] + ], + [ + [ + -8.566020011901855, + false + ] + ], + [ + [ + -10.961417198181152, + false + ] + ], + [ + [ + -10.11159896850586, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.941354751586914, + false + ], + [ + -8.566020011901855, + false + ], + [ + -10.961417198181152, + false + ], + [ + -10.11159896850586, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e1f9cfeac744ffec2dfa9e26688706bdd37d486ca5e583514ac532844115c757", + "prompt_hash": "a9cf649c16a1421d67dfe91a4f4556f9da722bf8231c08498c8e4ab673908ab6", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 23, + "doc": { + "question": "Which DE parameterization is characterized by a functional form w(a) = w_0 - w_a * ln a?", + "choices": [ + "CPL", + "Exponential", + "JBP", + "Logarithmic" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which DE parameterization is characterized by a functional form w(a) = w_0 - w_a * ln a?\nAnswer:", + " CPL" + ], + [ + "Question:Which DE parameterization is characterized by a functional form w(a) = w_0 - w_a * ln a?\nAnswer:", + " Exponential" + ], + [ + "Question:Which DE parameterization is characterized by a functional form w(a) = w_0 - w_a * ln a?\nAnswer:", + " JBP" + ], + [ + "Question:Which DE parameterization is characterized by a functional form w(a) = w_0 - w_a * ln a?\nAnswer:", + " Logarithmic" + ] + ], + "resps": [ + [ + [ + -11.294991493225098, + false + ] + ], + [ + [ + -8.75047779083252, + false + ] + ], + [ + [ + -15.90029525756836, + false + ] + ], + [ + [ + -10.654081344604492, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.294991493225098, + false + ], + [ + -8.75047779083252, + false + ], + [ + -15.90029525756836, + false + ], + [ + -10.654081344604492, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8dfd0f223423807451c6063af78c0afc900efca2f97c9c13ec34378122a50e73", + "prompt_hash": "709eb426ccdf71005995f3d6037e5013d2b95815c6f220f2f1c422762fb51d13", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 24, + "doc": { + "question": "Which supernova sample produced the largest statistical preference for Dynamical Dark Energy (DDE)?", + "choices": [ + "PantheonPlus", + "Union3", + "DESY5", + "SDSS-II" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which supernova sample produced the largest statistical preference for Dynamical Dark Energy (DDE)?\nAnswer:", + " PantheonPlus" + ], + [ + "Question:Which supernova sample produced the largest statistical preference for Dynamical Dark Energy (DDE)?\nAnswer:", + " Union3" + ], + [ + "Question:Which supernova sample produced the largest statistical preference for Dynamical Dark Energy (DDE)?\nAnswer:", + " DESY5" + ], + [ + "Question:Which supernova sample produced the largest statistical preference for Dynamical Dark Energy (DDE)?\nAnswer:", + " SDSS-II" + ] + ], + "resps": [ + [ + [ + -11.708585739135742, + false + ] + ], + [ + [ + -16.84764289855957, + false + ] + ], + [ + [ + -10.6693754196167, + false + ] + ], + [ + [ + -13.301567077636719, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.708585739135742, + false + ], + [ + -16.84764289855957, + false + ], + [ + -10.6693754196167, + false + ], + [ + -13.301567077636719, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "14d642bf381ec0530fe6821aa6181c2344c27825e93f56616bdcf94a79d6edd1", + "prompt_hash": "064d714bc1f66de31bbb6f9847b45cd804d589dfa7fa0fa1ef48a35905e1ba6b", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 25, + "doc": { + "question": "What is the primary challenge in global Structure-from-Motion related to translation averaging?", + "choices": [ + "Lack of feature detection methods", + "Scale ambiguity in relative translation", + "Insufficient camera rotation data", + "Inability to compute homographies" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the primary challenge in global Structure-from-Motion related to translation averaging?\nAnswer:", + " Lack of feature detection methods" + ], + [ + "Question:What is the primary challenge in global Structure-from-Motion related to translation averaging?\nAnswer:", + " Scale ambiguity in relative translation" + ], + [ + "Question:What is the primary challenge in global Structure-from-Motion related to translation averaging?\nAnswer:", + " Insufficient camera rotation data" + ], + [ + "Question:What is the primary challenge in global Structure-from-Motion related to translation averaging?\nAnswer:", + " Inability to compute homographies" + ] + ], + "resps": [ + [ + [ + -24.926162719726562, + false + ] + ], + [ + [ + -22.360336303710938, + false + ] + ], + [ + [ + -26.29901885986328, + false + ] + ], + [ + [ + -31.63737678527832, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.926162719726562, + false + ], + [ + -22.360336303710938, + false + ], + [ + -26.29901885986328, + false + ], + [ + -31.63737678527832, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f99d08f9316fce91b510f5d57417ce96c55cfbbbad46e9e0ed2510b687397869", + "prompt_hash": "a725097860ca0ad0446924e2aaecd9eeca8bbe5cbada9dca5897cb2f8366be48", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 26, + "doc": { + "question": "Which robust loss function is used in the global positioning optimization of GLOMAP?", + "choices": [ + "Mean squared error", + "Absolute error", + "Huber loss", + "Log-cosh loss" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which robust loss function is used in the global positioning optimization of GLOMAP?\nAnswer:", + " Mean squared error" + ], + [ + "Question:Which robust loss function is used in the global positioning optimization of GLOMAP?\nAnswer:", + " Absolute error" + ], + [ + "Question:Which robust loss function is used in the global positioning optimization of GLOMAP?\nAnswer:", + " Huber loss" + ], + [ + "Question:Which robust loss function is used in the global positioning optimization of GLOMAP?\nAnswer:", + " Log-cosh loss" + ] + ], + "resps": [ + [ + [ + -10.13412857055664, + false + ] + ], + [ + [ + -11.498472213745117, + false + ] + ], + [ + [ + -8.70973014831543, + false + ] + ], + [ + [ + -22.59506607055664, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.13412857055664, + false + ], + [ + -11.498472213745117, + false + ], + [ + -8.70973014831543, + false + ], + [ + -22.59506607055664, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "77c25d36e232925c9f49584d45188e8d9c9f8d946821699754d5bc19f98e8279", + "prompt_hash": "7eca0f7c3a1febfdd9b41f98198073faa3be68055d4641da35e96e285afc07dc", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 27, + "doc": { + "question": "Which datasets are used in the paper to evaluate performance on sequential data with millimeter-accurate ground truth?", + "choices": [ + "IMC 2023 and MIP360", + "ETH3D SLAM and ETH3D MVS (rig)", + "Strecha and KITTI", + "LaMAR and VisualSfM" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which datasets are used in the paper to evaluate performance on sequential data with millimeter-accurate ground truth?\nAnswer:", + " IMC 2023 and MIP360" + ], + [ + "Question:Which datasets are used in the paper to evaluate performance on sequential data with millimeter-accurate ground truth?\nAnswer:", + " ETH3D SLAM and ETH3D MVS (rig)" + ], + [ + "Question:Which datasets are used in the paper to evaluate performance on sequential data with millimeter-accurate ground truth?\nAnswer:", + " Strecha and KITTI" + ], + [ + "Question:Which datasets are used in the paper to evaluate performance on sequential data with millimeter-accurate ground truth?\nAnswer:", + " LaMAR and VisualSfM" + ] + ], + "resps": [ + [ + [ + -47.251590728759766, + false + ] + ], + [ + [ + -51.58158874511719, + false + ] + ], + [ + [ + -32.44324493408203, + false + ] + ], + [ + [ + -46.49724578857422, + false + ] + ] + ], + "filtered_resps": [ + [ + -47.251590728759766, + false + ], + [ + -51.58158874511719, + false + ], + [ + -32.44324493408203, + false + ], + [ + -46.49724578857422, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2df268151d36dc6a519f2de02a98aeffc1c67b48b4cc66a2316d317982acd099", + "prompt_hash": "2a3241ff08915e7beb5070a53126698c15aa77fe6abdadcb898f4ba796b9f07f", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 28, + "doc": { + "question": "Which optimization method is applied during global bundle adjustment in the GLOMAP pipeline?", + "choices": [ + "Gradient descent", + "Stochastic gradient descent", + "Levenberg–Marquardt", + "Newton-Raphson" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which optimization method is applied during global bundle adjustment in the GLOMAP pipeline?\nAnswer:", + " Gradient descent" + ], + [ + "Question:Which optimization method is applied during global bundle adjustment in the GLOMAP pipeline?\nAnswer:", + " Stochastic gradient descent" + ], + [ + "Question:Which optimization method is applied during global bundle adjustment in the GLOMAP pipeline?\nAnswer:", + " Levenberg–Marquardt" + ], + [ + "Question:Which optimization method is applied during global bundle adjustment in the GLOMAP pipeline?\nAnswer:", + " Newton-Raphson" + ] + ], + "resps": [ + [ + [ + -7.872706890106201, + false + ] + ], + [ + [ + -10.082393646240234, + false + ] + ], + [ + [ + -11.263614654541016, + false + ] + ], + [ + [ + -9.485450744628906, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.872706890106201, + false + ], + [ + -10.082393646240234, + false + ], + [ + -11.263614654541016, + false + ], + [ + -9.485450744628906, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "178654b53a9965290627138bc1a447ef847b56db76ca0dada65316f74dc3afc7", + "prompt_hash": "60df566c5bdf7cca2faa297f7a4005c0307ba8f475cd2f1c17d78c6af9629a03", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 29, + "doc": { + "question": "Which open-source global SfM library is noted for using the LUD method for translation averaging?", + "choices": [ + "OpenMVG", + "Theia", + "COLMAP", + "PixSfM" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which open-source global SfM library is noted for using the LUD method for translation averaging?\nAnswer:", + " OpenMVG" + ], + [ + "Question:Which open-source global SfM library is noted for using the LUD method for translation averaging?\nAnswer:", + " Theia" + ], + [ + "Question:Which open-source global SfM library is noted for using the LUD method for translation averaging?\nAnswer:", + " COLMAP" + ], + [ + "Question:Which open-source global SfM library is noted for using the LUD method for translation averaging?\nAnswer:", + " PixSfM" + ] + ], + "resps": [ + [ + [ + -15.288139343261719, + false + ] + ], + [ + [ + -18.35643768310547, + false + ] + ], + [ + [ + -7.876762390136719, + false + ] + ], + [ + [ + -8.51231575012207, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.288139343261719, + false + ], + [ + -18.35643768310547, + false + ], + [ + -7.876762390136719, + false + ], + [ + -8.51231575012207, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c1596459ab419391f5f7ede5691aa505a70638bb2de19354c58c6248c362e109", + "prompt_hash": "295125479a954ebe2e6c8d1fb4f3e35924196041365f10421b3e29a2c0b83a74", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 30, + "doc": { + "question": "What is the average success rate (ASR) of human red teaming on CYGNET before circuit breaking?", + "choices": [ + "A. 70.4%", + "B. 77.5%", + "C. 7.1%", + "D. 69.2%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the average success rate (ASR) of human red teaming on CYGNET before circuit breaking?\nAnswer:", + " A. 70.4%" + ], + [ + "Question:What is the average success rate (ASR) of human red teaming on CYGNET before circuit breaking?\nAnswer:", + " B. 77.5%" + ], + [ + "Question:What is the average success rate (ASR) of human red teaming on CYGNET before circuit breaking?\nAnswer:", + " C. 7.1%" + ], + [ + "Question:What is the average success rate (ASR) of human red teaming on CYGNET before circuit breaking?\nAnswer:", + " D. 69.2%" + ] + ], + "resps": [ + [ + [ + -22.771133422851562, + false + ] + ], + [ + [ + -26.42609405517578, + false + ] + ], + [ + [ + -26.433547973632812, + false + ] + ], + [ + [ + -27.742061614990234, + false + ] + ] + ], + "filtered_resps": [ + [ + -22.771133422851562, + false + ], + [ + -26.42609405517578, + false + ], + [ + -26.433547973632812, + false + ], + [ + -27.742061614990234, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2d216ad67d2d149e00496b388a15d9ac43439c72be3eca6ce5bec1c2f9a37076", + "prompt_hash": "e07b5748a3f14c1055fbf1e15ba87ad68c3b73db2b1981ba77687205852aa3f2", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 31, + "doc": { + "question": "Which tactic involves layering different fictional scenarios within a fictional scenario?", + "choices": [ + "A. Roleplay", + "B. Opposite Day", + "C. Recursion", + "D. Allegory" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which tactic involves layering different fictional scenarios within a fictional scenario?\nAnswer:", + " A. Roleplay" + ], + [ + "Question:Which tactic involves layering different fictional scenarios within a fictional scenario?\nAnswer:", + " B. Opposite Day" + ], + [ + "Question:Which tactic involves layering different fictional scenarios within a fictional scenario?\nAnswer:", + " C. Recursion" + ], + [ + "Question:Which tactic involves layering different fictional scenarios within a fictional scenario?\nAnswer:", + " D. Allegory" + ] + ], + "resps": [ + [ + [ + -15.934167861938477, + false + ] + ], + [ + [ + -27.397541046142578, + false + ] + ], + [ + [ + -21.708358764648438, + false + ] + ], + [ + [ + -20.459033966064453, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.934167861938477, + false + ], + [ + -27.397541046142578, + false + ], + [ + -21.708358764648438, + false + ], + [ + -20.459033966064453, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f723e49506add63585f0863f4da23e2998a1ba84e8c0c774b6979bb327c52286", + "prompt_hash": "0990249ca41f5db19b4b4e6d9ff3106bf97a20fcc4eba5a607f4ce2e5468fd41", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 32, + "doc": { + "question": "How many harmful behaviors are in HarmBench's standard and contextual test set?", + "choices": [ + "A. 120", + "B. 240", + "C. 360", + "D. 480" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many harmful behaviors are in HarmBench's standard and contextual test set?\nAnswer:", + " A. 120" + ], + [ + "Question:How many harmful behaviors are in HarmBench's standard and contextual test set?\nAnswer:", + " B. 240" + ], + [ + "Question:How many harmful behaviors are in HarmBench's standard and contextual test set?\nAnswer:", + " C. 360" + ], + [ + "Question:How many harmful behaviors are in HarmBench's standard and contextual test set?\nAnswer:", + " D. 480" + ] + ], + "resps": [ + [ + [ + -17.521814346313477, + false + ] + ], + [ + [ + -23.202856063842773, + false + ] + ], + [ + [ + -22.82879638671875, + false + ] + ], + [ + [ + -24.634403228759766, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.521814346313477, + false + ], + [ + -23.202856063842773, + false + ], + [ + -22.82879638671875, + false + ], + [ + -24.634403228759766, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e52d539c2546e24b6326d4ead53f0b17c674a1670d2c556d896fd19da605e7a1", + "prompt_hash": "39b763fbfc320c740fe3ca2c9f753f8bcbe864adfe0f3fa96efed184d3653b16", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 33, + "doc": { + "question": "Which defense mechanism is applied via API access and is not open-source?", + "choices": [ + "A. LAT", + "B. DERTA", + "C. CYGNET", + "D. RR" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which defense mechanism is applied via API access and is not open-source?\nAnswer:", + " A. LAT" + ], + [ + "Question:Which defense mechanism is applied via API access and is not open-source?\nAnswer:", + " B. DERTA" + ], + [ + "Question:Which defense mechanism is applied via API access and is not open-source?\nAnswer:", + " C. CYGNET" + ], + [ + "Question:Which defense mechanism is applied via API access and is not open-source?\nAnswer:", + " D. RR" + ] + ], + "resps": [ + [ + [ + -21.102420806884766, + false + ] + ], + [ + [ + -32.22483444213867, + false + ] + ], + [ + [ + -30.71807098388672, + false + ] + ], + [ + [ + -22.92583656311035, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.102420806884766, + false + ], + [ + -32.22483444213867, + false + ], + [ + -30.71807098388672, + false + ], + [ + -22.92583656311035, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c31587319e9576d92e88a14e3a5d10106bb01d195060f1c3c2701e592ca7768d", + "prompt_hash": "e7519de7eadcc94f3aa9761a55eecf5ff23e51af685818f4ceafcf5e1e7443ca", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 34, + "doc": { + "question": "Which unlearning method is applied on llama-3-8b-instruct in the evaluation?", + "choices": [ + "A. LAT", + "B. RR", + "C. RMU", + "D. CYGNET" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which unlearning method is applied on llama-3-8b-instruct in the evaluation?\nAnswer:", + " A. LAT" + ], + [ + "Question:Which unlearning method is applied on llama-3-8b-instruct in the evaluation?\nAnswer:", + " B. RR" + ], + [ + "Question:Which unlearning method is applied on llama-3-8b-instruct in the evaluation?\nAnswer:", + " C. RMU" + ], + [ + "Question:Which unlearning method is applied on llama-3-8b-instruct in the evaluation?\nAnswer:", + " D. CYGNET" + ] + ], + "resps": [ + [ + [ + -21.075618743896484, + false + ] + ], + [ + [ + -23.058002471923828, + false + ] + ], + [ + [ + -26.396194458007812, + false + ] + ], + [ + [ + -36.67958068847656, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.075618743896484, + false + ], + [ + -23.058002471923828, + false + ], + [ + -26.396194458007812, + false + ], + [ + -36.67958068847656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f4a7fceca1cd86a4f8a0d7d203d463a77882225cc6b1bf417ab9e93108068ad0", + "prompt_hash": "a7a82305a6edbec80a6ece0b8791b8462b52fe8847dfbb00d98e2346ed3ca779", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 35, + "doc": { + "question": "What is the total number of parameters in the GOT model?", + "choices": [ + "80M", + "500M", + "580M", + "1B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the total number of parameters in the GOT model?\nAnswer:", + " 80M" + ], + [ + "Question:What is the total number of parameters in the GOT model?\nAnswer:", + " 500M" + ], + [ + "Question:What is the total number of parameters in the GOT model?\nAnswer:", + " 580M" + ], + [ + "Question:What is the total number of parameters in the GOT model?\nAnswer:", + " 1B" + ] + ], + "resps": [ + [ + [ + -8.71826457977295, + false + ] + ], + [ + [ + -9.68232536315918, + false + ] + ], + [ + [ + -11.287875175476074, + false + ] + ], + [ + [ + -9.855000495910645, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.71826457977295, + false + ], + [ + -9.68232536315918, + false + ], + [ + -11.287875175476074, + false + ], + [ + -9.855000495910645, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7380787e1c057f5435f570a3bbc5da525962ce15f476f949afecb4e65921862e", + "prompt_hash": "5e469dc8fe8590f2dc84e502a359096b01f776c79608066fe25c79a1720b858d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 36, + "doc": { + "question": "Which dataset was used to extract pseudo ground truth for scene text OCR data in the GOT pre-training stage?", + "choices": [ + "SynthText", + "PaddleOCR", + "Tesseract", + "TextOCR" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset was used to extract pseudo ground truth for scene text OCR data in the GOT pre-training stage?\nAnswer:", + " SynthText" + ], + [ + "Question:Which dataset was used to extract pseudo ground truth for scene text OCR data in the GOT pre-training stage?\nAnswer:", + " PaddleOCR" + ], + [ + "Question:Which dataset was used to extract pseudo ground truth for scene text OCR data in the GOT pre-training stage?\nAnswer:", + " Tesseract" + ], + [ + "Question:Which dataset was used to extract pseudo ground truth for scene text OCR data in the GOT pre-training stage?\nAnswer:", + " TextOCR" + ] + ], + "resps": [ + [ + [ + -11.177651405334473, + false + ] + ], + [ + [ + -9.446195602416992, + false + ] + ], + [ + [ + -8.80298900604248, + false + ] + ], + [ + [ + -10.531850814819336, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.177651405334473, + false + ], + [ + -9.446195602416992, + false + ], + [ + -8.80298900604248, + false + ], + [ + -10.531850814819336, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "994deb6439a522548ebd5a163665b864cc2c0aa755b14b8cb6f0cc432c3a0309", + "prompt_hash": "8cac60a6fb45ae86eba482ac1556dee5bc0271a924826a110316a0f98106b562", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 37, + "doc": { + "question": "Which rendering tool was used to generate images for geometric shapes in GOT's dataset?", + "choices": [ + "Verovio", + "TikZ", + "Mathpix", + "Matplotlib" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which rendering tool was used to generate images for geometric shapes in GOT's dataset?\nAnswer:", + " Verovio" + ], + [ + "Question:Which rendering tool was used to generate images for geometric shapes in GOT's dataset?\nAnswer:", + " TikZ" + ], + [ + "Question:Which rendering tool was used to generate images for geometric shapes in GOT's dataset?\nAnswer:", + " Mathpix" + ], + [ + "Question:Which rendering tool was used to generate images for geometric shapes in GOT's dataset?\nAnswer:", + " Matplotlib" + ] + ], + "resps": [ + [ + [ + -15.639054298400879, + false + ] + ], + [ + [ + -11.802534103393555, + false + ] + ], + [ + [ + -16.534772872924805, + false + ] + ], + [ + [ + -11.40635871887207, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.639054298400879, + false + ], + [ + -11.802534103393555, + false + ], + [ + -16.534772872924805, + false + ], + [ + -11.40635871887207, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7cfda73fc25b4dd06193ffc4b04cfbaa74dc7b5eccaf4c5ae4c08835ab0a5c34", + "prompt_hash": "1dca76e7c2d4d469366a1fb76962c55c4f505b97dbfd9d4c9b94834dd1c7c65e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 38, + "doc": { + "question": "What decoder model is used in the final architecture of GOT?", + "choices": [ + "LLaMA-2 7B", + "OPT-125M", + "Qwen-0.5B", + "GPT-2" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What decoder model is used in the final architecture of GOT?\nAnswer:", + " LLaMA-2 7B" + ], + [ + "Question:What decoder model is used in the final architecture of GOT?\nAnswer:", + " OPT-125M" + ], + [ + "Question:What decoder model is used in the final architecture of GOT?\nAnswer:", + " Qwen-0.5B" + ], + [ + "Question:What decoder model is used in the final architecture of GOT?\nAnswer:", + " GPT-2" + ] + ], + "resps": [ + [ + [ + -14.056198120117188, + false + ] + ], + [ + [ + -14.997190475463867, + false + ] + ], + [ + [ + -16.97852897644043, + false + ] + ], + [ + [ + -7.084348678588867, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.056198120117188, + false + ], + [ + -14.997190475463867, + false + ], + [ + -16.97852897644043, + false + ], + [ + -7.084348678588867, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b6d651e38fd2fcd7c40a7bbb536c4c771234618ad561faf300b5f1e81aadd4c0", + "prompt_hash": "dcc0ed11c1112a6cf430d28aa636753877743b73b960590f90a3ac0c42185762", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 39, + "doc": { + "question": "How many image-text pairs are used in the pre-training of the GOT vision encoder?", + "choices": [ + "3 million", + "5 million", + "8 million", + "10 million" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many image-text pairs are used in the pre-training of the GOT vision encoder?\nAnswer:", + " 3 million" + ], + [ + "Question:How many image-text pairs are used in the pre-training of the GOT vision encoder?\nAnswer:", + " 5 million" + ], + [ + "Question:How many image-text pairs are used in the pre-training of the GOT vision encoder?\nAnswer:", + " 8 million" + ], + [ + "Question:How many image-text pairs are used in the pre-training of the GOT vision encoder?\nAnswer:", + " 10 million" + ] + ], + "resps": [ + [ + [ + -9.339093208312988, + false + ] + ], + [ + [ + -9.036966323852539, + false + ] + ], + [ + [ + -10.546751022338867, + false + ] + ], + [ + [ + -8.766083717346191, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.339093208312988, + false + ], + [ + -9.036966323852539, + false + ], + [ + -10.546751022338867, + false + ], + [ + -8.766083717346191, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8b7f7f937740eb1703167a07d32f5fca12fae238eb7612113ec02e2feb40ae65", + "prompt_hash": "d6b8f044cc916afe2580b1ac3d16140829d110e6f941af99d0bebf937f4a4aaf", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 40, + "doc": { + "question": "What does the ViewCrafter framework use to achieve precise control of 6 DoF camera poses?", + "choices": [ + "Text prompts and depth estimation", + "Latent space embeddings and video warping", + "Point cloud representations and video diffusion models", + "2D image-based monocular estimations" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does the ViewCrafter framework use to achieve precise control of 6 DoF camera poses?\nAnswer:", + " Text prompts and depth estimation" + ], + [ + "Question:What does the ViewCrafter framework use to achieve precise control of 6 DoF camera poses?\nAnswer:", + " Latent space embeddings and video warping" + ], + [ + "Question:What does the ViewCrafter framework use to achieve precise control of 6 DoF camera poses?\nAnswer:", + " Point cloud representations and video diffusion models" + ], + [ + "Question:What does the ViewCrafter framework use to achieve precise control of 6 DoF camera poses?\nAnswer:", + " 2D image-based monocular estimations" + ] + ], + "resps": [ + [ + [ + -23.642772674560547, + false + ] + ], + [ + [ + -31.760562896728516, + false + ] + ], + [ + [ + -31.7762451171875, + false + ] + ], + [ + [ + -29.73705291748047, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.642772674560547, + false + ], + [ + -31.760562896728516, + false + ], + [ + -31.7762451171875, + false + ], + [ + -29.73705291748047, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "20efdb6f39d6e770c732dcdf1c0b35244cedabe9d6dcd0701091d58548cf74ab", + "prompt_hash": "f01ce7509a42c001a8b4687c85922b07e4e5b3e7e3646bca906850209df2650a", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 41, + "doc": { + "question": "Which model is used as the base video diffusion model in ViewCrafter?", + "choices": [ + "AnimateDiff", + "DynamiCrafter", + "ZeroNVS", + "LucidDreamer" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which model is used as the base video diffusion model in ViewCrafter?\nAnswer:", + " AnimateDiff" + ], + [ + "Question:Which model is used as the base video diffusion model in ViewCrafter?\nAnswer:", + " DynamiCrafter" + ], + [ + "Question:Which model is used as the base video diffusion model in ViewCrafter?\nAnswer:", + " ZeroNVS" + ], + [ + "Question:Which model is used as the base video diffusion model in ViewCrafter?\nAnswer:", + " LucidDreamer" + ] + ], + "resps": [ + [ + [ + -11.66192626953125, + false + ] + ], + [ + [ + -11.834842681884766, + false + ] + ], + [ + [ + -18.940771102905273, + false + ] + ], + [ + [ + -12.462959289550781, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.66192626953125, + false + ], + [ + -11.834842681884766, + false + ], + [ + -18.940771102905273, + false + ], + [ + -12.462959289550781, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0dd03d82ddbc637e00a05439ac2ea544b4b5bcb60d72a3b24720bb57f23ce709", + "prompt_hash": "c533106d2d00d3188a020b9dcead6a9543bdb2b762a234e8f4deba5e3bf21ea3", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 42, + "doc": { + "question": "What algorithm does ViewCrafter employ to determine the next best camera pose during iterative synthesis?", + "choices": [ + "Random walk sampling", + "Fixed circular trajectory", + "Next-Best-View (NBV) planning", + "Gradient-based motion estimation" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What algorithm does ViewCrafter employ to determine the next best camera pose during iterative synthesis?\nAnswer:", + " Random walk sampling" + ], + [ + "Question:What algorithm does ViewCrafter employ to determine the next best camera pose during iterative synthesis?\nAnswer:", + " Fixed circular trajectory" + ], + [ + "Question:What algorithm does ViewCrafter employ to determine the next best camera pose during iterative synthesis?\nAnswer:", + " Next-Best-View (NBV) planning" + ], + [ + "Question:What algorithm does ViewCrafter employ to determine the next best camera pose during iterative synthesis?\nAnswer:", + " Gradient-based motion estimation" + ] + ], + "resps": [ + [ + [ + -17.710290908813477, + false + ] + ], + [ + [ + -23.67266082763672, + false + ] + ], + [ + [ + -24.032934188842773, + false + ] + ], + [ + [ + -19.99657440185547, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.710290908813477, + false + ], + [ + -23.67266082763672, + false + ], + [ + -24.032934188842773, + false + ], + [ + -19.99657440185547, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8a425b4867abc928d2df97796ef238fe6f5ed803bee6c22e708315eeb0f1e59b", + "prompt_hash": "8a872d35921534c9cd4fdd3e6bac16a8947410679a755b76e5afe39b674f49cc", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 43, + "doc": { + "question": "Which datasets are used to train the ViewCrafter model?", + "choices": [ + "ScanNet and CO3D", + "DL3DV and RealEstate10K", + "ShapeNet and COCO", + "Tanks-and-Temples and Matterport3D" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which datasets are used to train the ViewCrafter model?\nAnswer:", + " ScanNet and CO3D" + ], + [ + "Question:Which datasets are used to train the ViewCrafter model?\nAnswer:", + " DL3DV and RealEstate10K" + ], + [ + "Question:Which datasets are used to train the ViewCrafter model?\nAnswer:", + " ShapeNet and COCO" + ], + [ + "Question:Which datasets are used to train the ViewCrafter model?\nAnswer:", + " Tanks-and-Temples and Matterport3D" + ] + ], + "resps": [ + [ + [ + -17.817087173461914, + false + ] + ], + [ + [ + -42.86650466918945, + false + ] + ], + [ + [ + -19.645099639892578, + false + ] + ], + [ + [ + -27.194059371948242, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.817087173461914, + false + ], + [ + -42.86650466918945, + false + ], + [ + -19.645099639892578, + false + ], + [ + -27.194059371948242, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a80bdbf698650f3e6766a717561335bd14e5e7f1f2567eecbc983bb009bb4fb5", + "prompt_hash": "48f041efbc3663626df4f54a69d1bfcf3f8393fdef02d683f8686aba0b98cd46", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 44, + "doc": { + "question": "Which method does ViewCrafter use to reconstruct the point cloud from sparse input images?", + "choices": [ + "COLMAP", + "Structure-from-Motion", + "DUSt3R", + "Plücker coordinate regression" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which method does ViewCrafter use to reconstruct the point cloud from sparse input images?\nAnswer:", + " COLMAP" + ], + [ + "Question:Which method does ViewCrafter use to reconstruct the point cloud from sparse input images?\nAnswer:", + " Structure-from-Motion" + ], + [ + "Question:Which method does ViewCrafter use to reconstruct the point cloud from sparse input images?\nAnswer:", + " DUSt3R" + ], + [ + "Question:Which method does ViewCrafter use to reconstruct the point cloud from sparse input images?\nAnswer:", + " Plücker coordinate regression" + ] + ], + "resps": [ + [ + [ + -10.547996520996094, + false + ] + ], + [ + [ + -9.73310375213623, + false + ] + ], + [ + [ + -32.7406005859375, + false + ] + ], + [ + [ + -33.1044921875, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.547996520996094, + false + ], + [ + -9.73310375213623, + false + ], + [ + -32.7406005859375, + false + ], + [ + -33.1044921875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4e725f4f1dd36dad41d72a0ef19b65be68e912ec0bb12af51288fe5c9a242232", + "prompt_hash": "fbcbd75cc4ed82855cd63198990031d0dcf1bca655ab55f830aedb5b3bf69e46", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 45, + "doc": { + "question": "Which benchmark includes complex scenarios like parallel and multiple function calls in languages such as Java, JavaScript, and Python?", + "choices": [ + "AgentBench", + "ToolBench", + "ToolQuery", + "Berkeley Function-Calling Benchmark" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which benchmark includes complex scenarios like parallel and multiple function calls in languages such as Java, JavaScript, and Python?\nAnswer:", + " AgentBench" + ], + [ + "Question:Which benchmark includes complex scenarios like parallel and multiple function calls in languages such as Java, JavaScript, and Python?\nAnswer:", + " ToolBench" + ], + [ + "Question:Which benchmark includes complex scenarios like parallel and multiple function calls in languages such as Java, JavaScript, and Python?\nAnswer:", + " ToolQuery" + ], + [ + "Question:Which benchmark includes complex scenarios like parallel and multiple function calls in languages such as Java, JavaScript, and Python?\nAnswer:", + " Berkeley Function-Calling Benchmark" + ] + ], + "resps": [ + [ + [ + -12.500007629394531, + false + ] + ], + [ + [ + -13.430700302124023, + false + ] + ], + [ + [ + -22.2950439453125, + false + ] + ], + [ + [ + -26.18708610534668, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.500007629394531, + false + ], + [ + -13.430700302124023, + false + ], + [ + -22.2950439453125, + false + ], + [ + -26.18708610534668, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "50e437a64770455b952405b5a456fde6bac837d3fbeacbcde5eb5eacb46b1d1e", + "prompt_hash": "c7c05251e7a48d01e77bfe1c75830f826e9b4fc4fadf5119c27132a213fbbd0a", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 46, + "doc": { + "question": "Which data synthesis framework is used to generate verifiable datasets based on executable APIs?", + "choices": [ + "ToolBench", + "AgentOhana", + "APIGen", + "DialogStudio" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which data synthesis framework is used to generate verifiable datasets based on executable APIs?\nAnswer:", + " ToolBench" + ], + [ + "Question:Which data synthesis framework is used to generate verifiable datasets based on executable APIs?\nAnswer:", + " AgentOhana" + ], + [ + "Question:Which data synthesis framework is used to generate verifiable datasets based on executable APIs?\nAnswer:", + " APIGen" + ], + [ + "Question:Which data synthesis framework is used to generate verifiable datasets based on executable APIs?\nAnswer:", + " DialogStudio" + ] + ], + "resps": [ + [ + [ + -17.682613372802734, + false + ] + ], + [ + [ + -25.438610076904297, + false + ] + ], + [ + [ + -10.732872009277344, + false + ] + ], + [ + [ + -22.601285934448242, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.682613372802734, + false + ], + [ + -25.438610076904297, + false + ], + [ + -10.732872009277344, + false + ], + [ + -22.601285934448242, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6bdea1c9b0ac67637689735ba7b180e755376e3f84be30e1bf448190ed012d64", + "prompt_hash": "3c98548a57b1c17f173a9e5501a877d63d3d69d4993c2749ba04d30a62713361", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 47, + "doc": { + "question": "What type of augmentation involves rephrasing instructions and verifying them with LLMs to ensure consistency?", + "choices": [ + "Prompt Format Augmentation", + "Order Shuffling", + "Instruction-Following Augmentation", + "Concatenation Tokens" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What type of augmentation involves rephrasing instructions and verifying them with LLMs to ensure consistency?\nAnswer:", + " Prompt Format Augmentation" + ], + [ + "Question:What type of augmentation involves rephrasing instructions and verifying them with LLMs to ensure consistency?\nAnswer:", + " Order Shuffling" + ], + [ + "Question:What type of augmentation involves rephrasing instructions and verifying them with LLMs to ensure consistency?\nAnswer:", + " Instruction-Following Augmentation" + ], + [ + "Question:What type of augmentation involves rephrasing instructions and verifying them with LLMs to ensure consistency?\nAnswer:", + " Concatenation Tokens" + ] + ], + "resps": [ + [ + [ + -18.61235809326172, + false + ] + ], + [ + [ + -20.94488525390625, + false + ] + ], + [ + [ + -13.356648445129395, + false + ] + ], + [ + [ + -25.019317626953125, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.61235809326172, + false + ], + [ + -20.94488525390625, + false + ], + [ + -13.356648445129395, + false + ], + [ + -25.019317626953125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0c1b6a5d3000ee881d1e8dded9f31d9fb738710106c6552e7ab4f1d80cd1a4f1", + "prompt_hash": "2ae7793b6c6d78c339d6de0dc6bbb465497256c56f8735df1ef4cd965487f3b6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 48, + "doc": { + "question": "What is the smallest model in the xLAM series mentioned, and what is its approximate accuracy on the BFCL v2 benchmark?", + "choices": [ + "xLAM-1b-fc-r with 75.43%", + "xLAM-7b-r with 80.33%", + "xLAM-8x22b-r with 87.31%", + "xLAM-8x7b-r with 84.69%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the smallest model in the xLAM series mentioned, and what is its approximate accuracy on the BFCL v2 benchmark?\nAnswer:", + " xLAM-1b-fc-r with 75.43%" + ], + [ + "Question:What is the smallest model in the xLAM series mentioned, and what is its approximate accuracy on the BFCL v2 benchmark?\nAnswer:", + " xLAM-7b-r with 80.33%" + ], + [ + "Question:What is the smallest model in the xLAM series mentioned, and what is its approximate accuracy on the BFCL v2 benchmark?\nAnswer:", + " xLAM-8x22b-r with 87.31%" + ], + [ + "Question:What is the smallest model in the xLAM series mentioned, and what is its approximate accuracy on the BFCL v2 benchmark?\nAnswer:", + " xLAM-8x7b-r with 84.69%" + ] + ], + "resps": [ + [ + [ + -58.31315612792969, + false + ] + ], + [ + [ + -44.66631317138672, + false + ] + ], + [ + [ + -51.25954818725586, + false + ] + ], + [ + [ + -48.78382873535156, + false + ] + ] + ], + "filtered_resps": [ + [ + -58.31315612792969, + false + ], + [ + -44.66631317138672, + false + ], + [ + -51.25954818725586, + false + ], + [ + -48.78382873535156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6c0bc5fff5cb3e8d98cb7fa345607fe20ecaa1d0c3c3cc95554662f11a026b5f", + "prompt_hash": "bce0a721dac988be6371d832269753de01df1413eb3a30636480ef35325f7ab9", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 49, + "doc": { + "question": "Which technique is applied to avoid overfitting on a specific tool order in the training data?", + "choices": [ + "Format Instruction-Following", + "LoRA-based Fine-Tuning", + "Order Shuffling", + "Function-Calling Templates" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which technique is applied to avoid overfitting on a specific tool order in the training data?\nAnswer:", + " Format Instruction-Following" + ], + [ + "Question:Which technique is applied to avoid overfitting on a specific tool order in the training data?\nAnswer:", + " LoRA-based Fine-Tuning" + ], + [ + "Question:Which technique is applied to avoid overfitting on a specific tool order in the training data?\nAnswer:", + " Order Shuffling" + ], + [ + "Question:Which technique is applied to avoid overfitting on a specific tool order in the training data?\nAnswer:", + " Function-Calling Templates" + ] + ], + "resps": [ + [ + [ + -27.455965042114258, + false + ] + ], + [ + [ + -21.98909568786621, + false + ] + ], + [ + [ + -18.27690315246582, + false + ] + ], + [ + [ + -29.975255966186523, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.455965042114258, + false + ], + [ + -21.98909568786621, + false + ], + [ + -18.27690315246582, + false + ], + [ + -29.975255966186523, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7bf55761c45c0593f901523c8d8abff6c2a18e28c75c860b2cce9eb771f6ee34", + "prompt_hash": "52a3b30b08b978b46fc93a59102d2acf835be5a76f55ba7c83221c2637a6438a", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 50, + "doc": { + "question": "What is the internal code name of Baidu's next-generation query-ad matching project?", + "choices": [ + "Mobius", + "Phoenix Nest", + "OPQ", + "ANN" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the internal code name of Baidu's next-generation query-ad matching project?\nAnswer:", + " Mobius" + ], + [ + "Question:What is the internal code name of Baidu's next-generation query-ad matching project?\nAnswer:", + " Phoenix Nest" + ], + [ + "Question:What is the internal code name of Baidu's next-generation query-ad matching project?\nAnswer:", + " OPQ" + ], + [ + "Question:What is the internal code name of Baidu's next-generation query-ad matching project?\nAnswer:", + " ANN" + ] + ], + "resps": [ + [ + [ + -12.405713081359863, + false + ] + ], + [ + [ + -20.490856170654297, + false + ] + ], + [ + [ + -15.323873519897461, + false + ] + ], + [ + [ + -12.232057571411133, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.405713081359863, + false + ], + [ + -20.490856170654297, + false + ], + [ + -15.323873519897461, + false + ], + [ + -12.232057571411133, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "add4053f4575b941d604a1c65c539138615d9db0ae66f788419ffd4c83255035", + "prompt_hash": "aa88127607633251bb957100ec9d8356d119bc74bcdfc424d6633f46d0d4ef5f", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 51, + "doc": { + "question": "What is the business metric defined as CPM in Baidu's sponsored search?", + "choices": [ + "CPM = Bid × ROI", + "CPM = CTR × Bid", + "CPM = Clicks × ROI", + "CPM = CPC × CTR" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the business metric defined as CPM in Baidu's sponsored search?\nAnswer:", + " CPM = Bid × ROI" + ], + [ + "Question:What is the business metric defined as CPM in Baidu's sponsored search?\nAnswer:", + " CPM = CTR × Bid" + ], + [ + "Question:What is the business metric defined as CPM in Baidu's sponsored search?\nAnswer:", + " CPM = Clicks × ROI" + ], + [ + "Question:What is the business metric defined as CPM in Baidu's sponsored search?\nAnswer:", + " CPM = CPC × CTR" + ] + ], + "resps": [ + [ + [ + -32.9281120300293, + false + ] + ], + [ + [ + -25.26803207397461, + false + ] + ], + [ + [ + -30.90808868408203, + false + ] + ], + [ + [ + -18.589683532714844, + false + ] + ] + ], + "filtered_resps": [ + [ + -32.9281120300293, + false + ], + [ + -25.26803207397461, + false + ], + [ + -30.90808868408203, + false + ], + [ + -18.589683532714844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "86b40e187d2628ac02aa83671304f7187db6bccfc71eba9db2b342a4d7f78653", + "prompt_hash": "d52c15f9293fd3abce639a3ac5f5e00a7e01b38415210b2bbcce80b8baf5b40e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 52, + "doc": { + "question": "Which algorithm does Mobius-V1 use for fast ad indexing and retrieval?", + "choices": [ + "Random Partition Trees", + "Product Quantization", + "Optimized Product Quantization", + "K-Means Hashing" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which algorithm does Mobius-V1 use for fast ad indexing and retrieval?\nAnswer:", + " Random Partition Trees" + ], + [ + "Question:Which algorithm does Mobius-V1 use for fast ad indexing and retrieval?\nAnswer:", + " Product Quantization" + ], + [ + "Question:Which algorithm does Mobius-V1 use for fast ad indexing and retrieval?\nAnswer:", + " Optimized Product Quantization" + ], + [ + "Question:Which algorithm does Mobius-V1 use for fast ad indexing and retrieval?\nAnswer:", + " K-Means Hashing" + ] + ], + "resps": [ + [ + [ + -24.32925796508789, + false + ] + ], + [ + [ + -16.3418025970459, + false + ] + ], + [ + [ + -22.6213321685791, + false + ] + ], + [ + [ + -22.01201820373535, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.32925796508789, + false + ], + [ + -16.3418025970459, + false + ], + [ + -22.6213321685791, + false + ], + [ + -22.01201820373535, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "28685b874477d3310d55c078b762704fe9ce7e447e7417772159e91e24f5faed", + "prompt_hash": "eefc4ae19997e45fd3f48c0cf92fd7f448ac3545a5ae88edfa561cbd874ea861", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 53, + "doc": { + "question": "What performance improvement did Mobius-V1 achieve in CPM on the Baidu App?", + "choices": [ + "+1.0%", + "+2.9%", + "+3.8%", + "+5.0%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What performance improvement did Mobius-V1 achieve in CPM on the Baidu App?\nAnswer:", + " +1.0%" + ], + [ + "Question:What performance improvement did Mobius-V1 achieve in CPM on the Baidu App?\nAnswer:", + " +2.9%" + ], + [ + "Question:What performance improvement did Mobius-V1 achieve in CPM on the Baidu App?\nAnswer:", + " +3.8%" + ], + [ + "Question:What performance improvement did Mobius-V1 achieve in CPM on the Baidu App?\nAnswer:", + " +5.0%" + ] + ], + "resps": [ + [ + [ + -20.06861114501953, + false + ] + ], + [ + [ + -20.90535545349121, + false + ] + ], + [ + [ + -20.866792678833008, + false + ] + ], + [ + [ + -21.263343811035156, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.06861114501953, + false + ], + [ + -20.90535545349121, + false + ], + [ + -20.866792678833008, + false + ], + [ + -21.263343811035156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4e5118f4ffe10fb48bdfb0a855069d9e266c775ca731744ccd557342c37dff21", + "prompt_hash": "f3f24a3f3a904c41ff9d03ece9af0310d088dd1c3c69829375bcf4326424b28d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 54, + "doc": { + "question": "In the CTR model training, what is the role of the 'teacher' component in Mobius-V1?", + "choices": [ + "To predict final ad rankings", + "To simulate click behavior of users", + "To judge semantic relevance of query-ad pairs", + "To compress feature vectors for retrieval" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In the CTR model training, what is the role of the 'teacher' component in Mobius-V1?\nAnswer:", + " To predict final ad rankings" + ], + [ + "Question:In the CTR model training, what is the role of the 'teacher' component in Mobius-V1?\nAnswer:", + " To simulate click behavior of users" + ], + [ + "Question:In the CTR model training, what is the role of the 'teacher' component in Mobius-V1?\nAnswer:", + " To judge semantic relevance of query-ad pairs" + ], + [ + "Question:In the CTR model training, what is the role of the 'teacher' component in Mobius-V1?\nAnswer:", + " To compress feature vectors for retrieval" + ] + ], + "resps": [ + [ + [ + -35.07558059692383, + false + ] + ], + [ + [ + -27.24683380126953, + false + ] + ], + [ + [ + -40.58359909057617, + false + ] + ], + [ + [ + -33.87470245361328, + false + ] + ] + ], + "filtered_resps": [ + [ + -35.07558059692383, + false + ], + [ + -27.24683380126953, + false + ], + [ + -40.58359909057617, + false + ], + [ + -33.87470245361328, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7b2260f5f2914c1b0824e72b40840441445c51603206d3bc51ba5c7c90eef71b", + "prompt_hash": "a5715e3e45fb1e58df657cd7dfa7b3a20372d1b0f3222421b6e810d662ed1e9c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 55, + "doc": { + "question": "What is the maximum number of papers retrieved by the research ideation agent during the paper retrieval step?", + "choices": [ + "20", + "60", + "120", + "200" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the maximum number of papers retrieved by the research ideation agent during the paper retrieval step?\nAnswer:", + " 20" + ], + [ + "Question:What is the maximum number of papers retrieved by the research ideation agent during the paper retrieval step?\nAnswer:", + " 60" + ], + [ + "Question:What is the maximum number of papers retrieved by the research ideation agent during the paper retrieval step?\nAnswer:", + " 120" + ], + [ + "Question:What is the maximum number of papers retrieved by the research ideation agent during the paper retrieval step?\nAnswer:", + " 200" + ] + ], + "resps": [ + [ + [ + -3.57956862449646, + false + ] + ], + [ + [ + -5.394436836242676, + false + ] + ], + [ + [ + -6.261102199554443, + false + ] + ], + [ + [ + -5.285019397735596, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.57956862449646, + false + ], + [ + -5.394436836242676, + false + ], + [ + -6.261102199554443, + false + ], + [ + -5.285019397735596, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f646bbd461c4465b3ec2a1464705c9f7184cadc5d25e456f267250e6668570ce", + "prompt_hash": "1df44aa395f1715ec809617aad796f435ab7a39758dfc89b92b3d86bbdda0282", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 56, + "doc": { + "question": "Which model was selected as the ranker for project proposal ranking based on pairwise comparisons?", + "choices": [ + "GPT-4o", + "Claude-3.5-Sonnet", + "Claude-3-Opus", + "Gemini 1.5" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which model was selected as the ranker for project proposal ranking based on pairwise comparisons?\nAnswer:", + " GPT-4o" + ], + [ + "Question:Which model was selected as the ranker for project proposal ranking based on pairwise comparisons?\nAnswer:", + " Claude-3.5-Sonnet" + ], + [ + "Question:Which model was selected as the ranker for project proposal ranking based on pairwise comparisons?\nAnswer:", + " Claude-3-Opus" + ], + [ + "Question:Which model was selected as the ranker for project proposal ranking based on pairwise comparisons?\nAnswer:", + " Gemini 1.5" + ] + ], + "resps": [ + [ + [ + -9.801225662231445, + false + ] + ], + [ + [ + -18.31494903564453, + false + ] + ], + [ + [ + -13.818419456481934, + false + ] + ], + [ + [ + -11.413298606872559, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.801225662231445, + false + ], + [ + -18.31494903564453, + false + ], + [ + -13.818419456481934, + false + ], + [ + -11.413298606872559, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fc2ed4314dcee0565ca79572989aca6514627cfccf638ed7f981ebdb0c6b2078", + "prompt_hash": "ec596453bf625ac5a51f528394afe1d5e65a5cb337ed866606699df4b7014ee6", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 57, + "doc": { + "question": "How many seed ideas does the ideation agent generate per research topic?", + "choices": [ + "500", + "1000", + "2000", + "4000" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:How many seed ideas does the ideation agent generate per research topic?\nAnswer:", + " 500" + ], + [ + "Question:How many seed ideas does the ideation agent generate per research topic?\nAnswer:", + " 1000" + ], + [ + "Question:How many seed ideas does the ideation agent generate per research topic?\nAnswer:", + " 2000" + ], + [ + "Question:How many seed ideas does the ideation agent generate per research topic?\nAnswer:", + " 4000" + ] + ], + "resps": [ + [ + [ + -8.664156913757324, + false + ] + ], + [ + [ + -9.027714729309082, + false + ] + ], + [ + [ + -10.901153564453125, + false + ] + ], + [ + [ + -11.820749282836914, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.664156913757324, + false + ], + [ + -9.027714729309082, + false + ], + [ + -10.901153564453125, + false + ], + [ + -11.820749282836914, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9a34ce2f253785dd2e4ea6e29e06a175a7e030c54c846d40ebb91821f527e3fc", + "prompt_hash": "e932a0054f93b6ab828dd21259f0574bb1b42dabcbfde0b81e8a9d003b5d410b", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 58, + "doc": { + "question": "Which Sentence-Transformers model is used for deduplicating seed ideas?", + "choices": [ + "all-MiniLM-L6-v2", + "paraphrase-MiniLM-L12-v2", + "mpnet-base-v2", + "bert-base-nli-mean-tokens" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which Sentence-Transformers model is used for deduplicating seed ideas?\nAnswer:", + " all-MiniLM-L6-v2" + ], + [ + "Question:Which Sentence-Transformers model is used for deduplicating seed ideas?\nAnswer:", + " paraphrase-MiniLM-L12-v2" + ], + [ + "Question:Which Sentence-Transformers model is used for deduplicating seed ideas?\nAnswer:", + " mpnet-base-v2" + ], + [ + "Question:Which Sentence-Transformers model is used for deduplicating seed ideas?\nAnswer:", + " bert-base-nli-mean-tokens" + ] + ], + "resps": [ + [ + [ + -11.210933685302734, + false + ] + ], + [ + [ + -23.711734771728516, + false + ] + ], + [ + [ + -13.684947967529297, + false + ] + ], + [ + [ + -13.334312438964844, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.210933685302734, + false + ], + [ + -23.711734771728516, + false + ], + [ + -13.684947967529297, + false + ], + [ + -13.334312438964844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4485c7632925269c4b382a986bad0930a216df9f3a9b00c78f27ea87e5f0ed10", + "prompt_hash": "90afa19aaaee164fc1373ae78040ef1852c7c499388d136a5263d41f281cf1da", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 59, + "doc": { + "question": "What threshold is used for cosine similarity in the seed idea deduplication process?", + "choices": [ + "0.5", + "0.7", + "0.8", + "0.9" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What threshold is used for cosine similarity in the seed idea deduplication process?\nAnswer:", + " 0.5" + ], + [ + "Question:What threshold is used for cosine similarity in the seed idea deduplication process?\nAnswer:", + " 0.7" + ], + [ + "Question:What threshold is used for cosine similarity in the seed idea deduplication process?\nAnswer:", + " 0.8" + ], + [ + "Question:What threshold is used for cosine similarity in the seed idea deduplication process?\nAnswer:", + " 0.9" + ] + ], + "resps": [ + [ + [ + -4.469439506530762, + false + ] + ], + [ + [ + -4.292275428771973, + false + ] + ], + [ + [ + -4.272160530090332, + false + ] + ], + [ + [ + -3.832259178161621, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.469439506530762, + false + ], + [ + -4.292275428771973, + false + ], + [ + -4.272160530090332, + false + ], + [ + -3.832259178161621, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9c46621f2977b227ecc3f554d71640c222a63328eb6c7b6f54cffa5e90a36cfb", + "prompt_hash": "66edf2fc1c0a71ba26ed0fc6a7dce2a2e8f92579c61bb033bff3b3ea6d7b435c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 60, + "doc": { + "question": "What fundamental question lies at the core of large language models?", + "choices": [ + "How to encode images into text", + "How to simulate human reasoning", + "What word is likely to come next given a sequence of words", + "How to generate grammatically perfect sentences" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What fundamental question lies at the core of large language models?\nAnswer:", + " How to encode images into text" + ], + [ + "Question:What fundamental question lies at the core of large language models?\nAnswer:", + " How to simulate human reasoning" + ], + [ + "Question:What fundamental question lies at the core of large language models?\nAnswer:", + " What word is likely to come next given a sequence of words" + ], + [ + "Question:What fundamental question lies at the core of large language models?\nAnswer:", + " How to generate grammatically perfect sentences" + ] + ], + "resps": [ + [ + [ + -22.318111419677734, + false + ] + ], + [ + [ + -18.891098022460938, + false + ] + ], + [ + [ + -35.946693420410156, + false + ] + ], + [ + [ + -24.28699493408203, + false + ] + ] + ], + "filtered_resps": [ + [ + -22.318111419677734, + false + ], + [ + -18.891098022460938, + false + ], + [ + -35.946693420410156, + false + ], + [ + -24.28699493408203, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2bc651e9602b4caca953fde43181ea664f77a71d0f2d9c59a2be5be71d706f8b", + "prompt_hash": "6036288fedf64bb453b55ddbeaabaeabc0c5d4993693333928f10a56655d8f20", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 61, + "doc": { + "question": "What problem does the Mamba model aim to solve in sequence modeling?", + "choices": [ + "High memory usage of RNNs", + "Low accuracy on short sequences", + "Quadratic bottleneck in the attention mechanism", + "Lack of generalization in convolutional models" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What problem does the Mamba model aim to solve in sequence modeling?\nAnswer:", + " High memory usage of RNNs" + ], + [ + "Question:What problem does the Mamba model aim to solve in sequence modeling?\nAnswer:", + " Low accuracy on short sequences" + ], + [ + "Question:What problem does the Mamba model aim to solve in sequence modeling?\nAnswer:", + " Quadratic bottleneck in the attention mechanism" + ], + [ + "Question:What problem does the Mamba model aim to solve in sequence modeling?\nAnswer:", + " Lack of generalization in convolutional models" + ] + ], + "resps": [ + [ + [ + -26.314756393432617, + false + ] + ], + [ + [ + -24.300613403320312, + false + ] + ], + [ + [ + -31.88043975830078, + false + ] + ], + [ + [ + -26.176528930664062, + false + ] + ] + ], + "filtered_resps": [ + [ + -26.314756393432617, + false + ], + [ + -24.300613403320312, + false + ], + [ + -31.88043975830078, + false + ], + [ + -26.176528930664062, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3938cd1105c8c6addb1747a25a4e9533a4491d59d54415b0095e03dfac182ec6", + "prompt_hash": "a123996b1c5197ad6764012ae08b4e0a72db1f108396c12b875028b51eafd764", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 62, + "doc": { + "question": "What does the BitFit fine-tuning method update?", + "choices": [ + "All transformer layers", + "Only the output layer weights", + "Bias terms and the task-specific classification layer", + "The entire embedding matrix" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does the BitFit fine-tuning method update?\nAnswer:", + " All transformer layers" + ], + [ + "Question:What does the BitFit fine-tuning method update?\nAnswer:", + " Only the output layer weights" + ], + [ + "Question:What does the BitFit fine-tuning method update?\nAnswer:", + " Bias terms and the task-specific classification layer" + ], + [ + "Question:What does the BitFit fine-tuning method update?\nAnswer:", + " The entire embedding matrix" + ] + ], + "resps": [ + [ + [ + -15.890267372131348, + false + ] + ], + [ + [ + -15.56641960144043, + false + ] + ], + [ + [ + -37.99882507324219, + false + ] + ], + [ + [ + -15.27823543548584, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.890267372131348, + false + ], + [ + -15.56641960144043, + false + ], + [ + -37.99882507324219, + false + ], + [ + -15.27823543548584, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f2207e170fc43164dbf1d42e40561c3f4250ca00745940935f99d5dc20438dc0", + "prompt_hash": "01c9f1436cd9999ff1800b6eebc4567206d2b4ecdf5da6dad52c30ed2d2f1319", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 63, + "doc": { + "question": "What is a characteristic of Kolmogorov-Arnold Networks (KANs) compared to traditional neural networks?", + "choices": [ + "They use fixed activation functions on nodes", + "They apply learnable activation functions on edges", + "They avoid using activation functions entirely", + "They use convolutional filters for feature learning" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is a characteristic of Kolmogorov-Arnold Networks (KANs) compared to traditional neural networks?\nAnswer:", + " They use fixed activation functions on nodes" + ], + [ + "Question:What is a characteristic of Kolmogorov-Arnold Networks (KANs) compared to traditional neural networks?\nAnswer:", + " They apply learnable activation functions on edges" + ], + [ + "Question:What is a characteristic of Kolmogorov-Arnold Networks (KANs) compared to traditional neural networks?\nAnswer:", + " They avoid using activation functions entirely" + ], + [ + "Question:What is a characteristic of Kolmogorov-Arnold Networks (KANs) compared to traditional neural networks?\nAnswer:", + " They use convolutional filters for feature learning" + ] + ], + "resps": [ + [ + [ + -27.779762268066406, + false + ] + ], + [ + [ + -32.78925323486328, + false + ] + ], + [ + [ + -26.899110794067383, + false + ] + ], + [ + [ + -23.58316421508789, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.779762268066406, + false + ], + [ + -32.78925323486328, + false + ], + [ + -26.899110794067383, + false + ], + [ + -23.58316421508789, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7a4b192cfda32cda5e83b7be84297fb65f0922104d79f4ed508b830e34f12957", + "prompt_hash": "5c76c2773ed113ea2061bda0694c74214eb2e6f96b091a4fb1d3f9547e65de70", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 64, + "doc": { + "question": "What does Chain-of-Thought (CoT) prompting aim to improve in LLMs?", + "choices": [ + "Memory efficiency during training", + "Logical consistency and reduction of hallucinations", + "Vocabulary expansion", + "Encoding of positional information" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What does Chain-of-Thought (CoT) prompting aim to improve in LLMs?\nAnswer:", + " Memory efficiency during training" + ], + [ + "Question:What does Chain-of-Thought (CoT) prompting aim to improve in LLMs?\nAnswer:", + " Logical consistency and reduction of hallucinations" + ], + [ + "Question:What does Chain-of-Thought (CoT) prompting aim to improve in LLMs?\nAnswer:", + " Vocabulary expansion" + ], + [ + "Question:What does Chain-of-Thought (CoT) prompting aim to improve in LLMs?\nAnswer:", + " Encoding of positional information" + ] + ], + "resps": [ + [ + [ + -23.194095611572266, + false + ] + ], + [ + [ + -23.975643157958984, + false + ] + ], + [ + [ + -16.475173950195312, + false + ] + ], + [ + [ + -24.811058044433594, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.194095611572266, + false + ], + [ + -23.975643157958984, + false + ], + [ + -16.475173950195312, + false + ], + [ + -24.811058044433594, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "780cbf0fabbb53d71f45195aadaf088049f8c1fc307a088210da70f2f04fec4c", + "prompt_hash": "ab10be7a2b6fdc7339decae8a574a1da022b8713ad30ecf09512ac21ac410ccd", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 65, + "doc": { + "question": "What type of license is the gsplat library released under?", + "choices": [ + "MIT License", + "Apache License 2.0", + "GPLv3", + "BSD 3-Clause" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What type of license is the gsplat library released under?\nAnswer:", + " MIT License" + ], + [ + "Question:What type of license is the gsplat library released under?\nAnswer:", + " Apache License 2.0" + ], + [ + "Question:What type of license is the gsplat library released under?\nAnswer:", + " GPLv3" + ], + [ + "Question:What type of license is the gsplat library released under?\nAnswer:", + " BSD 3-Clause" + ] + ], + "resps": [ + [ + [ + -3.371497392654419, + false + ] + ], + [ + [ + -4.135261058807373, + false + ] + ], + [ + [ + -5.155677795410156, + false + ] + ], + [ + [ + -7.361147880554199, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.371497392654419, + false + ], + [ + -4.135261058807373, + false + ], + [ + -5.155677795410156, + false + ], + [ + -7.361147880554199, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d889ba1674f868cc5cfc85b5b1827d87aabbd6426f6e15ba997cd07ac9eb3859", + "prompt_hash": "caf3c756d70db75fe1205e6839156b37c2667f3ddb4a08ac10da2bab6dfef009", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 66, + "doc": { + "question": "Which densification strategy in gsplat is based on a Bayesian perspective using SGLD?", + "choices": [ + "ADC", + "Absgrad", + "MCMC", + "SHG" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which densification strategy in gsplat is based on a Bayesian perspective using SGLD?\nAnswer:", + " ADC" + ], + [ + "Question:Which densification strategy in gsplat is based on a Bayesian perspective using SGLD?\nAnswer:", + " Absgrad" + ], + [ + "Question:Which densification strategy in gsplat is based on a Bayesian perspective using SGLD?\nAnswer:", + " MCMC" + ], + [ + "Question:Which densification strategy in gsplat is based on a Bayesian perspective using SGLD?\nAnswer:", + " SHG" + ] + ], + "resps": [ + [ + [ + -15.597640991210938, + false + ] + ], + [ + [ + -21.404687881469727, + false + ] + ], + [ + [ + -8.992700576782227, + false + ] + ], + [ + [ + -13.896909713745117, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.597640991210938, + false + ], + [ + -21.404687881469727, + false + ], + [ + -8.992700576782227, + false + ], + [ + -13.896909713745117, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9d49ca94b540e249c2679eaa12be0fa663ff1ee9035b86504f1a8d460014df12", + "prompt_hash": "a30c1008a89164e906dd377775ee686a1b43a3d9ce66da3b1d847ab0aa19aae0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 67, + "doc": { + "question": "Which CUDA-accelerated library supports both RGB and depth map rendering using differentiable rasterization?", + "choices": [ + "torch-splatting", + "gsplat", + "GauStudio", + "instant-ngp" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which CUDA-accelerated library supports both RGB and depth map rendering using differentiable rasterization?\nAnswer:", + " torch-splatting" + ], + [ + "Question:Which CUDA-accelerated library supports both RGB and depth map rendering using differentiable rasterization?\nAnswer:", + " gsplat" + ], + [ + "Question:Which CUDA-accelerated library supports both RGB and depth map rendering using differentiable rasterization?\nAnswer:", + " GauStudio" + ], + [ + "Question:Which CUDA-accelerated library supports both RGB and depth map rendering using differentiable rasterization?\nAnswer:", + " instant-ngp" + ] + ], + "resps": [ + [ + [ + -20.24197006225586, + false + ] + ], + [ + [ + -21.16048812866211, + false + ] + ], + [ + [ + -22.701190948486328, + false + ] + ], + [ + [ + -14.450319290161133, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.24197006225586, + false + ], + [ + -21.16048812866211, + false + ], + [ + -22.701190948486328, + false + ], + [ + -14.450319290161133, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3a5eb42470b10857233e816971280246b52c1ab8353efd055aab5c2f4ef7c48b", + "prompt_hash": "80efea59a81b004277094513555a8d3fea5085fa20b2bd149bf0a251c372755e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 68, + "doc": { + "question": "What is the default threshold for culling Gaussians based on opacity in the ADC strategy?", + "choices": [ + "0.01", + "0.001", + "0.005", + "0.02" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the default threshold for culling Gaussians based on opacity in the ADC strategy?\nAnswer:", + " 0.01" + ], + [ + "Question:What is the default threshold for culling Gaussians based on opacity in the ADC strategy?\nAnswer:", + " 0.001" + ], + [ + "Question:What is the default threshold for culling Gaussians based on opacity in the ADC strategy?\nAnswer:", + " 0.005" + ], + [ + "Question:What is the default threshold for culling Gaussians based on opacity in the ADC strategy?\nAnswer:", + " 0.02" + ] + ], + "resps": [ + [ + [ + -4.799712181091309, + false + ] + ], + [ + [ + -5.712863922119141, + false + ] + ], + [ + [ + -6.453279495239258, + false + ] + ], + [ + [ + -5.598419666290283, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.799712181091309, + false + ], + [ + -5.712863922119141, + false + ], + [ + -6.453279495239258, + false + ], + [ + -5.598419666290283, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2753c252c20ee778d0651fa89f051f9ce493d0a7f6758d12c43c280951658466", + "prompt_hash": "5f66e60bcb2b0eb7eee2efda17ca71e50bdd8e0ce36d7877dbbabe8d2a5667dc", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 69, + "doc": { + "question": "What is the default value of the anti-aliasing parameter 's' in gsplat?", + "choices": [ + "0.1", + "0.2", + "0.3", + "0.5" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the default value of the anti-aliasing parameter 's' in gsplat?\nAnswer:", + " 0.1" + ], + [ + "Question:What is the default value of the anti-aliasing parameter 's' in gsplat?\nAnswer:", + " 0.2" + ], + [ + "Question:What is the default value of the anti-aliasing parameter 's' in gsplat?\nAnswer:", + " 0.3" + ], + [ + "Question:What is the default value of the anti-aliasing parameter 's' in gsplat?\nAnswer:", + " 0.5" + ] + ], + "resps": [ + [ + [ + -5.626408576965332, + false + ] + ], + [ + [ + -5.924612998962402, + false + ] + ], + [ + [ + -6.593517303466797, + false + ] + ], + [ + [ + -5.611193656921387, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.626408576965332, + false + ], + [ + -5.924612998962402, + false + ], + [ + -6.593517303466797, + false + ], + [ + -5.611193656921387, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d9ef7a2b2cb18d913274bf908caa43fe2c0e3d868458ce345eb5de473f7e9baf", + "prompt_hash": "01b84fcc2036f104fa4e0df9fa2ac7ca74feb1a427a13cd782bfb7759396c2f5", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 70, + "doc": { + "question": "What is the smallest positive integer \\( t \\) such that there exist integers \\( x_1, x_2, \\ldots, x_t \\) satisfying \\( x_1^3 + x_2^3 + \\cdots + x_t^3 = 2002^{2002} \\)?", + "choices": [ + "A. 3", + "B. 4", + "C. 5", + "D. 6" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the smallest positive integer \\( t \\) such that there exist integers \\( x_1, x_2, \\ldots, x_t \\) satisfying \\( x_1^3 + x_2^3 + \\cdots + x_t^3 = 2002^{2002} \\)?\nAnswer:", + " A. 3" + ], + [ + "Question:What is the smallest positive integer \\( t \\) such that there exist integers \\( x_1, x_2, \\ldots, x_t \\) satisfying \\( x_1^3 + x_2^3 + \\cdots + x_t^3 = 2002^{2002} \\)?\nAnswer:", + " B. 4" + ], + [ + "Question:What is the smallest positive integer \\( t \\) such that there exist integers \\( x_1, x_2, \\ldots, x_t \\) satisfying \\( x_1^3 + x_2^3 + \\cdots + x_t^3 = 2002^{2002} \\)?\nAnswer:", + " C. 5" + ], + [ + "Question:What is the smallest positive integer \\( t \\) such that there exist integers \\( x_1, x_2, \\ldots, x_t \\) satisfying \\( x_1^3 + x_2^3 + \\cdots + x_t^3 = 2002^{2002} \\)?\nAnswer:", + " D. 6" + ] + ], + "resps": [ + [ + [ + -12.540119171142578, + false + ] + ], + [ + [ + -17.282943725585938, + false + ] + ], + [ + [ + -18.027965545654297, + false + ] + ], + [ + [ + -17.628211975097656, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.540119171142578, + false + ], + [ + -17.282943725585938, + false + ], + [ + -18.027965545654297, + false + ], + [ + -17.628211975097656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ff1d2def68ec018b36ac8c6dfa4ce07df301cfbc19e9c315d5e20acf8f8bcd9b", + "prompt_hash": "58bd525ae7ffa87e25e53c0d1d63c3f6e27f16d75ff862d206a4b58fffed811f", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 71, + "doc": { + "question": "Which function satisfies the property that for each \\( x \\in \\mathbb{R}^+ \\), there exists exactly one \\( y \\in \\mathbb{R}^+ \\) such that \\( xf(y) + yf(x) \\leq 2 \\)?", + "choices": [ + "A.\\( f(x) = x^2 \\)", + "B. \\( f(x) = 1 \\)", + "C. \\( f(x) = \\frac{1}{x} \\)", + "D. \\( f(x) = x + 1 \\)" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which function satisfies the property that for each \\( x \\in \\mathbb{R}^+ \\), there exists exactly one \\( y \\in \\mathbb{R}^+ \\) such that \\( xf(y) + yf(x) \\leq 2 \\)?\nAnswer:", + " A.\\( f(x) = x^2 \\)" + ], + [ + "Question:Which function satisfies the property that for each \\( x \\in \\mathbb{R}^+ \\), there exists exactly one \\( y \\in \\mathbb{R}^+ \\) such that \\( xf(y) + yf(x) \\leq 2 \\)?\nAnswer:", + " B. \\( f(x) = 1 \\)" + ], + [ + "Question:Which function satisfies the property that for each \\( x \\in \\mathbb{R}^+ \\), there exists exactly one \\( y \\in \\mathbb{R}^+ \\) such that \\( xf(y) + yf(x) \\leq 2 \\)?\nAnswer:", + " C. \\( f(x) = \\frac{1}{x} \\)" + ], + [ + "Question:Which function satisfies the property that for each \\( x \\in \\mathbb{R}^+ \\), there exists exactly one \\( y \\in \\mathbb{R}^+ \\) such that \\( xf(y) + yf(x) \\leq 2 \\)?\nAnswer:", + " D. \\( f(x) = x + 1 \\)" + ] + ], + "resps": [ + [ + [ + -19.827207565307617, + false + ] + ], + [ + [ + -20.322336196899414, + false + ] + ], + [ + [ + -21.418210983276367, + false + ] + ], + [ + [ + -22.060394287109375, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.827207565307617, + false + ], + [ + -20.322336196899414, + false + ], + [ + -21.418210983276367, + false + ], + [ + -22.060394287109375, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "aef2aa8bcc1161172ca855d45b2b7026cbb5b6215ec2b4835db010afe650bff4", + "prompt_hash": "5ba85c492e639300e8f33662c7377e2dd6cfa7c4c8201da4c4e2fb94ab263f27", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 72, + "doc": { + "question": "What is the answer to whether it is possible that \\( s(k) \\) and \\( s(P(k)) \\) have the same parity for all positive integers \\( k \\), where \\( s(n) \\) denotes the sum of digits of \\( n \\)?", + "choices": [ + "A. Yes", + "B. No", + "C. Sometimes", + "D. Cannot be determined" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the answer to whether it is possible that \\( s(k) \\) and \\( s(P(k)) \\) have the same parity for all positive integers \\( k \\), where \\( s(n) \\) denotes the sum of digits of \\( n \\)?\nAnswer:", + " A. Yes" + ], + [ + "Question:What is the answer to whether it is possible that \\( s(k) \\) and \\( s(P(k)) \\) have the same parity for all positive integers \\( k \\), where \\( s(n) \\) denotes the sum of digits of \\( n \\)?\nAnswer:", + " B. No" + ], + [ + "Question:What is the answer to whether it is possible that \\( s(k) \\) and \\( s(P(k)) \\) have the same parity for all positive integers \\( k \\), where \\( s(n) \\) denotes the sum of digits of \\( n \\)?\nAnswer:", + " C. Sometimes" + ], + [ + "Question:What is the answer to whether it is possible that \\( s(k) \\) and \\( s(P(k)) \\) have the same parity for all positive integers \\( k \\), where \\( s(n) \\) denotes the sum of digits of \\( n \\)?\nAnswer:", + " D. Cannot be determined" + ] + ], + "resps": [ + [ + [ + -10.544660568237305, + false + ] + ], + [ + [ + -14.121866226196289, + false + ] + ], + [ + [ + -20.223602294921875, + false + ] + ], + [ + [ + -25.502727508544922, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.544660568237305, + false + ], + [ + -14.121866226196289, + false + ], + [ + -20.223602294921875, + false + ], + [ + -25.502727508544922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e52014fa502465af56621ee5459f41c8d9ef996e16731caf76c60c39f71a0572", + "prompt_hash": "b751d112055d2f216ddab13abfb425a9e4b66761413232e32738a2ffc59a0d5a", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 73, + "doc": { + "question": "What is the maximum number of students \\( k \\) such that any two have at least one course in common, and no outsider can be classmates with all \\( k \\)?", + "choices": [ + "A. 256", + "B. 512", + "C. 1023", + "D. 1024" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the maximum number of students \\( k \\) such that any two have at least one course in common, and no outsider can be classmates with all \\( k \\)?\nAnswer:", + " A. 256" + ], + [ + "Question:What is the maximum number of students \\( k \\) such that any two have at least one course in common, and no outsider can be classmates with all \\( k \\)?\nAnswer:", + " B. 512" + ], + [ + "Question:What is the maximum number of students \\( k \\) such that any two have at least one course in common, and no outsider can be classmates with all \\( k \\)?\nAnswer:", + " C. 1023" + ], + [ + "Question:What is the maximum number of students \\( k \\) such that any two have at least one course in common, and no outsider can be classmates with all \\( k \\)?\nAnswer:", + " D. 1024" + ] + ], + "resps": [ + [ + [ + -19.086959838867188, + false + ] + ], + [ + [ + -24.054302215576172, + false + ] + ], + [ + [ + -28.452510833740234, + false + ] + ], + [ + [ + -27.108470916748047, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.086959838867188, + false + ], + [ + -24.054302215576172, + false + ], + [ + -28.452510833740234, + false + ], + [ + -27.108470916748047, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "69a9652202d73079db7a9bd7ed603c9584a9846c30e7d73e2aa2a0cd8723d7b9", + "prompt_hash": "eedfc36fc0e2bfd5bc91cf0e269066cc1dec19da4fa2ffd9332c0dbbc7320d8d", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 74, + "doc": { + "question": "Who wins the game where Joe and Penny remove stones with increasing limits per move, starting from 5000 stones and alternating turns?", + "choices": [ + "A. Joe", + "B. Penny", + "C. It's a draw", + "D. Cannot be determined" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Who wins the game where Joe and Penny remove stones with increasing limits per move, starting from 5000 stones and alternating turns?\nAnswer:", + " A. Joe" + ], + [ + "Question:Who wins the game where Joe and Penny remove stones with increasing limits per move, starting from 5000 stones and alternating turns?\nAnswer:", + " B. Penny" + ], + [ + "Question:Who wins the game where Joe and Penny remove stones with increasing limits per move, starting from 5000 stones and alternating turns?\nAnswer:", + " C. It's a draw" + ], + [ + "Question:Who wins the game where Joe and Penny remove stones with increasing limits per move, starting from 5000 stones and alternating turns?\nAnswer:", + " D. Cannot be determined" + ] + ], + "resps": [ + [ + [ + -8.734781265258789, + false + ] + ], + [ + [ + -11.859012603759766, + false + ] + ], + [ + [ + -23.23371124267578, + false + ] + ], + [ + [ + -28.801799774169922, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.734781265258789, + false + ], + [ + -11.859012603759766, + false + ], + [ + -23.23371124267578, + false + ], + [ + -28.801799774169922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "384f7ec867d80108c16e91e7b85808724733266fff2fdce9a3462149734ba984", + "prompt_hash": "05de3ecb83cbdc67d7a9928100b0a19366b0596238a98cb064bca07239cbc3d9", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 75, + "doc": { + "question": "What does LoRA update during fine-tuning?", + "choices": [ + "Only the original model weights", + "Only the bias terms", + "Only the low-rank matrices A and B", + "All parameters of the model" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does LoRA update during fine-tuning?\nAnswer:", + " Only the original model weights" + ], + [ + "Question:What does LoRA update during fine-tuning?\nAnswer:", + " Only the bias terms" + ], + [ + "Question:What does LoRA update during fine-tuning?\nAnswer:", + " Only the low-rank matrices A and B" + ], + [ + "Question:What does LoRA update during fine-tuning?\nAnswer:", + " All parameters of the model" + ] + ], + "resps": [ + [ + [ + -16.489402770996094, + false + ] + ], + [ + [ + -14.673160552978516, + false + ] + ], + [ + [ + -19.73063087463379, + false + ] + ], + [ + [ + -12.429987907409668, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.489402770996094, + false + ], + [ + -14.673160552978516, + false + ], + [ + -19.73063087463379, + false + ], + [ + -12.429987907409668, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c9ddacc3de73b039949f28da7796fef39b37af97348266e2cddc327310dc2b12", + "prompt_hash": "45f5e6babc5b65261709ef37a6a4ba781194093d2b182e9b94babfe251f5005b", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 76, + "doc": { + "question": "What is the primary limitation of FedIT when clients have different LoRA ranks?", + "choices": [ + "It cannot communicate with more than 10 clients", + "It introduces delays due to encryption", + "It cannot aggregate local LoRAs with heterogeneous ranks", + "It lacks support for any LoRA rank" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the primary limitation of FedIT when clients have different LoRA ranks?\nAnswer:", + " It cannot communicate with more than 10 clients" + ], + [ + "Question:What is the primary limitation of FedIT when clients have different LoRA ranks?\nAnswer:", + " It introduces delays due to encryption" + ], + [ + "Question:What is the primary limitation of FedIT when clients have different LoRA ranks?\nAnswer:", + " It cannot aggregate local LoRAs with heterogeneous ranks" + ], + [ + "Question:What is the primary limitation of FedIT when clients have different LoRA ranks?\nAnswer:", + " It lacks support for any LoRA rank" + ] + ], + "resps": [ + [ + [ + -28.548847198486328, + false + ] + ], + [ + [ + -30.07623863220215, + false + ] + ], + [ + [ + -34.82258605957031, + false + ] + ], + [ + [ + -23.75527572631836, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.548847198486328, + false + ], + [ + -30.07623863220215, + false + ], + [ + -34.82258605957031, + false + ], + [ + -23.75527572631836, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5f313bde67b9e18bca0d640d237389a6a41cdcc68166bbc185ac0dcc9d6d1c74", + "prompt_hash": "54439a6ab5141dff92cc07e799f1f6ac1f9d6a0a970c1aa8d9ed283468651382", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 77, + "doc": { + "question": "Which mathematical inequality illustrates the aggregation noise in FedIT?", + "choices": [ + "ΣA × ΣB = ΣA×B", + "ΣA + ΣB = Σ(A+B)", + "ΣA × ΣB ≠ ΣA×B", + "A × B = B × A" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which mathematical inequality illustrates the aggregation noise in FedIT?\nAnswer:", + " ΣA × ΣB = ΣA×B" + ], + [ + "Question:Which mathematical inequality illustrates the aggregation noise in FedIT?\nAnswer:", + " ΣA + ΣB = Σ(A+B)" + ], + [ + "Question:Which mathematical inequality illustrates the aggregation noise in FedIT?\nAnswer:", + " ΣA × ΣB ≠ ΣA×B" + ], + [ + "Question:Which mathematical inequality illustrates the aggregation noise in FedIT?\nAnswer:", + " A × B = B × A" + ] + ], + "resps": [ + [ + [ + -46.70803451538086, + false + ] + ], + [ + [ + -35.111167907714844, + false + ] + ], + [ + [ + -52.45105743408203, + false + ] + ], + [ + [ + -23.067930221557617, + false + ] + ] + ], + "filtered_resps": [ + [ + -46.70803451538086, + false + ], + [ + -35.111167907714844, + false + ], + [ + -52.45105743408203, + false + ], + [ + -23.067930221557617, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c38660c2b5fdfbe620a25e1771ca70b1c0e8f7a6d749d26a1a8eb72abcd55f5f", + "prompt_hash": "b23c2a87198b56c5a1b9b032fb8f76f58cc1b32740c848d8b3d6c77e185a9d49", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 78, + "doc": { + "question": "What is the dimension of LoRA matrix A when the rank is set to 16 for a LLaMA-7B model?", + "choices": [ + "4096 × 4096", + "4096 × 16", + "16 × 4096", + "16 × 16" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the dimension of LoRA matrix A when the rank is set to 16 for a LLaMA-7B model?\nAnswer:", + " 4096 × 4096" + ], + [ + "Question:What is the dimension of LoRA matrix A when the rank is set to 16 for a LLaMA-7B model?\nAnswer:", + " 4096 × 16" + ], + [ + "Question:What is the dimension of LoRA matrix A when the rank is set to 16 for a LLaMA-7B model?\nAnswer:", + " 16 × 4096" + ], + [ + "Question:What is the dimension of LoRA matrix A when the rank is set to 16 for a LLaMA-7B model?\nAnswer:", + " 16 × 16" + ] + ], + "resps": [ + [ + [ + -14.631674766540527, + false + ] + ], + [ + [ + -16.16933250427246, + false + ] + ], + [ + [ + -16.902545928955078, + false + ] + ], + [ + [ + -10.089134216308594, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.631674766540527, + false + ], + [ + -16.16933250427246, + false + ], + [ + -16.902545928955078, + false + ], + [ + -10.089134216308594, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7b37553842b931dc0bc0dd2ce757625784693d0ca565acd95c7612158325c33f", + "prompt_hash": "fa9872f95ee586fc82ac541c3731438cf2389ee3d88d1488bd93cd963667165d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 79, + "doc": { + "question": "What operation is used in FLoRA to aggregate LoRA modules from multiple clients?", + "choices": [ + "Matrix inversion", + "Module duplication", + "Stacking", + "Element-wise multiplication" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What operation is used in FLoRA to aggregate LoRA modules from multiple clients?\nAnswer:", + " Matrix inversion" + ], + [ + "Question:What operation is used in FLoRA to aggregate LoRA modules from multiple clients?\nAnswer:", + " Module duplication" + ], + [ + "Question:What operation is used in FLoRA to aggregate LoRA modules from multiple clients?\nAnswer:", + " Stacking" + ], + [ + "Question:What operation is used in FLoRA to aggregate LoRA modules from multiple clients?\nAnswer:", + " Element-wise multiplication" + ] + ], + "resps": [ + [ + [ + -13.950897216796875, + false + ] + ], + [ + [ + -15.547689437866211, + false + ] + ], + [ + [ + -9.502487182617188, + false + ] + ], + [ + [ + -8.775291442871094, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.950897216796875, + false + ], + [ + -15.547689437866211, + false + ], + [ + -9.502487182617188, + false + ], + [ + -8.775291442871094, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "18616f7d416b734bb496239b878291e49111f7c0ab679eeddadaf7a1fc1f9fdc", + "prompt_hash": "cd7bd01f285b3dea733e29ccaddf2c1e488215351be7baf12819bc3c33d3a944", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 80, + "doc": { + "question": "What is the highest accuracy achieved by LLaMA 3.1 405B on the regular Blocksworld dataset?", + "choices": [ + "62.6%", + "52.8%", + "82.0%", + "97.8%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the highest accuracy achieved by LLaMA 3.1 405B on the regular Blocksworld dataset?\nAnswer:", + " 62.6%" + ], + [ + "Question:What is the highest accuracy achieved by LLaMA 3.1 405B on the regular Blocksworld dataset?\nAnswer:", + " 52.8%" + ], + [ + "Question:What is the highest accuracy achieved by LLaMA 3.1 405B on the regular Blocksworld dataset?\nAnswer:", + " 82.0%" + ], + [ + "Question:What is the highest accuracy achieved by LLaMA 3.1 405B on the regular Blocksworld dataset?\nAnswer:", + " 97.8%" + ] + ], + "resps": [ + [ + [ + -14.532271385192871, + false + ] + ], + [ + [ + -14.77785587310791, + false + ] + ], + [ + [ + -13.464859962463379, + false + ] + ], + [ + [ + -13.568106651306152, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.532271385192871, + false + ], + [ + -14.77785587310791, + false + ], + [ + -13.464859962463379, + false + ], + [ + -13.568106651306152, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4d3a6a1e8cfc25ac671d4c12e53eb1831e93fe6c7b2b3858fb109b324f5acd4a", + "prompt_hash": "5cf63224372e80012ab0d364dde0ac06edc4a61d4bf4f677d8f899c76df107de", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 81, + "doc": { + "question": "How many steps do the Blocksworld problems in the PlanBench dataset require to solve?", + "choices": [ + "2 to 16 steps", + "10 to 30 steps", + "3 to 15 steps", + "5 to 20 steps" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:How many steps do the Blocksworld problems in the PlanBench dataset require to solve?\nAnswer:", + " 2 to 16 steps" + ], + [ + "Question:How many steps do the Blocksworld problems in the PlanBench dataset require to solve?\nAnswer:", + " 10 to 30 steps" + ], + [ + "Question:How many steps do the Blocksworld problems in the PlanBench dataset require to solve?\nAnswer:", + " 3 to 15 steps" + ], + [ + "Question:How many steps do the Blocksworld problems in the PlanBench dataset require to solve?\nAnswer:", + " 5 to 20 steps" + ] + ], + "resps": [ + [ + [ + -16.07807159423828, + false + ] + ], + [ + [ + -13.862556457519531, + false + ] + ], + [ + [ + -14.577719688415527, + false + ] + ], + [ + [ + -13.827674865722656, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.07807159423828, + false + ], + [ + -13.862556457519531, + false + ], + [ + -14.577719688415527, + false + ], + [ + -13.827674865722656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3759c5f8da919bb1d2ea2b0722011fe9d55074e9f7991c5078ee4b9970fc6eae", + "prompt_hash": "63490735f777659af0d2ebe263086aa21b8e16fce98f28321ed0d880b0693ff4", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 82, + "doc": { + "question": "How much did the researchers report spending on o1 model experiments?", + "choices": [ + "$1897.55", + "$975.50", + "$2783.40", + "$1325.00" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:How much did the researchers report spending on o1 model experiments?\nAnswer:", + " $1897.55" + ], + [ + "Question:How much did the researchers report spending on o1 model experiments?\nAnswer:", + " $975.50" + ], + [ + "Question:How much did the researchers report spending on o1 model experiments?\nAnswer:", + " $2783.40" + ], + [ + "Question:How much did the researchers report spending on o1 model experiments?\nAnswer:", + " $1325.00" + ] + ], + "resps": [ + [ + [ + -19.34306526184082, + false + ] + ], + [ + [ + -17.32545280456543, + false + ] + ], + [ + [ + -19.518150329589844, + false + ] + ], + [ + [ + -15.12232494354248, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.34306526184082, + false + ], + [ + -17.32545280456543, + false + ], + [ + -19.518150329589844, + false + ], + [ + -15.12232494354248, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3f1f4e50c43d044f5631966b4bcaa790ad7a0c82e22086d1f50d5ecd2cb50a41", + "prompt_hash": "ce84816c19ef6f5dbd25e89ad489bf610fb24cd3b759bdebb4f1ab53474df311", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 83, + "doc": { + "question": "What percentage of randomized Mystery Blocksworld unsolvable instances were falsely claimed as impossible by the model?", + "choices": [ + "11.5%", + "5%", + "16%", + "27%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What percentage of randomized Mystery Blocksworld unsolvable instances were falsely claimed as impossible by the model?\nAnswer:", + " 11.5%" + ], + [ + "Question:What percentage of randomized Mystery Blocksworld unsolvable instances were falsely claimed as impossible by the model?\nAnswer:", + " 5%" + ], + [ + "Question:What percentage of randomized Mystery Blocksworld unsolvable instances were falsely claimed as impossible by the model?\nAnswer:", + " 16%" + ], + [ + "Question:What percentage of randomized Mystery Blocksworld unsolvable instances were falsely claimed as impossible by the model?\nAnswer:", + " 27%" + ] + ], + "resps": [ + [ + [ + -12.747138023376465, + false + ] + ], + [ + [ + -9.35422134399414, + false + ] + ], + [ + [ + -10.66763973236084, + false + ] + ], + [ + [ + -11.035689353942871, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.747138023376465, + false + ], + [ + -9.35422134399414, + false + ], + [ + -10.66763973236084, + false + ], + [ + -11.035689353942871, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "dfb7671ceb3b0e0d7d62d30d1f081a8b650635dc785392ee16b600aab062f7ec", + "prompt_hash": "21053b00326cf3ea0175f96b68b215af89dbc864bf99d0da0d1654df70ab6699", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 84, + "doc": { + "question": "How long did Fast Downward take on average to solve a Blocksworld instance?", + "choices": [ + "0.265 seconds", + "2.65 seconds", + "1.25 seconds", + "0.562 seconds" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:How long did Fast Downward take on average to solve a Blocksworld instance?\nAnswer:", + " 0.265 seconds" + ], + [ + "Question:How long did Fast Downward take on average to solve a Blocksworld instance?\nAnswer:", + " 2.65 seconds" + ], + [ + "Question:How long did Fast Downward take on average to solve a Blocksworld instance?\nAnswer:", + " 1.25 seconds" + ], + [ + "Question:How long did Fast Downward take on average to solve a Blocksworld instance?\nAnswer:", + " 0.562 seconds" + ] + ], + "resps": [ + [ + [ + -12.132086753845215, + false + ] + ], + [ + [ + -11.014654159545898, + false + ] + ], + [ + [ + -9.894144058227539, + false + ] + ], + [ + [ + -12.297073364257812, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.132086753845215, + false + ], + [ + -11.014654159545898, + false + ], + [ + -9.894144058227539, + false + ], + [ + -12.297073364257812, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "019764cc7d216dc4c71ecb9c7b035d108cddded072ed5a1536a2314957381b76", + "prompt_hash": "6ed5a165c77e6410e7ce5b5e3b46407a205fa6a4ccbe9298c7ab249929ed554e", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 85, + "doc": { + "question": "What is the main reason OCR-based text extraction is unsuitable for high-quality document data extraction?", + "choices": [ + "It is too slow for large documents.", + "It requires labeled training data.", + "It introduces noise when processing images, tables, and formulas.", + "It only supports English text." + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the main reason OCR-based text extraction is unsuitable for high-quality document data extraction?\nAnswer:", + " It is too slow for large documents." + ], + [ + "Question:What is the main reason OCR-based text extraction is unsuitable for high-quality document data extraction?\nAnswer:", + " It requires labeled training data." + ], + [ + "Question:What is the main reason OCR-based text extraction is unsuitable for high-quality document data extraction?\nAnswer:", + " It introduces noise when processing images, tables, and formulas." + ], + [ + "Question:What is the main reason OCR-based text extraction is unsuitable for high-quality document data extraction?\nAnswer:", + " It only supports English text." + ] + ], + "resps": [ + [ + [ + -19.091449737548828, + false + ] + ], + [ + [ + -16.6263370513916, + false + ] + ], + [ + [ + -46.70049285888672, + false + ] + ], + [ + [ + -18.63629913330078, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.091449737548828, + false + ], + [ + -16.6263370513916, + false + ], + [ + -46.70049285888672, + false + ], + [ + -18.63629913330078, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "466404aaa1d54d80fa94d7b2cfe084ca64f9489abe8d2f9a933a9acb1b08236e", + "prompt_hash": "4bf51dce5b41b771f49adf61efe7e776d91851252c2d335140a9c3c53bcdb4f3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 86, + "doc": { + "question": "Which model does MinerU use for formula recognition?", + "choices": [ + "Pix2tex", + "Texify", + "UniMERNet", + "Mathpix" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model does MinerU use for formula recognition?\nAnswer:", + " Pix2tex" + ], + [ + "Question:Which model does MinerU use for formula recognition?\nAnswer:", + " Texify" + ], + [ + "Question:Which model does MinerU use for formula recognition?\nAnswer:", + " UniMERNet" + ], + [ + "Question:Which model does MinerU use for formula recognition?\nAnswer:", + " Mathpix" + ] + ], + "resps": [ + [ + [ + -24.9783935546875, + false + ] + ], + [ + [ + -19.929180145263672, + false + ] + ], + [ + [ + -27.404014587402344, + false + ] + ], + [ + [ + -20.638700485229492, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.9783935546875, + false + ], + [ + -19.929180145263672, + false + ], + [ + -27.404014587402344, + false + ], + [ + -20.638700485229492, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5f798d7eefabe6752366a0de70c102ffe7012c3160888e78084fd0aaab2b2d3d", + "prompt_hash": "4ab0e7faeb3668d22081b105fca0cb00395e4b53703df6f58970034e5b60fc8d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 87, + "doc": { + "question": "What layout detection model was fine-tuned on MinerU's internal dataset?", + "choices": [ + "DocXchain", + "LayoutLMv3", + "Surya", + "360LayoutAnalysis-Paper" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What layout detection model was fine-tuned on MinerU's internal dataset?\nAnswer:", + " DocXchain" + ], + [ + "Question:What layout detection model was fine-tuned on MinerU's internal dataset?\nAnswer:", + " LayoutLMv3" + ], + [ + "Question:What layout detection model was fine-tuned on MinerU's internal dataset?\nAnswer:", + " Surya" + ], + [ + "Question:What layout detection model was fine-tuned on MinerU's internal dataset?\nAnswer:", + " 360LayoutAnalysis-Paper" + ] + ], + "resps": [ + [ + [ + -27.757598876953125, + false + ] + ], + [ + [ + -11.168102264404297, + false + ] + ], + [ + [ + -17.636960983276367, + false + ] + ], + [ + [ + -40.06000518798828, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.757598876953125, + false + ], + [ + -11.168102264404297, + false + ], + [ + -17.636960983276367, + false + ], + [ + -40.06000518798828, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "78a8f26b79e106deb0cda4f1dc69d21ee3b0cfed8e5d23011f39e98d8778ffa3", + "prompt_hash": "7447f7e97610c9d135bffcb6bedb3973cfcb18b24a95c9b56c025f12be6fc215", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 88, + "doc": { + "question": "Which two models are used by MinerU for table recognition tasks?", + "choices": [ + "StructEqTable and TableMaster", + "DocXchain and TableMaster", + "Pix2tex and StructEqTable", + "Texify and TableMaster" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which two models are used by MinerU for table recognition tasks?\nAnswer:", + " StructEqTable and TableMaster" + ], + [ + "Question:Which two models are used by MinerU for table recognition tasks?\nAnswer:", + " DocXchain and TableMaster" + ], + [ + "Question:Which two models are used by MinerU for table recognition tasks?\nAnswer:", + " Pix2tex and StructEqTable" + ], + [ + "Question:Which two models are used by MinerU for table recognition tasks?\nAnswer:", + " Texify and TableMaster" + ] + ], + "resps": [ + [ + [ + -41.12918472290039, + false + ] + ], + [ + [ + -46.08926773071289, + false + ] + ], + [ + [ + -45.41828918457031, + false + ] + ], + [ + [ + -38.265869140625, + false + ] + ] + ], + "filtered_resps": [ + [ + -41.12918472290039, + false + ], + [ + -46.08926773071289, + false + ], + [ + -45.41828918457031, + false + ], + [ + -38.265869140625, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7ebc7a53eeb4822f4633bb50c81307b737b7e5d75cedec0d907633580d14d0c7", + "prompt_hash": "94f14cd39b28fc71d7d77acfeda9bce9541e04e522aeb95d34315422b2fb94c5", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 89, + "doc": { + "question": "What document attribute determines whether MinerU applies OCR or direct text extraction?", + "choices": [ + "Document file size", + "Language of the document", + "Presence of formulas", + "Scanned vs. text-based PDF" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What document attribute determines whether MinerU applies OCR or direct text extraction?\nAnswer:", + " Document file size" + ], + [ + "Question:What document attribute determines whether MinerU applies OCR or direct text extraction?\nAnswer:", + " Language of the document" + ], + [ + "Question:What document attribute determines whether MinerU applies OCR or direct text extraction?\nAnswer:", + " Presence of formulas" + ], + [ + "Question:What document attribute determines whether MinerU applies OCR or direct text extraction?\nAnswer:", + " Scanned vs. text-based PDF" + ] + ], + "resps": [ + [ + [ + -15.031461715698242, + false + ] + ], + [ + [ + -13.443138122558594, + false + ] + ], + [ + [ + -21.756549835205078, + false + ] + ], + [ + [ + -32.97414016723633, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.031461715698242, + false + ], + [ + -13.443138122558594, + false + ], + [ + -21.756549835205078, + false + ], + [ + -32.97414016723633, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b65253960a8d398702513c3de6c03b58763a782ea4c8eb76d51069f4c1a83a16", + "prompt_hash": "6ec9281ad15979a5e1a2f3d8be4a535782184f3f378b9de562997926427b1b65", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 90, + "doc": { + "question": "What is the codebook size used by the vision tokenizer?", + "choices": [ + "16384", + "32768", + "65536", + "4096" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the codebook size used by the vision tokenizer?\nAnswer:", + " 16384" + ], + [ + "Question:What is the codebook size used by the vision tokenizer?\nAnswer:", + " 32768" + ], + [ + "Question:What is the codebook size used by the vision tokenizer?\nAnswer:", + " 65536" + ], + [ + "Question:What is the codebook size used by the vision tokenizer?\nAnswer:", + " 4096" + ] + ], + "resps": [ + [ + [ + -9.472506523132324, + false + ] + ], + [ + [ + -9.326908111572266, + false + ] + ], + [ + [ + -8.893693923950195, + false + ] + ], + [ + [ + -6.730905532836914, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.472506523132324, + false + ], + [ + -9.326908111572266, + false + ], + [ + -8.893693923950195, + false + ], + [ + -6.730905532836914, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "921138505deac46d1dc26ce3e7d95d3173d37af399a55da377c25799cf397256", + "prompt_hash": "8c13fc6441acf5a97158c0331c0abc98f3f5d45d7793efd374c10a6dab8654bb", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 91, + "doc": { + "question": "What is the total context length supported by the model during video training?", + "choices": [ + "4096", + "131072", + "2048", + "65536" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the total context length supported by the model during video training?\nAnswer:", + " 4096" + ], + [ + "Question:What is the total context length supported by the model during video training?\nAnswer:", + " 131072" + ], + [ + "Question:What is the total context length supported by the model during video training?\nAnswer:", + " 2048" + ], + [ + "Question:What is the total context length supported by the model during video training?\nAnswer:", + " 65536" + ] + ], + "resps": [ + [ + [ + -6.252519130706787, + false + ] + ], + [ + [ + -11.824712753295898, + false + ] + ], + [ + [ + -6.172224998474121, + false + ] + ], + [ + [ + -10.215291976928711, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.252519130706787, + false + ], + [ + -11.824712753295898, + false + ], + [ + -6.172224998474121, + false + ], + [ + -10.215291976928711, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "bfeee6c195dc0a3c16023724818ceb8649450dbbb88b1962eab3682c072102b1", + "prompt_hash": "9c81cc8c1465e209432661336484c3e2cfd0aafef5bb044705a5b1a7e7ec3ce2", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 92, + "doc": { + "question": "What filtering threshold was used with the LAION-AI aesthetic predictor for images?", + "choices": [ + "5.5", + "4.0", + "6.0", + "5.0" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What filtering threshold was used with the LAION-AI aesthetic predictor for images?\nAnswer:", + " 5.5" + ], + [ + "Question:What filtering threshold was used with the LAION-AI aesthetic predictor for images?\nAnswer:", + " 4.0" + ], + [ + "Question:What filtering threshold was used with the LAION-AI aesthetic predictor for images?\nAnswer:", + " 6.0" + ], + [ + "Question:What filtering threshold was used with the LAION-AI aesthetic predictor for images?\nAnswer:", + " 5.0" + ] + ], + "resps": [ + [ + [ + -8.747995376586914, + false + ] + ], + [ + [ + -8.179380416870117, + false + ] + ], + [ + [ + -9.052521705627441, + false + ] + ], + [ + [ + -7.995424270629883, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.747995376586914, + false + ], + [ + -8.179380416870117, + false + ], + [ + -9.052521705627441, + false + ], + [ + -7.995424270629883, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e58253cf7bf1b42c0e218db5cc1e700afe70d68ab4c4326e0ca1c1393d3b2508", + "prompt_hash": "f104f738b2da0ffbac987df020d34e8922606c468203a1c53463f2ffb1220886", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 93, + "doc": { + "question": "Which activation function is used in the model architecture?", + "choices": [ + "ReLU", + "GELU", + "SwiGLU", + "LeakyReLU" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which activation function is used in the model architecture?\nAnswer:", + " ReLU" + ], + [ + "Question:Which activation function is used in the model architecture?\nAnswer:", + " GELU" + ], + [ + "Question:Which activation function is used in the model architecture?\nAnswer:", + " SwiGLU" + ], + [ + "Question:Which activation function is used in the model architecture?\nAnswer:", + " LeakyReLU" + ] + ], + "resps": [ + [ + [ + -2.701192855834961, + false + ] + ], + [ + [ + -6.409379482269287, + false + ] + ], + [ + [ + -12.06315803527832, + false + ] + ], + [ + [ + -4.720894813537598, + false + ] + ] + ], + "filtered_resps": [ + [ + -2.701192855834961, + false + ], + [ + -6.409379482269287, + false + ], + [ + -12.06315803527832, + false + ], + [ + -4.720894813537598, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fa9b33bb481b548158c9744834bef4ed62e97a0d0893c4e5c69840958b0b0189", + "prompt_hash": "13ba185649417990c269bbd6c34c95cbfc797c8f4234e5923414f85e38ab18ae", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 94, + "doc": { + "question": "What are the dimensions of the visual tokens that the vision tokenizer compresses a 512x512 image into?", + "choices": [ + "64", + "4096", + "512", + "32" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What are the dimensions of the visual tokens that the vision tokenizer compresses a 512x512 image into?\nAnswer:", + " 64" + ], + [ + "Question:What are the dimensions of the visual tokens that the vision tokenizer compresses a 512x512 image into?\nAnswer:", + " 4096" + ], + [ + "Question:What are the dimensions of the visual tokens that the vision tokenizer compresses a 512x512 image into?\nAnswer:", + " 512" + ], + [ + "Question:What are the dimensions of the visual tokens that the vision tokenizer compresses a 512x512 image into?\nAnswer:", + " 32" + ] + ], + "resps": [ + [ + [ + -5.736916542053223, + false + ] + ], + [ + [ + -7.617210388183594, + false + ] + ], + [ + [ + -5.7705512046813965, + false + ] + ], + [ + [ + -5.485168933868408, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.736916542053223, + false + ], + [ + -7.617210388183594, + false + ], + [ + -5.7705512046813965, + false + ], + [ + -5.485168933868408, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d480b3c765bc0fd2dfba85e7928b9e6256f3d92f03d35f0c172d8ed3889bead1", + "prompt_hash": "84fe528602ee3059051e8e68a2b6b11ca0e4eda52876bde6554ba7e13bc5c0d5", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 95, + "doc": { + "question": "What simulator does the PhysGen system use for 2D rigid body dynamics?", + "choices": [ + "PyBullet", + "MuJoCo", + "Pymunk", + "PhysX" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What simulator does the PhysGen system use for 2D rigid body dynamics?\nAnswer:", + " PyBullet" + ], + [ + "Question:What simulator does the PhysGen system use for 2D rigid body dynamics?\nAnswer:", + " MuJoCo" + ], + [ + "Question:What simulator does the PhysGen system use for 2D rigid body dynamics?\nAnswer:", + " Pymunk" + ], + [ + "Question:What simulator does the PhysGen system use for 2D rigid body dynamics?\nAnswer:", + " PhysX" + ] + ], + "resps": [ + [ + [ + -6.650575160980225, + false + ] + ], + [ + [ + -8.297380447387695, + false + ] + ], + [ + [ + -6.8455810546875, + false + ] + ], + [ + [ + -8.160148620605469, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.650575160980225, + false + ], + [ + -8.297380447387695, + false + ], + [ + -6.8455810546875, + false + ], + [ + -8.160148620605469, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e89640744aa137f59ddd183fee7a69b64a815e0ea1445e6748f305d2f2d35874", + "prompt_hash": "1babfd91f4788bc86342e6f0f603c60b26327f96ac47d8dcd91b228980968ac0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 96, + "doc": { + "question": "What is the resolution used for video generation in the experiments?", + "choices": [ + "256 × 256", + "768 × 768", + "1024 × 1024", + "512 × 512" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the resolution used for video generation in the experiments?\nAnswer:", + " 256 × 256" + ], + [ + "Question:What is the resolution used for video generation in the experiments?\nAnswer:", + " 768 × 768" + ], + [ + "Question:What is the resolution used for video generation in the experiments?\nAnswer:", + " 1024 × 1024" + ], + [ + "Question:What is the resolution used for video generation in the experiments?\nAnswer:", + " 512 × 512" + ] + ], + "resps": [ + [ + [ + -7.873867511749268, + false + ] + ], + [ + [ + -10.801594734191895, + false + ] + ], + [ + [ + -8.7615966796875, + false + ] + ], + [ + [ + -8.544112205505371, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.873867511749268, + false + ], + [ + -10.801594734191895, + false + ], + [ + -8.7615966796875, + false + ], + [ + -8.544112205505371, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c1f9aac27b758c6c9bbe16854de2f0da20a0063ce56679800a51b8a47ec0a5de", + "prompt_hash": "98db683fc1ff393d1165bf42f2254a7124c024e7a9fbb969eac3f5662de74532", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 97, + "doc": { + "question": "What metric is used to evaluate image appearance in the quantitative evaluation?", + "choices": [ + "Fréchet Video Distance", + "Peak Signal-to-Noise Ratio", + "Image-FID", + "SSIM" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What metric is used to evaluate image appearance in the quantitative evaluation?\nAnswer:", + " Fréchet Video Distance" + ], + [ + "Question:What metric is used to evaluate image appearance in the quantitative evaluation?\nAnswer:", + " Peak Signal-to-Noise Ratio" + ], + [ + "Question:What metric is used to evaluate image appearance in the quantitative evaluation?\nAnswer:", + " Image-FID" + ], + [ + "Question:What metric is used to evaluate image appearance in the quantitative evaluation?\nAnswer:", + " SSIM" + ] + ], + "resps": [ + [ + [ + -14.589303016662598, + false + ] + ], + [ + [ + -7.220948219299316, + false + ] + ], + [ + [ + -19.263450622558594, + false + ] + ], + [ + [ + -6.307013034820557, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.589303016662598, + false + ], + [ + -7.220948219299316, + false + ], + [ + -19.263450622558594, + false + ], + [ + -6.307013034820557, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b2c8febd871fe21d664c54b2885582a829ee64ca1ee70a739f09058a533511f8", + "prompt_hash": "417d34dd90cf897e51f620454efe10802be46cc668007e267dd4afa55052b0d9", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 98, + "doc": { + "question": "How many human participants were involved in the physical-realism and photo-realism evaluation?", + "choices": [ + "10", + "12", + "14", + "20" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many human participants were involved in the physical-realism and photo-realism evaluation?\nAnswer:", + " 10" + ], + [ + "Question:How many human participants were involved in the physical-realism and photo-realism evaluation?\nAnswer:", + " 12" + ], + [ + "Question:How many human participants were involved in the physical-realism and photo-realism evaluation?\nAnswer:", + " 14" + ], + [ + "Question:How many human participants were involved in the physical-realism and photo-realism evaluation?\nAnswer:", + " 20" + ] + ], + "resps": [ + [ + [ + -4.0728678703308105, + true + ] + ], + [ + [ + -4.49506139755249, + false + ] + ], + [ + [ + -5.605117321014404, + false + ] + ], + [ + [ + -4.566987037658691, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.0728678703308105, + true + ], + [ + -4.49506139755249, + false + ], + [ + -5.605117321014404, + false + ], + [ + -4.566987037658691, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2e057fc17dce303739e665d9ffdd88f0cb75faaee88f577b0132ad378807aa61", + "prompt_hash": "0b4f201db05858d3bae3b02a7724eeb3f7961656c58f1ad4fc98cf12ff486654", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 99, + "doc": { + "question": "Which model is used to estimate surface normals for intrinsic decomposition?", + "choices": [ + "OmniData", + "GeoWizard", + "DINOv2", + "SAM" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which model is used to estimate surface normals for intrinsic decomposition?\nAnswer:", + " OmniData" + ], + [ + "Question:Which model is used to estimate surface normals for intrinsic decomposition?\nAnswer:", + " GeoWizard" + ], + [ + "Question:Which model is used to estimate surface normals for intrinsic decomposition?\nAnswer:", + " DINOv2" + ], + [ + "Question:Which model is used to estimate surface normals for intrinsic decomposition?\nAnswer:", + " SAM" + ] + ], + "resps": [ + [ + [ + -20.465606689453125, + false + ] + ], + [ + [ + -20.510021209716797, + false + ] + ], + [ + [ + -10.1165771484375, + false + ] + ], + [ + [ + -10.176116943359375, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.465606689453125, + false + ], + [ + -20.510021209716797, + false + ], + [ + -10.1165771484375, + false + ], + [ + -10.176116943359375, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "be853caa9fc164305d79a18a3a24573ea1963c2fb005183ae8ce413eebebab46", + "prompt_hash": "28f78e448df786530b10b8c768afc1db85a16371e1d6f722e67ba3779c48fb82", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + } + ], + "arxiv_mc_2410": [ + { + "doc_id": 0, + "doc": { + "question": "What are the three main phases of the TECH approach for fixing software issues?", + "choices": [ + "Planning, Execution, Review", + "Localization, Repair, Patch Validation", + "Input Parsing, Compilation, Testing", + "Scanning, Refactoring, Deployment" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What are the three main phases of the TECH approach for fixing software issues?\nAnswer:", + " Planning, Execution, Review" + ], + [ + "Question:What are the three main phases of the TECH approach for fixing software issues?\nAnswer:", + " Localization, Repair, Patch Validation" + ], + [ + "Question:What are the three main phases of the TECH approach for fixing software issues?\nAnswer:", + " Input Parsing, Compilation, Testing" + ], + [ + "Question:What are the three main phases of the TECH approach for fixing software issues?\nAnswer:", + " Scanning, Refactoring, Deployment" + ] + ], + "resps": [ + [ + [ + -11.142742156982422, + false + ] + ], + [ + [ + -31.907644271850586, + false + ] + ], + [ + [ + -28.914268493652344, + false + ] + ], + [ + [ + -23.669246673583984, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.142742156982422, + false + ], + [ + -31.907644271850586, + false + ], + [ + -28.914268493652344, + false + ], + [ + -23.669246673583984, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8f30c31f2913f7a7945de0d4dc3242e4f453933de09ccf31eb6e210083d748c4", + "prompt_hash": "d785864492154380bf1f9f3e96c251bc75978f55738006395216d3242e45aff9", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 1, + "doc": { + "question": "What format does TECH use to concisely represent a repository's file and directory structure?", + "choices": [ + "Markdown format", + "Skeleton format", + "Tree-like structure called repostructureformat", + "JSON representation" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What format does TECH use to concisely represent a repository's file and directory structure?\nAnswer:", + " Markdown format" + ], + [ + "Question:What format does TECH use to concisely represent a repository's file and directory structure?\nAnswer:", + " Skeleton format" + ], + [ + "Question:What format does TECH use to concisely represent a repository's file and directory structure?\nAnswer:", + " Tree-like structure called repostructureformat" + ], + [ + "Question:What format does TECH use to concisely represent a repository's file and directory structure?\nAnswer:", + " JSON representation" + ] + ], + "resps": [ + [ + [ + -14.080405235290527, + false + ] + ], + [ + [ + -19.98662567138672, + false + ] + ], + [ + [ + -49.63202667236328, + false + ] + ], + [ + [ + -11.538562774658203, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.080405235290527, + false + ], + [ + -19.98662567138672, + false + ], + [ + -49.63202667236328, + false + ], + [ + -11.538562774658203, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "61bf0927a6d4fb046ae2fe9787c168288efc1338d0565d93b1e933f503142de9", + "prompt_hash": "c2e4f29b8ab34d9218f5de34f173a908265aa157bce548d6b3e07479e957b9c3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 2, + "doc": { + "question": "What type of test does TECH synthesize to verify whether an issue has been fixed?", + "choices": [ + "Regression test", + "Smoke test", + "Reproduction test", + "Integration test" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What type of test does TECH synthesize to verify whether an issue has been fixed?\nAnswer:", + " Regression test" + ], + [ + "Question:What type of test does TECH synthesize to verify whether an issue has been fixed?\nAnswer:", + " Smoke test" + ], + [ + "Question:What type of test does TECH synthesize to verify whether an issue has been fixed?\nAnswer:", + " Reproduction test" + ], + [ + "Question:What type of test does TECH synthesize to verify whether an issue has been fixed?\nAnswer:", + " Integration test" + ] + ], + "resps": [ + [ + [ + -9.417646408081055, + false + ] + ], + [ + [ + -14.601350784301758, + false + ] + ], + [ + [ + -12.118673324584961, + false + ] + ], + [ + [ + -10.362127304077148, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.417646408081055, + false + ], + [ + -14.601350784301758, + false + ], + [ + -12.118673324584961, + false + ], + [ + -10.362127304077148, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1f4bf8b5c4ff2a11454e56e263d85156cae6700871a341481e534898e5687dff", + "prompt_hash": "023e52b0faef56033cb02985f46f356e563fb5058f7ea12bade09233d6921c5e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 3, + "doc": { + "question": "What diff format does TECH use during the repair phase to apply patches?", + "choices": [ + "Context-based patch", + "Unified diff format", + "Search/Replace edit format", + "Reverse diff format" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What diff format does TECH use during the repair phase to apply patches?\nAnswer:", + " Context-based patch" + ], + [ + "Question:What diff format does TECH use during the repair phase to apply patches?\nAnswer:", + " Unified diff format" + ], + [ + "Question:What diff format does TECH use during the repair phase to apply patches?\nAnswer:", + " Search/Replace edit format" + ], + [ + "Question:What diff format does TECH use during the repair phase to apply patches?\nAnswer:", + " Reverse diff format" + ] + ], + "resps": [ + [ + [ + -18.5589599609375, + false + ] + ], + [ + [ + -16.782909393310547, + false + ] + ], + [ + [ + -37.64253616333008, + false + ] + ], + [ + [ + -20.763954162597656, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.5589599609375, + false + ], + [ + -16.782909393310547, + false + ], + [ + -37.64253616333008, + false + ], + [ + -20.763954162597656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f2c5708f345133a398dfca01d690074eb387df7eb91fb481c8f5c47295e2e96e", + "prompt_hash": "79bfec124e28d4bf05de9556619b459d58f71b627bcd66101040b5c37e5b58a0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 4, + "doc": { + "question": "Which benchmark did OpenAI release to address concerns about problem quality in SWE-bench Lite?", + "choices": [ + "SWE-bench Pro", + "SWE-bench Alpha", + "SWE-bench Verified", + "SWE-bench Debug" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which benchmark did OpenAI release to address concerns about problem quality in SWE-bench Lite?\nAnswer:", + " SWE-bench Pro" + ], + [ + "Question:Which benchmark did OpenAI release to address concerns about problem quality in SWE-bench Lite?\nAnswer:", + " SWE-bench Alpha" + ], + [ + "Question:Which benchmark did OpenAI release to address concerns about problem quality in SWE-bench Lite?\nAnswer:", + " SWE-bench Verified" + ], + [ + "Question:Which benchmark did OpenAI release to address concerns about problem quality in SWE-bench Lite?\nAnswer:", + " SWE-bench Debug" + ] + ], + "resps": [ + [ + [ + -11.525687217712402, + false + ] + ], + [ + [ + -13.013885498046875, + false + ] + ], + [ + [ + -15.520681381225586, + false + ] + ], + [ + [ + -14.712057113647461, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.525687217712402, + false + ], + [ + -13.013885498046875, + false + ], + [ + -15.520681381225586, + false + ], + [ + -14.712057113647461, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "55486ad2ef065f2cbbcca4b5038e34a1a0b0d244820634cea34a7dc17c5da3b5", + "prompt_hash": "aca54400c5b3ff7d1572b9cc03a80d7966fd32951ea88e9eca8558bdaa3734c3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 5, + "doc": { + "question": "What type of image encoder does PaliGemma use?", + "choices": [ + "A. CLIP", + "B. SigLIP", + "C. BLIP-2", + "D. Flamingo" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What type of image encoder does PaliGemma use?\nAnswer:", + " A. CLIP" + ], + [ + "Question:What type of image encoder does PaliGemma use?\nAnswer:", + " B. SigLIP" + ], + [ + "Question:What type of image encoder does PaliGemma use?\nAnswer:", + " C. BLIP-2" + ], + [ + "Question:What type of image encoder does PaliGemma use?\nAnswer:", + " D. Flamingo" + ] + ], + "resps": [ + [ + [ + -9.815264701843262, + false + ] + ], + [ + [ + -21.23035430908203, + false + ] + ], + [ + [ + -16.70880126953125, + false + ] + ], + [ + [ + -20.90701675415039, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.815264701843262, + false + ], + [ + -21.23035430908203, + false + ], + [ + -16.70880126953125, + false + ], + [ + -20.90701675415039, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "faf0292756db6246476c88adf99a11951deb3b1c6118964ba5bd87de03a8fc52", + "prompt_hash": "a43e47ae912a3783d89098b4568202ae426e79d14917bf43d0505e67195ab643", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 6, + "doc": { + "question": "How many new tokens were added to support PaliGemma’s structured computer vision tasks?", + "choices": [ + "A. 1024", + "B. 128", + "C. 1152", + "D. 2048" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many new tokens were added to support PaliGemma’s structured computer vision tasks?\nAnswer:", + " A. 1024" + ], + [ + "Question:How many new tokens were added to support PaliGemma’s structured computer vision tasks?\nAnswer:", + " B. 128" + ], + [ + "Question:How many new tokens were added to support PaliGemma’s structured computer vision tasks?\nAnswer:", + " C. 1152" + ], + [ + "Question:How many new tokens were added to support PaliGemma’s structured computer vision tasks?\nAnswer:", + " D. 2048" + ] + ], + "resps": [ + [ + [ + -17.1997013092041, + false + ] + ], + [ + [ + -19.620742797851562, + false + ] + ], + [ + [ + -23.257474899291992, + false + ] + ], + [ + [ + -23.265338897705078, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.1997013092041, + false + ], + [ + -19.620742797851562, + false + ], + [ + -23.257474899291992, + false + ], + [ + -23.265338897705078, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b86bf0c3962ea72631ecfbb84ea987beeb0eea7dec02c015c5666bd94d1e99ef", + "prompt_hash": "6a3f272c72a41669cf58da1d9c24c62d760682550711922ffbd1ffbb7289b811", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 7, + "doc": { + "question": "What resolution does PaliGemma use during Stage 1 pretraining?", + "choices": [ + "A. 224px", + "B. 448px", + "C. 896px", + "D. 1024px" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What resolution does PaliGemma use during Stage 1 pretraining?\nAnswer:", + " A. 224px" + ], + [ + "Question:What resolution does PaliGemma use during Stage 1 pretraining?\nAnswer:", + " B. 448px" + ], + [ + "Question:What resolution does PaliGemma use during Stage 1 pretraining?\nAnswer:", + " C. 896px" + ], + [ + "Question:What resolution does PaliGemma use during Stage 1 pretraining?\nAnswer:", + " D. 1024px" + ] + ], + "resps": [ + [ + [ + -18.129924774169922, + false + ] + ], + [ + [ + -24.31297492980957, + false + ] + ], + [ + [ + -24.918224334716797, + false + ] + ], + [ + [ + -21.301815032958984, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.129924774169922, + false + ], + [ + -24.31297492980957, + false + ], + [ + -24.918224334716797, + false + ], + [ + -21.301815032958984, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b998212e8cc1fb9f15a5fb0fc5d89b81af35494c6ac9464feb23ec42ff9de43e", + "prompt_hash": "33be460a43eec891167ef62e7ecffcb88a69dcf9f4538a1bafd197a9dd31f241", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 8, + "doc": { + "question": "Which language model does PaliGemma use as its decoder?", + "choices": [ + "A. Gemma-7B", + "B. PaLM-E", + "C. Gemma-2B", + "D. mT5" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which language model does PaliGemma use as its decoder?\nAnswer:", + " A. Gemma-7B" + ], + [ + "Question:Which language model does PaliGemma use as its decoder?\nAnswer:", + " B. PaLM-E" + ], + [ + "Question:Which language model does PaliGemma use as its decoder?\nAnswer:", + " C. Gemma-2B" + ], + [ + "Question:Which language model does PaliGemma use as its decoder?\nAnswer:", + " D. mT5" + ] + ], + "resps": [ + [ + [ + -13.4105863571167, + false + ] + ], + [ + [ + -25.003814697265625, + false + ] + ], + [ + [ + -18.639402389526367, + false + ] + ], + [ + [ + -19.03719139099121, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.4105863571167, + false + ], + [ + -25.003814697265625, + false + ], + [ + -18.639402389526367, + false + ], + [ + -19.03719139099121, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8ea584d27437c754402c4878ff8038a7b28b01ee7f36e783ec652978aed23dee", + "prompt_hash": "e8dd45cd5bb9e2e28d6ae138f4fe19b7ae51a6ff9b895ae462dba64b3f5cda97", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 9, + "doc": { + "question": "What kind of connector is used between SigLIP and Gemma in PaliGemma?", + "choices": [ + "A. MLP with attention", + "B. Q-Former", + "C. Linear projection", + "D. Convolutional adapter" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What kind of connector is used between SigLIP and Gemma in PaliGemma?\nAnswer:", + " A. MLP with attention" + ], + [ + "Question:What kind of connector is used between SigLIP and Gemma in PaliGemma?\nAnswer:", + " B. Q-Former" + ], + [ + "Question:What kind of connector is used between SigLIP and Gemma in PaliGemma?\nAnswer:", + " C. Linear projection" + ], + [ + "Question:What kind of connector is used between SigLIP and Gemma in PaliGemma?\nAnswer:", + " D. Convolutional adapter" + ] + ], + "resps": [ + [ + [ + -24.83165740966797, + false + ] + ], + [ + [ + -24.223936080932617, + false + ] + ], + [ + [ + -24.006492614746094, + false + ] + ], + [ + [ + -31.108749389648438, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.83165740966797, + false + ], + [ + -24.223936080932617, + false + ], + [ + -24.006492614746094, + false + ], + [ + -31.108749389648438, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c3b705e13367b24fe44ddee6efc65b4f95aee365c48c1376151412963b520013", + "prompt_hash": "373ffce3f8709c674ae833152b3ed551fb570bec602b3187cee3b62910992334", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 10, + "doc": { + "question": "What is the size of the vocabulary used in the Gemma 2 models?", + "choices": [ + "128k", + "256k", + "512k", + "64k" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the size of the vocabulary used in the Gemma 2 models?\nAnswer:", + " 128k" + ], + [ + "Question:What is the size of the vocabulary used in the Gemma 2 models?\nAnswer:", + " 256k" + ], + [ + "Question:What is the size of the vocabulary used in the Gemma 2 models?\nAnswer:", + " 512k" + ], + [ + "Question:What is the size of the vocabulary used in the Gemma 2 models?\nAnswer:", + " 64k" + ] + ], + "resps": [ + [ + [ + -7.427914619445801, + false + ] + ], + [ + [ + -8.645819664001465, + false + ] + ], + [ + [ + -9.469427108764648, + false + ] + ], + [ + [ + -7.870913505554199, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.427914619445801, + false + ], + [ + -8.645819664001465, + false + ], + [ + -9.469427108764648, + false + ], + [ + -7.870913505554199, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5b9d9b32959d1bd9a3838b7e692d123eb3256ed4d79f1b5f6341ab4833443bec", + "prompt_hash": "c21c35818306c97b4e6bc5c857e0494f9b0ab1ca1aac07486760d842b4679594", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 11, + "doc": { + "question": "What non-linearity function is used in all Gemma 2 models?", + "choices": [ + "ReLU", + "Swish", + "GeGLU", + "Sigmoid" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What non-linearity function is used in all Gemma 2 models?\nAnswer:", + " ReLU" + ], + [ + "Question:What non-linearity function is used in all Gemma 2 models?\nAnswer:", + " Swish" + ], + [ + "Question:What non-linearity function is used in all Gemma 2 models?\nAnswer:", + " GeGLU" + ], + [ + "Question:What non-linearity function is used in all Gemma 2 models?\nAnswer:", + " Sigmoid" + ] + ], + "resps": [ + [ + [ + -4.497497081756592, + false + ] + ], + [ + [ + -9.737420082092285, + false + ] + ], + [ + [ + -24.105804443359375, + false + ] + ], + [ + [ + -4.997573375701904, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.497497081756592, + false + ], + [ + -9.737420082092285, + false + ], + [ + -24.105804443359375, + false + ], + [ + -4.997573375701904, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0b8e6c5eb1b5746b793ab356c9c33eecaa945b7e5f7a27fe29d9ff7058c46623", + "prompt_hash": "b2f86be38f5913ba3b9cef331d432f708e5a912ecb9ea64dc3dd2fe675a3ba25", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 12, + "doc": { + "question": "Which mechanism does Gemma 2 use to improve inference speed while maintaining downstream performance?", + "choices": [ + "Multi-Head Attention", + "Rotary Position Embedding", + "Grouped-Query Attention", + "Sparse Transformer" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which mechanism does Gemma 2 use to improve inference speed while maintaining downstream performance?\nAnswer:", + " Multi-Head Attention" + ], + [ + "Question:Which mechanism does Gemma 2 use to improve inference speed while maintaining downstream performance?\nAnswer:", + " Rotary Position Embedding" + ], + [ + "Question:Which mechanism does Gemma 2 use to improve inference speed while maintaining downstream performance?\nAnswer:", + " Grouped-Query Attention" + ], + [ + "Question:Which mechanism does Gemma 2 use to improve inference speed while maintaining downstream performance?\nAnswer:", + " Sparse Transformer" + ] + ], + "resps": [ + [ + [ + -15.394296646118164, + false + ] + ], + [ + [ + -20.19460678100586, + false + ] + ], + [ + [ + -15.833362579345703, + false + ] + ], + [ + [ + -17.08722686767578, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.394296646118164, + false + ], + [ + -20.19460678100586, + false + ], + [ + -15.833362579345703, + false + ], + [ + -17.08722686767578, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "785e76f9c2316a0ed775048f820bed3227d85f88e8e34ccc1792f82dce848c38", + "prompt_hash": "461d52f64e24115c0807730ff58a90a5c0bdbaed6f3fe71b3b9907b17dea6b4e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 13, + "doc": { + "question": "How many tokens was the Gemma 2 27B model trained on?", + "choices": [ + "3 trillion", + "8 trillion", + "13 trillion", + "15 trillion" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many tokens was the Gemma 2 27B model trained on?\nAnswer:", + " 3 trillion" + ], + [ + "Question:How many tokens was the Gemma 2 27B model trained on?\nAnswer:", + " 8 trillion" + ], + [ + "Question:How many tokens was the Gemma 2 27B model trained on?\nAnswer:", + " 13 trillion" + ], + [ + "Question:How many tokens was the Gemma 2 27B model trained on?\nAnswer:", + " 15 trillion" + ] + ], + "resps": [ + [ + [ + -10.128683090209961, + false + ] + ], + [ + [ + -12.361979484558105, + false + ] + ], + [ + [ + -12.137228012084961, + false + ] + ], + [ + [ + -12.41441822052002, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.128683090209961, + false + ], + [ + -12.361979484558105, + false + ], + [ + -12.137228012084961, + false + ], + [ + -12.41441822052002, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7eb7aedba0cadb9438df3db05265fd3b52d42e554ef64a857b9b1b71a4218dcc", + "prompt_hash": "f0593255989772c2c18a7e5a6b143ae507e838dcfe8a0b3b5200723bd40eba70", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 14, + "doc": { + "question": "What soft cap value is used for logits in the final layer of Gemma 2 models?", + "choices": [ + "10.0", + "30.0", + "50.0", + "100.0" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What soft cap value is used for logits in the final layer of Gemma 2 models?\nAnswer:", + " 10.0" + ], + [ + "Question:What soft cap value is used for logits in the final layer of Gemma 2 models?\nAnswer:", + " 30.0" + ], + [ + "Question:What soft cap value is used for logits in the final layer of Gemma 2 models?\nAnswer:", + " 50.0" + ], + [ + "Question:What soft cap value is used for logits in the final layer of Gemma 2 models?\nAnswer:", + " 100.0" + ] + ], + "resps": [ + [ + [ + -6.458868026733398, + false + ] + ], + [ + [ + -8.575843811035156, + false + ] + ], + [ + [ + -8.410329818725586, + false + ] + ], + [ + [ + -8.538220405578613, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.458868026733398, + false + ], + [ + -8.575843811035156, + false + ], + [ + -8.410329818725586, + false + ], + [ + -8.538220405578613, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "95dc03314caef85c92c77964e354165d2854429f6ba048362ff298af844c0c30", + "prompt_hash": "02aba881c3e2b1813dc814c5c6f5e4b4af64d6f6de73eda92da54299c1a38d22", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 15, + "doc": { + "question": "Which language model is used as the LLM in the architecture?", + "choices": [ + "Qwen-2", + "GPT-3.5", + "LLaMA-2", + "BLOOM" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which language model is used as the LLM in the architecture?\nAnswer:", + " Qwen-2" + ], + [ + "Question:Which language model is used as the LLM in the architecture?\nAnswer:", + " GPT-3.5" + ], + [ + "Question:Which language model is used as the LLM in the architecture?\nAnswer:", + " LLaMA-2" + ], + [ + "Question:Which language model is used as the LLM in the architecture?\nAnswer:", + " BLOOM" + ] + ], + "resps": [ + [ + [ + -10.288427352905273, + false + ] + ], + [ + [ + -4.0216217041015625, + false + ] + ], + [ + [ + -8.288175582885742, + false + ] + ], + [ + [ + -7.530017375946045, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.288427352905273, + false + ], + [ + -4.0216217041015625, + false + ], + [ + -8.288175582885742, + false + ], + [ + -7.530017375946045, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "763e2d4590d8631d47a491547a5c25bc18e8897ebad0deb508b89de5df974438", + "prompt_hash": "a251e3cc0a376c8158837560ea2b3cbf69857501397f91508554761288bf785f", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 16, + "doc": { + "question": "What visual encoder is used to generate visual features from images?", + "choices": [ + "CLIP", + "SigLIP", + "ViT-B/32", + "BLIP-2" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What visual encoder is used to generate visual features from images?\nAnswer:", + " CLIP" + ], + [ + "Question:What visual encoder is used to generate visual features from images?\nAnswer:", + " SigLIP" + ], + [ + "Question:What visual encoder is used to generate visual features from images?\nAnswer:", + " ViT-B/32" + ], + [ + "Question:What visual encoder is used to generate visual features from images?\nAnswer:", + " BLIP-2" + ] + ], + "resps": [ + [ + [ + -4.400332927703857, + false + ] + ], + [ + [ + -12.483972549438477, + false + ] + ], + [ + [ + -12.72900390625, + false + ] + ], + [ + [ + -9.399051666259766, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.400332927703857, + false + ], + [ + -12.483972549438477, + false + ], + [ + -12.72900390625, + false + ], + [ + -9.399051666259766, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1fed2d0f3f806f0ede4ac3dcd2e24da5af46bbe8c681c54f27c660efdae75d64", + "prompt_hash": "76a68c139bf7c017822524c8f45b702cf3e64b99258e8479dcefddc1bcfb3aae", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 17, + "doc": { + "question": "How many samples are included in the Re-Captioned Detailed Description Data?", + "choices": [ + "2.1M", + "3.5M", + "5.0M", + "1.8M" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many samples are included in the Re-Captioned Detailed Description Data?\nAnswer:", + " 2.1M" + ], + [ + "Question:How many samples are included in the Re-Captioned Detailed Description Data?\nAnswer:", + " 3.5M" + ], + [ + "Question:How many samples are included in the Re-Captioned Detailed Description Data?\nAnswer:", + " 5.0M" + ], + [ + "Question:How many samples are included in the Re-Captioned Detailed Description Data?\nAnswer:", + " 1.8M" + ] + ], + "resps": [ + [ + [ + -12.48883056640625, + false + ] + ], + [ + [ + -12.085700988769531, + false + ] + ], + [ + [ + -13.646947860717773, + false + ] + ], + [ + [ + -11.971969604492188, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.48883056640625, + false + ], + [ + -12.085700988769531, + false + ], + [ + -13.646947860717773, + false + ], + [ + -11.971969604492188, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "432a3bf5c6810f0e0bba9d1541cb9434bfb62e6a117cdf56357d718e4ba521a0", + "prompt_hash": "2f9de16b807907efad345b576e9fce196e96e0b258d7e05eae7be249c75cc7ab", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 18, + "doc": { + "question": "What is the maximum number of visual tokens used in Stage-2 of training?", + "choices": [ + "729", + "729 × 5", + "729 × 10", + "729 × 2" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the maximum number of visual tokens used in Stage-2 of training?\nAnswer:", + " 729" + ], + [ + "Question:What is the maximum number of visual tokens used in Stage-2 of training?\nAnswer:", + " 729 × 5" + ], + [ + "Question:What is the maximum number of visual tokens used in Stage-2 of training?\nAnswer:", + " 729 × 10" + ], + [ + "Question:What is the maximum number of visual tokens used in Stage-2 of training?\nAnswer:", + " 729 × 2" + ] + ], + "resps": [ + [ + [ + -9.973566055297852, + false + ] + ], + [ + [ + -23.389095306396484, + false + ] + ], + [ + [ + -22.504749298095703, + false + ] + ], + [ + [ + -22.001413345336914, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.973566055297852, + false + ], + [ + -23.389095306396484, + false + ], + [ + -22.504749298095703, + false + ], + [ + -22.001413345336914, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2bc24be57e51015d15ac5e3e51f61ed93196c209592ae538e612d4f932ea7920", + "prompt_hash": "b48df395d7a06051c3af142362fcfe394f8cfea66edbf90cf90c664dd7904b14", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 19, + "doc": { + "question": "Which dataset was used to generate 92K detailed Chinese caption samples?", + "choices": [ + "COCO", + "ShareGPT4V", + "SynDOG EN", + "Evo-Instruct" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset was used to generate 92K detailed Chinese caption samples?\nAnswer:", + " COCO" + ], + [ + "Question:Which dataset was used to generate 92K detailed Chinese caption samples?\nAnswer:", + " ShareGPT4V" + ], + [ + "Question:Which dataset was used to generate 92K detailed Chinese caption samples?\nAnswer:", + " SynDOG EN" + ], + [ + "Question:Which dataset was used to generate 92K detailed Chinese caption samples?\nAnswer:", + " Evo-Instruct" + ] + ], + "resps": [ + [ + [ + -7.290433406829834, + false + ] + ], + [ + [ + -12.020796775817871, + false + ] + ], + [ + [ + -36.19415283203125, + false + ] + ], + [ + [ + -20.265243530273438, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.290433406829834, + false + ], + [ + -12.020796775817871, + false + ], + [ + -36.19415283203125, + false + ], + [ + -20.265243530273438, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c48e39a64b13161ee1bea7e39ea13c045f0e59c4f0a902a9841855dd286bcb0f", + "prompt_hash": "29c6ecd1445476dff678dd07862a504e9a47079055c8023ffbad73b57364430d", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 20, + "doc": { + "question": "What is the primary reason for the latency in the pre-filling stage of long-context LLMs?", + "choices": [ + "Model initialization", + "Token decoding speed", + "Self-attention computation overhead", + "Memory swapping between GPUs" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the primary reason for the latency in the pre-filling stage of long-context LLMs?\nAnswer:", + " Model initialization" + ], + [ + "Question:What is the primary reason for the latency in the pre-filling stage of long-context LLMs?\nAnswer:", + " Token decoding speed" + ], + [ + "Question:What is the primary reason for the latency in the pre-filling stage of long-context LLMs?\nAnswer:", + " Self-attention computation overhead" + ], + [ + "Question:What is the primary reason for the latency in the pre-filling stage of long-context LLMs?\nAnswer:", + " Memory swapping between GPUs" + ] + ], + "resps": [ + [ + [ + -11.29690170288086, + false + ] + ], + [ + [ + -21.47367286682129, + false + ] + ], + [ + [ + -19.235584259033203, + false + ] + ], + [ + [ + -26.47637367248535, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.29690170288086, + false + ], + [ + -21.47367286682129, + false + ], + [ + -19.235584259033203, + false + ], + [ + -26.47637367248535, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c35537323672d6352eef7842a5fdcd331b963c254c0b1e10fc871a1b0979fa8a", + "prompt_hash": "50c6e338256e19846b03fddb143a281480b7c983a1f11d369bd847a728c3f144", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 21, + "doc": { + "question": "What recall percentage is achieved when only the top 4k columns are retained in a 128k attention matrix?", + "choices": [ + "89.3%", + "92.5%", + "96.8%", + "99.1%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What recall percentage is achieved when only the top 4k columns are retained in a 128k attention matrix?\nAnswer:", + " 89.3%" + ], + [ + "Question:What recall percentage is achieved when only the top 4k columns are retained in a 128k attention matrix?\nAnswer:", + " 92.5%" + ], + [ + "Question:What recall percentage is achieved when only the top 4k columns are retained in a 128k attention matrix?\nAnswer:", + " 96.8%" + ], + [ + "Question:What recall percentage is achieved when only the top 4k columns are retained in a 128k attention matrix?\nAnswer:", + " 99.1%" + ] + ], + "resps": [ + [ + [ + -14.330986022949219, + false + ] + ], + [ + [ + -13.229812622070312, + false + ] + ], + [ + [ + -13.767045974731445, + false + ] + ], + [ + [ + -14.044456481933594, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.330986022949219, + false + ], + [ + -13.229812622070312, + false + ], + [ + -13.767045974731445, + false + ], + [ + -14.044456481933594, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5438925b93c9e52efeef96c678d215f83f32bbc7b162644c312d3fd33139500b", + "prompt_hash": "0ddd6c551b6b0dc2c533e3a6f6d1db7231e1c899b2128d3c6b8f18ca58fda0e8", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 22, + "doc": { + "question": "Which attention pattern is characterized by tokens focusing on initial tokens and local windows?", + "choices": [ + "Block-Sparse pattern", + "Vertical-Slash pattern", + "A-shape pattern", + "Global-Context pattern" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which attention pattern is characterized by tokens focusing on initial tokens and local windows?\nAnswer:", + " Block-Sparse pattern" + ], + [ + "Question:Which attention pattern is characterized by tokens focusing on initial tokens and local windows?\nAnswer:", + " Vertical-Slash pattern" + ], + [ + "Question:Which attention pattern is characterized by tokens focusing on initial tokens and local windows?\nAnswer:", + " A-shape pattern" + ], + [ + "Question:Which attention pattern is characterized by tokens focusing on initial tokens and local windows?\nAnswer:", + " Global-Context pattern" + ] + ], + "resps": [ + [ + [ + -28.077165603637695, + false + ] + ], + [ + [ + -36.12141418457031, + false + ] + ], + [ + [ + -24.157867431640625, + false + ] + ], + [ + [ + -23.051002502441406, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.077165603637695, + false + ], + [ + -36.12141418457031, + false + ], + [ + -24.157867431640625, + false + ], + [ + -23.051002502441406, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7a0b7e981e36e32875a92973f2a7923dd3b684ec424f172722d92441aac32bf0", + "prompt_hash": "fe87432365ef44cd6f6feca83292c45aeb4e6405c58e10d3f0256bf28770f6d3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 23, + "doc": { + "question": "What is the reported speedup using MInference on 1M-token prompts with LLaMA-3-8B on a single A100 GPU?", + "choices": [ + "Up to 3×", + "Up to 5×", + "Up to 8×", + "Up to 10×" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the reported speedup using MInference on 1M-token prompts with LLaMA-3-8B on a single A100 GPU?\nAnswer:", + " Up to 3×" + ], + [ + "Question:What is the reported speedup using MInference on 1M-token prompts with LLaMA-3-8B on a single A100 GPU?\nAnswer:", + " Up to 5×" + ], + [ + "Question:What is the reported speedup using MInference on 1M-token prompts with LLaMA-3-8B on a single A100 GPU?\nAnswer:", + " Up to 8×" + ], + [ + "Question:What is the reported speedup using MInference on 1M-token prompts with LLaMA-3-8B on a single A100 GPU?\nAnswer:", + " Up to 10×" + ] + ], + "resps": [ + [ + [ + -18.81794548034668, + false + ] + ], + [ + [ + -19.063770294189453, + false + ] + ], + [ + [ + -18.887496948242188, + false + ] + ], + [ + [ + -18.622156143188477, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.81794548034668, + false + ], + [ + -19.063770294189453, + false + ], + [ + -18.887496948242188, + false + ], + [ + -18.622156143188477, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "be1003f886be358b8d164b50e52a8632a02e9892a114e08f9b2a19a29a6ea779", + "prompt_hash": "493f78c96da1d5cb7247238dd51eec7334cc3d4421166683f854892037fc3bbd", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 24, + "doc": { + "question": "Which method is used in MInference to approximate sparse indices for Block-Sparse heads?", + "choices": [ + "Top-K ranking of global tokens", + "Mean pooling of query and key vectors", + "Nearest-neighbor clustering", + "Random subsampling of attention weights" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which method is used in MInference to approximate sparse indices for Block-Sparse heads?\nAnswer:", + " Top-K ranking of global tokens" + ], + [ + "Question:Which method is used in MInference to approximate sparse indices for Block-Sparse heads?\nAnswer:", + " Mean pooling of query and key vectors" + ], + [ + "Question:Which method is used in MInference to approximate sparse indices for Block-Sparse heads?\nAnswer:", + " Nearest-neighbor clustering" + ], + [ + "Question:Which method is used in MInference to approximate sparse indices for Block-Sparse heads?\nAnswer:", + " Random subsampling of attention weights" + ] + ], + "resps": [ + [ + [ + -31.907312393188477, + false + ] + ], + [ + [ + -30.570072174072266, + false + ] + ], + [ + [ + -22.7277889251709, + false + ] + ], + [ + [ + -26.002704620361328, + false + ] + ] + ], + "filtered_resps": [ + [ + -31.907312393188477, + false + ], + [ + -30.570072174072266, + false + ], + [ + -22.7277889251709, + false + ], + [ + -26.002704620361328, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1ee9007973ea9d762c1d28efd631ea2df296bb5722e13f9b73476089e17f593c", + "prompt_hash": "dec85aaaced9973622c82d55bd48a6590d9c4c61c214320bc06f4ecb773982a4", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 25, + "doc": { + "question": "What was the absolute self-correction gain of the proposed method on reasoning problems from the MATH dataset?", + "choices": [ + "15.6%", + "9.1%", + "23.0%", + "12.2%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What was the absolute self-correction gain of the proposed method on reasoning problems from the MATH dataset?\nAnswer:", + " 15.6%" + ], + [ + "Question:What was the absolute self-correction gain of the proposed method on reasoning problems from the MATH dataset?\nAnswer:", + " 9.1%" + ], + [ + "Question:What was the absolute self-correction gain of the proposed method on reasoning problems from the MATH dataset?\nAnswer:", + " 23.0%" + ], + [ + "Question:What was the absolute self-correction gain of the proposed method on reasoning problems from the MATH dataset?\nAnswer:", + " 12.2%" + ] + ], + "resps": [ + [ + [ + -15.868279457092285, + false + ] + ], + [ + [ + -15.192159652709961, + false + ] + ], + [ + [ + -15.974629402160645, + false + ] + ], + [ + [ + -15.19686508178711, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.868279457092285, + false + ], + [ + -15.192159652709961, + false + ], + [ + -15.974629402160645, + false + ], + [ + -15.19686508178711, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f647626b6c2effe8ca7302ca4613cf8129fd899143a80c08ab8f03289e257337", + "prompt_hash": "3868ab363e5f76c379bfc768131c3b6ffe214eaae3d6b1244836f37f8ce8a071", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 26, + "doc": { + "question": "Which base model was fine-tuned for experiments on coding problems?", + "choices": [ + "Gemini 1.0 Pro", + "Gemini 1.5 Flash", + "GPT-3.5", + "PaLM 2" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which base model was fine-tuned for experiments on coding problems?\nAnswer:", + " Gemini 1.0 Pro" + ], + [ + "Question:Which base model was fine-tuned for experiments on coding problems?\nAnswer:", + " Gemini 1.5 Flash" + ], + [ + "Question:Which base model was fine-tuned for experiments on coding problems?\nAnswer:", + " GPT-3.5" + ], + [ + "Question:Which base model was fine-tuned for experiments on coding problems?\nAnswer:", + " PaLM 2" + ] + ], + "resps": [ + [ + [ + -11.975496292114258, + false + ] + ], + [ + [ + -11.706640243530273, + false + ] + ], + [ + [ + -4.315305233001709, + false + ] + ], + [ + [ + -9.331165313720703, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.975496292114258, + false + ], + [ + -11.706640243530273, + false + ], + [ + -4.315305233001709, + false + ], + [ + -9.331165313720703, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b417fc051bd16e505060fabea6737d135a4be1170c992653e9da7ec5ec2a682b", + "prompt_hash": "07d264febd521ba3080badaf41a97ae084f718159df7ec066c962b7cc8de4507", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 27, + "doc": { + "question": "In Stage I of the training approach, what constraint is applied to the model’s first attempt?", + "choices": [ + "It must maximize accuracy directly", + "It is constrained to be close to the base model using a KL-divergence", + "It is discarded and only the second attempt is optimized", + "It is forced to match ground-truth answers" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:In Stage I of the training approach, what constraint is applied to the model’s first attempt?\nAnswer:", + " It must maximize accuracy directly" + ], + [ + "Question:In Stage I of the training approach, what constraint is applied to the model’s first attempt?\nAnswer:", + " It is constrained to be close to the base model using a KL-divergence" + ], + [ + "Question:In Stage I of the training approach, what constraint is applied to the model’s first attempt?\nAnswer:", + " It is discarded and only the second attempt is optimized" + ], + [ + "Question:In Stage I of the training approach, what constraint is applied to the model’s first attempt?\nAnswer:", + " It is forced to match ground-truth answers" + ] + ], + "resps": [ + [ + [ + -30.358684539794922, + false + ] + ], + [ + [ + -41.4800910949707, + false + ] + ], + [ + [ + -34.299049377441406, + false + ] + ], + [ + [ + -29.237268447875977, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.358684539794922, + false + ], + [ + -41.4800910949707, + false + ], + [ + -34.299049377441406, + false + ], + [ + -29.237268447875977, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b8c8119741d755b6d2421c475321038ba54dc019dc31248a9d83b976c63b9bed", + "prompt_hash": "7f31f11aac28d018d0a6f0802cd3bce644fe5839dd3195d20e3521da243f905c", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 28, + "doc": { + "question": "What hyperparameter value was used for the learning rate when training the method on the MATH dataset?", + "choices": [ + "5e-6", + "1e-5", + "1e-4", + "5e-5" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What hyperparameter value was used for the learning rate when training the method on the MATH dataset?\nAnswer:", + " 5e-6" + ], + [ + "Question:What hyperparameter value was used for the learning rate when training the method on the MATH dataset?\nAnswer:", + " 1e-5" + ], + [ + "Question:What hyperparameter value was used for the learning rate when training the method on the MATH dataset?\nAnswer:", + " 1e-4" + ], + [ + "Question:What hyperparameter value was used for the learning rate when training the method on the MATH dataset?\nAnswer:", + " 5e-5" + ] + ], + "resps": [ + [ + [ + -8.413949012756348, + false + ] + ], + [ + [ + -6.770974636077881, + false + ] + ], + [ + [ + -6.059501647949219, + false + ] + ], + [ + [ + -7.30908727645874, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.413949012756348, + false + ], + [ + -6.770974636077881, + false + ], + [ + -6.059501647949219, + false + ], + [ + -7.30908727645874, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "37abbc46b91aca2efa4127a035b6201304e388062228a9add526d282010ccb0f", + "prompt_hash": "ec4777367a3cfe0380696c09d06021224bebb270ab0c973687ae04831ce2411c", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 29, + "doc": { + "question": "What was the accuracy@t2 of the method on the HumanEval coding benchmark?", + "choices": [ + "52.5%", + "54.3%", + "56.7%", + "64.6%" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What was the accuracy@t2 of the method on the HumanEval coding benchmark?\nAnswer:", + " 52.5%" + ], + [ + "Question:What was the accuracy@t2 of the method on the HumanEval coding benchmark?\nAnswer:", + " 54.3%" + ], + [ + "Question:What was the accuracy@t2 of the method on the HumanEval coding benchmark?\nAnswer:", + " 56.7%" + ], + [ + "Question:What was the accuracy@t2 of the method on the HumanEval coding benchmark?\nAnswer:", + " 64.6%" + ] + ], + "resps": [ + [ + [ + -13.624635696411133, + false + ] + ], + [ + [ + -14.403515815734863, + false + ] + ], + [ + [ + -14.236377716064453, + false + ] + ], + [ + [ + -14.352609634399414, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.624635696411133, + false + ], + [ + -14.403515815734863, + false + ], + [ + -14.236377716064453, + false + ], + [ + -14.352609634399414, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d34e36f1883b08b4e303e1155a8c5e4b00af362834b850e5aa13120ee8e57bb5", + "prompt_hash": "9a16d45757cca256ff8ad1d89fa972d6db23ca8bbfb942bab51659565a081b8d", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 30, + "doc": { + "question": "What is the default number of GPUs per machine (U) used in the auto-parallelism algorithm?", + "choices": [ + "4", + "8", + "16", + "32" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the default number of GPUs per machine (U) used in the auto-parallelism algorithm?\nAnswer:", + " 4" + ], + [ + "Question:What is the default number of GPUs per machine (U) used in the auto-parallelism algorithm?\nAnswer:", + " 8" + ], + [ + "Question:What is the default number of GPUs per machine (U) used in the auto-parallelism algorithm?\nAnswer:", + " 16" + ], + [ + "Question:What is the default number of GPUs per machine (U) used in the auto-parallelism algorithm?\nAnswer:", + " 32" + ] + ], + "resps": [ + [ + [ + -3.4775350093841553, + false + ] + ], + [ + [ + -3.663968801498413, + false + ] + ], + [ + [ + -5.262546539306641, + false + ] + ], + [ + [ + -5.171002388000488, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.4775350093841553, + false + ], + [ + -3.663968801498413, + false + ], + [ + -5.262546539306641, + false + ], + [ + -5.171002388000488, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f612f5a0f7dce9c3cab1f9b7264350f535e58c3ff60f197a3ec58f4c3ad32fa7", + "prompt_hash": "67a6515518bc1d2f3d9ffa96539cd1f9a1c48faf0b6a78b868320df3db076bd1", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 31, + "doc": { + "question": "Which workload type is described as memory-bound in the auto-parallelism algorithm?", + "choices": [ + "Training", + "Inference", + "Generation", + "Backward pass" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which workload type is described as memory-bound in the auto-parallelism algorithm?\nAnswer:", + " Training" + ], + [ + "Question:Which workload type is described as memory-bound in the auto-parallelism algorithm?\nAnswer:", + " Inference" + ], + [ + "Question:Which workload type is described as memory-bound in the auto-parallelism algorithm?\nAnswer:", + " Generation" + ], + [ + "Question:Which workload type is described as memory-bound in the auto-parallelism algorithm?\nAnswer:", + " Backward pass" + ] + ], + "resps": [ + [ + [ + -7.63580322265625, + false + ] + ], + [ + [ + -7.492424011230469, + false + ] + ], + [ + [ + -9.588844299316406, + false + ] + ], + [ + [ + -12.772279739379883, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.63580322265625, + false + ], + [ + -7.492424011230469, + false + ], + [ + -9.588844299316406, + false + ], + [ + -12.772279739379883, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1d7916337d09f81d90fd24ab466d07533e48fc1e51f50d4c5f7b39ae5d0a4c43", + "prompt_hash": "b30c24ce196072cfa65f45b1664ae3504848f4675842802db20232d1095eabb1", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 32, + "doc": { + "question": "In RLHF, what operation does the actor model perform during Stage 1 (Generation)?", + "choices": [ + "Compute squared-error loss", + "Produce responses from prompts using auto-regressive generation", + "Evaluate divergence with the reference policy", + "Assign scalar rewards" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:In RLHF, what operation does the actor model perform during Stage 1 (Generation)?\nAnswer:", + " Compute squared-error loss" + ], + [ + "Question:In RLHF, what operation does the actor model perform during Stage 1 (Generation)?\nAnswer:", + " Produce responses from prompts using auto-regressive generation" + ], + [ + "Question:In RLHF, what operation does the actor model perform during Stage 1 (Generation)?\nAnswer:", + " Evaluate divergence with the reference policy" + ], + [ + "Question:In RLHF, what operation does the actor model perform during Stage 1 (Generation)?\nAnswer:", + " Assign scalar rewards" + ] + ], + "resps": [ + [ + [ + -30.3663330078125, + false + ] + ], + [ + [ + -35.097373962402344, + false + ] + ], + [ + [ + -35.44439697265625, + false + ] + ], + [ + [ + -18.711971282958984, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.3663330078125, + false + ], + [ + -35.097373962402344, + false + ], + [ + -35.44439697265625, + false + ], + [ + -18.711971282958984, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "25894f47d56bb4ce7d2bca2fa551786ae042ad88121e6af68f653ffb9c286461", + "prompt_hash": "0e3494099025b55e286623a7eb4a440a3500f83f64bc70e16f2b9f1aaad7aeac", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 33, + "doc": { + "question": "What parallelism strategy does ZeRO optimize?", + "choices": [ + "Pipeline parallelism", + "Tensor parallelism", + "Data parallelism", + "Model offloading" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What parallelism strategy does ZeRO optimize?\nAnswer:", + " Pipeline parallelism" + ], + [ + "Question:What parallelism strategy does ZeRO optimize?\nAnswer:", + " Tensor parallelism" + ], + [ + "Question:What parallelism strategy does ZeRO optimize?\nAnswer:", + " Data parallelism" + ], + [ + "Question:What parallelism strategy does ZeRO optimize?\nAnswer:", + " Model offloading" + ] + ], + "resps": [ + [ + [ + -10.252677917480469, + false + ] + ], + [ + [ + -8.313484191894531, + false + ] + ], + [ + [ + -8.054891586303711, + false + ] + ], + [ + [ + -20.16019058227539, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.252677917480469, + false + ], + [ + -8.313484191894531, + false + ], + [ + -8.054891586303711, + false + ], + [ + -20.16019058227539, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "14ad4867fd1a7cf0861481030428370605924da8a0b1bdfef6dcad706397b60e", + "prompt_hash": "eb1a8d1f7a2617e8d515800faefd480a234027205dbf2f3ce3be026c087b3bf8", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 34, + "doc": { + "question": "Which system colocates all models on the same set of devices for RLHF execution?", + "choices": [ + "OpenRLHF", + "NeMo-Aligner", + "DeepSpeed-Chat", + "HybridFlow" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which system colocates all models on the same set of devices for RLHF execution?\nAnswer:", + " OpenRLHF" + ], + [ + "Question:Which system colocates all models on the same set of devices for RLHF execution?\nAnswer:", + " NeMo-Aligner" + ], + [ + "Question:Which system colocates all models on the same set of devices for RLHF execution?\nAnswer:", + " DeepSpeed-Chat" + ], + [ + "Question:Which system colocates all models on the same set of devices for RLHF execution?\nAnswer:", + " HybridFlow" + ] + ], + "resps": [ + [ + [ + -15.416034698486328, + false + ] + ], + [ + [ + -25.671110153198242, + false + ] + ], + [ + [ + -13.71575927734375, + false + ] + ], + [ + [ + -17.46190643310547, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.416034698486328, + false + ], + [ + -25.671110153198242, + false + ], + [ + -13.71575927734375, + false + ], + [ + -17.46190643310547, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b22608c501506d3c57e6cb329f8d20566e5179d924ace4a6fc6c75413cbe34f4", + "prompt_hash": "2666f9983d03c6b0a75e6c5668e2439459f3f798b55909a9b1f0154d256c84eb", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 35, + "doc": { + "question": "What is the parameter size of the Helium text language model?", + "choices": [ + "3B", + "7B", + "13B", + "70B" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the parameter size of the Helium text language model?\nAnswer:", + " 3B" + ], + [ + "Question:What is the parameter size of the Helium text language model?\nAnswer:", + " 7B" + ], + [ + "Question:What is the parameter size of the Helium text language model?\nAnswer:", + " 13B" + ], + [ + "Question:What is the parameter size of the Helium text language model?\nAnswer:", + " 70B" + ] + ], + "resps": [ + [ + [ + -7.6070075035095215, + false + ] + ], + [ + [ + -6.104134559631348, + false + ] + ], + [ + [ + -7.87722110748291, + false + ] + ], + [ + [ + -8.18855094909668, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.6070075035095215, + false + ], + [ + -6.104134559631348, + false + ], + [ + -7.87722110748291, + false + ], + [ + -8.18855094909668, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "295a79dec253df87092a645dbb7aca11949e95802697013694a6fefe4ef6fb0f", + "prompt_hash": "88a6e62e96d34cc3fdbceaa93f9adfb56c8fc7cd6357984ce3d11b6f81de46b5", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 36, + "doc": { + "question": "Which neural audio codec is introduced to encode audio into discrete units?", + "choices": [ + "AudioLM", + "EnCodec", + "Mimi", + "SoundStream" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which neural audio codec is introduced to encode audio into discrete units?\nAnswer:", + " AudioLM" + ], + [ + "Question:Which neural audio codec is introduced to encode audio into discrete units?\nAnswer:", + " EnCodec" + ], + [ + "Question:Which neural audio codec is introduced to encode audio into discrete units?\nAnswer:", + " Mimi" + ], + [ + "Question:Which neural audio codec is introduced to encode audio into discrete units?\nAnswer:", + " SoundStream" + ] + ], + "resps": [ + [ + [ + -12.109827041625977, + false + ] + ], + [ + [ + -10.259915351867676, + false + ] + ], + [ + [ + -15.565141677856445, + false + ] + ], + [ + [ + -8.1206693649292, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.109827041625977, + false + ], + [ + -10.259915351867676, + false + ], + [ + -15.565141677856445, + false + ], + [ + -8.1206693649292, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "43046273a2d2929dfdb9fd277f085207ba600a9e1c95cc1d19a4e80097ad625f", + "prompt_hash": "16864ae4d828b91f371b7546d2dea0d21184075e15af6c995d72587a4c372ee6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 37, + "doc": { + "question": "What theoretical latency is reported for the real-time spoken dialogue system?", + "choices": [ + "80ms", + "120ms", + "160ms", + "230ms" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What theoretical latency is reported for the real-time spoken dialogue system?\nAnswer:", + " 80ms" + ], + [ + "Question:What theoretical latency is reported for the real-time spoken dialogue system?\nAnswer:", + " 120ms" + ], + [ + "Question:What theoretical latency is reported for the real-time spoken dialogue system?\nAnswer:", + " 160ms" + ], + [ + "Question:What theoretical latency is reported for the real-time spoken dialogue system?\nAnswer:", + " 230ms" + ] + ], + "resps": [ + [ + [ + -8.884298324584961, + false + ] + ], + [ + [ + -9.397584915161133, + false + ] + ], + [ + [ + -10.101129531860352, + false + ] + ], + [ + [ + -11.414660453796387, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.884298324584961, + false + ], + [ + -9.397584915161133, + false + ], + [ + -10.101129531860352, + false + ], + [ + -11.414660453796387, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b8850f8666028d7b8f55ed2a688a466e983ffda9a1ed39029e5a160dbf58d025", + "prompt_hash": "116a2f49d3531f67838d68a074089b4c452bc3eb568b8d8592054d40e1bbfaf3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 38, + "doc": { + "question": "What benchmark is adopted to evaluate safety under multiple categories?", + "choices": [ + "SUPERB", + "ALERT", + "MMLU", + "TriviaQA" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What benchmark is adopted to evaluate safety under multiple categories?\nAnswer:", + " SUPERB" + ], + [ + "Question:What benchmark is adopted to evaluate safety under multiple categories?\nAnswer:", + " ALERT" + ], + [ + "Question:What benchmark is adopted to evaluate safety under multiple categories?\nAnswer:", + " MMLU" + ], + [ + "Question:What benchmark is adopted to evaluate safety under multiple categories?\nAnswer:", + " TriviaQA" + ] + ], + "resps": [ + [ + [ + -13.114148139953613, + false + ] + ], + [ + [ + -17.356304168701172, + false + ] + ], + [ + [ + -10.324151039123535, + false + ] + ], + [ + [ + -13.553725242614746, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.114148139953613, + false + ], + [ + -17.356304168701172, + false + ], + [ + -10.324151039123535, + false + ], + [ + -13.553725242614746, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0ca451912f9e74fac5218c521b184aa7c06447916cec32fd5d97e838a9996769", + "prompt_hash": "a8a2c111e702c31f90d4990bce06e574afe60bd4a866b7c3a99fb16dd02deb62", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 39, + "doc": { + "question": "Which method is introduced to improve speech-to-speech generation by predicting text tokens before audio tokens?", + "choices": [ + "Chain-of-Modality", + "Parallel Generation", + "Interleaving", + "Residual Quantization" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which method is introduced to improve speech-to-speech generation by predicting text tokens before audio tokens?\nAnswer:", + " Chain-of-Modality" + ], + [ + "Question:Which method is introduced to improve speech-to-speech generation by predicting text tokens before audio tokens?\nAnswer:", + " Parallel Generation" + ], + [ + "Question:Which method is introduced to improve speech-to-speech generation by predicting text tokens before audio tokens?\nAnswer:", + " Interleaving" + ], + [ + "Question:Which method is introduced to improve speech-to-speech generation by predicting text tokens before audio tokens?\nAnswer:", + " Residual Quantization" + ] + ], + "resps": [ + [ + [ + -25.59917640686035, + false + ] + ], + [ + [ + -14.614076614379883, + false + ] + ], + [ + [ + -15.206499099731445, + false + ] + ], + [ + [ + -21.050691604614258, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.59917640686035, + false + ], + [ + -14.614076614379883, + false + ], + [ + -15.206499099731445, + false + ], + [ + -21.050691604614258, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a36384f4676d8978afe402e7f8d9b947f3a2a83db5673f38eedfc8e1f5b7a14b", + "prompt_hash": "9744f7a13faf7938951b7ec848c88d16aad947fc8ca2eaf7419692e7f13a5cbc", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 40, + "doc": { + "question": "How many videos are included in the \\\\DataName dataset?", + "choices": [ + "A. 178,510", + "B. 40,000", + "C. 900,000", + "D. 136 million" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:How many videos are included in the \\\\DataName dataset?\nAnswer:", + " A. 178,510" + ], + [ + "Question:How many videos are included in the \\\\DataName dataset?\nAnswer:", + " B. 40,000" + ], + [ + "Question:How many videos are included in the \\\\DataName dataset?\nAnswer:", + " C. 900,000" + ], + [ + "Question:How many videos are included in the \\\\DataName dataset?\nAnswer:", + " D. 136 million" + ] + ], + "resps": [ + [ + [ + -27.44942855834961, + false + ] + ], + [ + [ + -23.27755355834961, + false + ] + ], + [ + [ + -26.529640197753906, + false + ] + ], + [ + [ + -31.16094970703125, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.44942855834961, + false + ], + [ + -23.27755355834961, + false + ], + [ + -26.529640197753906, + false + ], + [ + -31.16094970703125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0fda4662f1ff0aa970514e5fdee88c432ac56b5b26263747812aa13614a7e8be", + "prompt_hash": "e05a2e8b0fc242c5cbb8630ceea6316bb72350bf467322d8cd62ea60708ab0e0", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 41, + "doc": { + "question": "What frame sampling rate is used in the \\\\DataName dataset?", + "choices": [ + "A. 0.008 fps", + "B. 0.15 fps", + "C. 1 fps", + "D. 10 fps" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What frame sampling rate is used in the \\\\DataName dataset?\nAnswer:", + " A. 0.008 fps" + ], + [ + "Question:What frame sampling rate is used in the \\\\DataName dataset?\nAnswer:", + " B. 0.15 fps" + ], + [ + "Question:What frame sampling rate is used in the \\\\DataName dataset?\nAnswer:", + " C. 1 fps" + ], + [ + "Question:What frame sampling rate is used in the \\\\DataName dataset?\nAnswer:", + " D. 10 fps" + ] + ], + "resps": [ + [ + [ + -27.462284088134766, + false + ] + ], + [ + [ + -26.013099670410156, + false + ] + ], + [ + [ + -18.49521255493164, + false + ] + ], + [ + [ + -19.096094131469727, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.462284088134766, + false + ], + [ + -26.013099670410156, + false + ], + [ + -18.49521255493164, + false + ], + [ + -19.096094131469727, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "08cf43a87aad451e03891dd8f52fd7dfae8127eb484b177c09ec9bd855170d55", + "prompt_hash": "9bd3fa5265e2205ee7e9a48d0e0fb0cb57a83361fe54e184b51355af1111300a", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 42, + "doc": { + "question": "How many question types are defined for generating question-answer pairs in \\\\DataName?", + "choices": [ + "A. 8", + "B. 10", + "C. 12", + "D. 16" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:How many question types are defined for generating question-answer pairs in \\\\DataName?\nAnswer:", + " A. 8" + ], + [ + "Question:How many question types are defined for generating question-answer pairs in \\\\DataName?\nAnswer:", + " B. 10" + ], + [ + "Question:How many question types are defined for generating question-answer pairs in \\\\DataName?\nAnswer:", + " C. 12" + ], + [ + "Question:How many question types are defined for generating question-answer pairs in \\\\DataName?\nAnswer:", + " D. 16" + ] + ], + "resps": [ + [ + [ + -15.037141799926758, + false + ] + ], + [ + [ + -18.426780700683594, + false + ] + ], + [ + [ + -19.0015869140625, + false + ] + ], + [ + [ + -20.956581115722656, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.037141799926758, + false + ], + [ + -18.426780700683594, + false + ], + [ + -19.0015869140625, + false + ], + [ + -20.956581115722656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "82e2376e9f9d4a079f3834fa42be02a1d9f23fa5015ce69c8c1a39ddfabf8ede", + "prompt_hash": "e3ee7893fc0206f90a0deae683266051c043888688f960c70aa659dcf549087b", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 43, + "doc": { + "question": "What is the total number of instruction-following samples in the \\\\DataName dataset?", + "choices": [ + "A. 1.3 million", + "B. 960,000", + "C. 178,000", + "D. 215,000" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the total number of instruction-following samples in the \\\\DataName dataset?\nAnswer:", + " A. 1.3 million" + ], + [ + "Question:What is the total number of instruction-following samples in the \\\\DataName dataset?\nAnswer:", + " B. 960,000" + ], + [ + "Question:What is the total number of instruction-following samples in the \\\\DataName dataset?\nAnswer:", + " C. 178,000" + ], + [ + "Question:What is the total number of instruction-following samples in the \\\\DataName dataset?\nAnswer:", + " D. 215,000" + ] + ], + "resps": [ + [ + [ + -20.036428451538086, + false + ] + ], + [ + [ + -28.441104888916016, + false + ] + ], + [ + [ + -27.614681243896484, + false + ] + ], + [ + [ + -28.703224182128906, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.036428451538086, + false + ], + [ + -28.441104888916016, + false + ], + [ + -27.614681243896484, + false + ], + [ + -28.703224182128906, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f128563c09c7fddc52e38a1b51567a61be91ea0faac794d2e8c55bc39075ccda", + "prompt_hash": "7f89cba8f9e1fe131fbecce3bcebe2aec104d7f3f37be4c619240b74381f8fc9", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 44, + "doc": { + "question": "Which dataset is reported to have an average sampling rate of 0.15 fps, sometimes sampling only 2 frames from a 30-second video?", + "choices": [ + "A. ShareGPT4Video", + "B. LLaVA-Hound", + "C. HowTo100M", + "D. ACAV-100M" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which dataset is reported to have an average sampling rate of 0.15 fps, sometimes sampling only 2 frames from a 30-second video?\nAnswer:", + " A. ShareGPT4Video" + ], + [ + "Question:Which dataset is reported to have an average sampling rate of 0.15 fps, sometimes sampling only 2 frames from a 30-second video?\nAnswer:", + " B. LLaVA-Hound" + ], + [ + "Question:Which dataset is reported to have an average sampling rate of 0.15 fps, sometimes sampling only 2 frames from a 30-second video?\nAnswer:", + " C. HowTo100M" + ], + [ + "Question:Which dataset is reported to have an average sampling rate of 0.15 fps, sometimes sampling only 2 frames from a 30-second video?\nAnswer:", + " D. ACAV-100M" + ] + ], + "resps": [ + [ + [ + -25.649900436401367, + false + ] + ], + [ + [ + -42.51259231567383, + false + ] + ], + [ + [ + -24.319427490234375, + false + ] + ], + [ + [ + -43.85147476196289, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.649900436401367, + false + ], + [ + -42.51259231567383, + false + ], + [ + -24.319427490234375, + false + ], + [ + -43.85147476196289, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "91ef16f651f4594ab7a458e54ea4022facbb0e5c3631264f316b197bae204fdd", + "prompt_hash": "ea07f7b02166866b6f0032e0c9168d65d23f09c6eed5c894629afaea30d2b9b7", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 45, + "doc": { + "question": "How many parameters does the 0.6B version of the model contain?", + "choices": [ + "1152M", + "590M", + "2240M", + "1604M" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many parameters does the 0.6B version of the model contain?\nAnswer:", + " 1152M" + ], + [ + "Question:How many parameters does the 0.6B version of the model contain?\nAnswer:", + " 590M" + ], + [ + "Question:How many parameters does the 0.6B version of the model contain?\nAnswer:", + " 2240M" + ], + [ + "Question:How many parameters does the 0.6B version of the model contain?\nAnswer:", + " 1604M" + ] + ], + "resps": [ + [ + [ + -14.291555404663086, + false + ] + ], + [ + [ + -10.522879600524902, + false + ] + ], + [ + [ + -16.00899887084961, + false + ] + ], + [ + [ + -16.75801658630371, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.291555404663086, + false + ], + [ + -10.522879600524902, + false + ], + [ + -16.00899887084961, + false + ], + [ + -16.75801658630371, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a4761b93b20496334d111187f778f76dfea66671563930bea6e0b3980c216e20", + "prompt_hash": "14882c4529bd7fe9c8ee9d792e71cfb63a0c12ca8a2bcec7afb3bfdba8d5d00a", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 46, + "doc": { + "question": "Which dataset contains 30K images from Midjourney and is used for evaluating FID and Clip Score?", + "choices": [ + "ImageNet-1K", + "MSCOCO", + "MJHQ-30K", + "DPG-Bench" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which dataset contains 30K images from Midjourney and is used for evaluating FID and Clip Score?\nAnswer:", + " ImageNet-1K" + ], + [ + "Question:Which dataset contains 30K images from Midjourney and is used for evaluating FID and Clip Score?\nAnswer:", + " MSCOCO" + ], + [ + "Question:Which dataset contains 30K images from Midjourney and is used for evaluating FID and Clip Score?\nAnswer:", + " MJHQ-30K" + ], + [ + "Question:Which dataset contains 30K images from Midjourney and is used for evaluating FID and Clip Score?\nAnswer:", + " DPG-Bench" + ] + ], + "resps": [ + [ + [ + -7.996962070465088, + false + ] + ], + [ + [ + -9.08751106262207, + false + ] + ], + [ + [ + -22.040189743041992, + false + ] + ], + [ + [ + -22.54903793334961, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.996962070465088, + false + ], + [ + -9.08751106262207, + false + ], + [ + -22.040189743041992, + false + ], + [ + -22.54903793334961, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "47b0bbebc2ea01a95a6021dfa70c6f6543afdbccfcee49014c890c3b647ddbb9", + "prompt_hash": "fc97e1336510e8fe920082f9a2bc83ac3e73e2c24d4c22c872b4669e00d5f1e3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 47, + "doc": { + "question": "What is the depth (number of layers) of the 1.6B version of the model?", + "choices": [ + "28", + "36", + "20", + "32" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the depth (number of layers) of the 1.6B version of the model?\nAnswer:", + " 28" + ], + [ + "Question:What is the depth (number of layers) of the 1.6B version of the model?\nAnswer:", + " 36" + ], + [ + "Question:What is the depth (number of layers) of the 1.6B version of the model?\nAnswer:", + " 20" + ], + [ + "Question:What is the depth (number of layers) of the 1.6B version of the model?\nAnswer:", + " 32" + ] + ], + "resps": [ + [ + [ + -7.322317123413086, + false + ] + ], + [ + [ + -6.984606742858887, + false + ] + ], + [ + [ + -5.454815864562988, + false + ] + ], + [ + [ + -5.265756607055664, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.322317123413086, + false + ], + [ + -6.984606742858887, + false + ], + [ + -5.454815864562988, + false + ], + [ + -5.265756607055664, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d283b2c2f551cd9a4491d5d44ffb10ddc54033ecfdcbe377a306d7ce6d105f13", + "prompt_hash": "f9e760f45616b073e48293f49d2ea32c9244abe3751176a0834a1a204b0a5391", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 48, + "doc": { + "question": "Which evaluation metric assesses human preference performance and includes 100 prompts?", + "choices": [ + "GenEval", + "FID", + "ImageReward", + "Clip Score" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which evaluation metric assesses human preference performance and includes 100 prompts?\nAnswer:", + " GenEval" + ], + [ + "Question:Which evaluation metric assesses human preference performance and includes 100 prompts?\nAnswer:", + " FID" + ], + [ + "Question:Which evaluation metric assesses human preference performance and includes 100 prompts?\nAnswer:", + " ImageReward" + ], + [ + "Question:Which evaluation metric assesses human preference performance and includes 100 prompts?\nAnswer:", + " Clip Score" + ] + ], + "resps": [ + [ + [ + -12.850953102111816, + false + ] + ], + [ + [ + -7.501878261566162, + false + ] + ], + [ + [ + -11.249866485595703, + false + ] + ], + [ + [ + -14.445904731750488, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.850953102111816, + false + ], + [ + -7.501878261566162, + false + ], + [ + -11.249866485595703, + false + ], + [ + -14.445904731750488, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b1c842171a30cabea8a0daa58584624e1982962bfeabe9757318a71179c49389", + "prompt_hash": "75cc4f9f1859d330acf6b765f1cb11c2974f5c217b2ea2829083badbb4a07185", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 49, + "doc": { + "question": "When replacing full attention with linear attention using AE-F8C4P2, what happens to latency?", + "choices": [ + "Increases from 1931ms to 2250ms", + "Reduces from 2250ms to 1931ms", + "Remains constant at 2250ms", + "Remains constant at 1931ms" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:When replacing full attention with linear attention using AE-F8C4P2, what happens to latency?\nAnswer:", + " Increases from 1931ms to 2250ms" + ], + [ + "Question:When replacing full attention with linear attention using AE-F8C4P2, what happens to latency?\nAnswer:", + " Reduces from 2250ms to 1931ms" + ], + [ + "Question:When replacing full attention with linear attention using AE-F8C4P2, what happens to latency?\nAnswer:", + " Remains constant at 2250ms" + ], + [ + "Question:When replacing full attention with linear attention using AE-F8C4P2, what happens to latency?\nAnswer:", + " Remains constant at 1931ms" + ] + ], + "resps": [ + [ + [ + -31.620426177978516, + false + ] + ], + [ + [ + -32.93737030029297, + false + ] + ], + [ + [ + -28.212650299072266, + false + ] + ], + [ + [ + -29.66581916809082, + false + ] + ] + ], + "filtered_resps": [ + [ + -31.620426177978516, + false + ], + [ + -32.93737030029297, + false + ], + [ + -28.212650299072266, + false + ], + [ + -29.66581916809082, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "820644bfcd2dc3aa158a97ab3742cb3b502046a277e8e05c499e6791d335730b", + "prompt_hash": "52f48f53d1dea326532a265611d26026a39424eef577534ac9a7d91b62926f97", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 50, + "doc": { + "question": "What is the main architectural simplification that makes the method 27% faster than LocoTrack?", + "choices": [ + "The use of 3D convolution for feature aggregation", + "A simplified MLP to process 4D correlation features", + "The removal of Fourier encoding in track embeddings", + "The integration of optical flow refinement modules" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main architectural simplification that makes the method 27% faster than LocoTrack?\nAnswer:", + " The use of 3D convolution for feature aggregation" + ], + [ + "Question:What is the main architectural simplification that makes the method 27% faster than LocoTrack?\nAnswer:", + " A simplified MLP to process 4D correlation features" + ], + [ + "Question:What is the main architectural simplification that makes the method 27% faster than LocoTrack?\nAnswer:", + " The removal of Fourier encoding in track embeddings" + ], + [ + "Question:What is the main architectural simplification that makes the method 27% faster than LocoTrack?\nAnswer:", + " The integration of optical flow refinement modules" + ] + ], + "resps": [ + [ + [ + -27.528322219848633, + false + ] + ], + [ + [ + -49.46475601196289, + false + ] + ], + [ + [ + -44.16676330566406, + false + ] + ], + [ + [ + -33.746856689453125, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.528322219848633, + false + ], + [ + -49.46475601196289, + false + ], + [ + -44.16676330566406, + false + ], + [ + -33.746856689453125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "85e26dc0e27fe3b1dfcc11f0663c09675a0372c7d6e26e9e33790b5f9e0e702e", + "prompt_hash": "f72f4336249fca531698c3403b55235f59beb77b5ba270ac8e97493c4efa36c0", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 51, + "doc": { + "question": "How many real videos does the proposed method use to outperform BootsTAPIR?", + "choices": [ + "15 billion", + "15 million", + "15 thousand", + "1.5 million" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many real videos does the proposed method use to outperform BootsTAPIR?\nAnswer:", + " 15 billion" + ], + [ + "Question:How many real videos does the proposed method use to outperform BootsTAPIR?\nAnswer:", + " 15 million" + ], + [ + "Question:How many real videos does the proposed method use to outperform BootsTAPIR?\nAnswer:", + " 15 thousand" + ], + [ + "Question:How many real videos does the proposed method use to outperform BootsTAPIR?\nAnswer:", + " 1.5 million" + ] + ], + "resps": [ + [ + [ + -15.632142066955566, + false + ] + ], + [ + [ + -12.80514907836914, + false + ] + ], + [ + [ + -14.469340324401855, + false + ] + ], + [ + [ + -11.69805908203125, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.632142066955566, + false + ], + [ + -12.80514907836914, + false + ], + [ + -14.469340324401855, + false + ], + [ + -11.69805908203125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "30dca43e6d67cfa022d80dec0a71ed2d08cc4128cb249ae5a5b2ba1c2ed39b69", + "prompt_hash": "d236ad1ed85bba8b379b7cdcfb1b7270884351d9aee04d2796646686f9e6d2bc", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 52, + "doc": { + "question": "Which teacher models are used to generate pseudo-labels in the student training protocol?", + "choices": [ + "RAFT, PIPs++, and SuperPoint", + "TAPIR, CoTracker, method online and method offline", + "MAST, TAPTR, and LightGlue", + "VGGSfM, PointOdyssey, and DISK" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which teacher models are used to generate pseudo-labels in the student training protocol?\nAnswer:", + " RAFT, PIPs++, and SuperPoint" + ], + [ + "Question:Which teacher models are used to generate pseudo-labels in the student training protocol?\nAnswer:", + " TAPIR, CoTracker, method online and method offline" + ], + [ + "Question:Which teacher models are used to generate pseudo-labels in the student training protocol?\nAnswer:", + " MAST, TAPTR, and LightGlue" + ], + [ + "Question:Which teacher models are used to generate pseudo-labels in the student training protocol?\nAnswer:", + " VGGSfM, PointOdyssey, and DISK" + ] + ], + "resps": [ + [ + [ + -52.90093231201172, + false + ] + ], + [ + [ + -59.762210845947266, + false + ] + ], + [ + [ + -49.665283203125, + false + ] + ], + [ + [ + -56.55097961425781, + false + ] + ] + ], + "filtered_resps": [ + [ + -52.90093231201172, + false + ], + [ + -59.762210845947266, + false + ], + [ + -49.665283203125, + false + ], + [ + -56.55097961425781, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e7096069f795a6b31354d56171bab7cb64e9729dc77a21a7023c3f21fbe830b0", + "prompt_hash": "4c8a81ecb2d25b0ecf479c4c24bc416df2057eb1ddeb420626479cc5a847f079", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 53, + "doc": { + "question": "What type of videos are primarily used in the large-scale dataset for training with pseudo-labels?", + "choices": [ + "Synthetic videos of static indoor scenes", + "Videos of computer-generated natural phenomena", + "Internet-like videos of humans and animals", + "Labeled surveillance videos from street cameras" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What type of videos are primarily used in the large-scale dataset for training with pseudo-labels?\nAnswer:", + " Synthetic videos of static indoor scenes" + ], + [ + "Question:What type of videos are primarily used in the large-scale dataset for training with pseudo-labels?\nAnswer:", + " Videos of computer-generated natural phenomena" + ], + [ + "Question:What type of videos are primarily used in the large-scale dataset for training with pseudo-labels?\nAnswer:", + " Internet-like videos of humans and animals" + ], + [ + "Question:What type of videos are primarily used in the large-scale dataset for training with pseudo-labels?\nAnswer:", + " Labeled surveillance videos from street cameras" + ] + ], + "resps": [ + [ + [ + -27.919666290283203, + false + ] + ], + [ + [ + -31.976137161254883, + false + ] + ], + [ + [ + -31.44933319091797, + false + ] + ], + [ + [ + -32.28932571411133, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.919666290283203, + false + ], + [ + -31.976137161254883, + false + ], + [ + -31.44933319091797, + false + ], + [ + -32.28932571411133, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "77d962a9d88756c197df6db26cb0ed6c97ac99c1f27d634b07cff078c08e3a2d", + "prompt_hash": "e272316b8094350cdcb23a77e849572f4541ca0a2b1dd54b397b2fff3b54db05", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 54, + "doc": { + "question": "Which metric measures both tracking and occlusion prediction accuracy together?", + "choices": [ + "Occlusion Accuracy (OA)", + "Average Jaccard (AJ)", + "Average Displacement for Visible Points", + "Confidence Threshold Score" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which metric measures both tracking and occlusion prediction accuracy together?\nAnswer:", + " Occlusion Accuracy (OA)" + ], + [ + "Question:Which metric measures both tracking and occlusion prediction accuracy together?\nAnswer:", + " Average Jaccard (AJ)" + ], + [ + "Question:Which metric measures both tracking and occlusion prediction accuracy together?\nAnswer:", + " Average Displacement for Visible Points" + ], + [ + "Question:Which metric measures both tracking and occlusion prediction accuracy together?\nAnswer:", + " Confidence Threshold Score" + ] + ], + "resps": [ + [ + [ + -10.50698471069336, + false + ] + ], + [ + [ + -15.57904052734375, + false + ] + ], + [ + [ + -29.477798461914062, + false + ] + ], + [ + [ + -17.279052734375, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.50698471069336, + false + ], + [ + -15.57904052734375, + false + ], + [ + -29.477798461914062, + false + ], + [ + -17.279052734375, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5b3399c26f3ef69cd385333b0e5bdc026bd5a64550372075f4e5fffebb9cb69d", + "prompt_hash": "9ced7ba0deba30498839b18d23a59085f5cf39c21d548e34267f77208b17eade", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 55, + "doc": { + "question": "What is the resolution of images used in the main visualization examples of Janus?", + "choices": [ + "256 × 256", + "384 × 384", + "512 × 512", + "1024 × 1024" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the resolution of images used in the main visualization examples of Janus?\nAnswer:", + " 256 × 256" + ], + [ + "Question:What is the resolution of images used in the main visualization examples of Janus?\nAnswer:", + " 384 × 384" + ], + [ + "Question:What is the resolution of images used in the main visualization examples of Janus?\nAnswer:", + " 512 × 512" + ], + [ + "Question:What is the resolution of images used in the main visualization examples of Janus?\nAnswer:", + " 1024 × 1024" + ] + ], + "resps": [ + [ + [ + -7.799862861633301, + false + ] + ], + [ + [ + -10.63082504272461, + false + ] + ], + [ + [ + -8.216829299926758, + false + ] + ], + [ + [ + -8.296860694885254, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.799862861633301, + false + ], + [ + -10.63082504272461, + false + ], + [ + -8.216829299926758, + false + ], + [ + -8.296860694885254, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "08275decf2faf58dfa291473ce53e1cee174d18aa5857c75541400d82863e845", + "prompt_hash": "4c3b81178c9839409766d7953ec4c181088f4afbf8459303d49eaec3b2600f82", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 56, + "doc": { + "question": "Which model does Janus use as its base large language model (LLM)?", + "choices": [ + "Qwen-VL", + "GPT-Neo", + "DeepSeek-LLM", + "LLaMA 2" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model does Janus use as its base large language model (LLM)?\nAnswer:", + " Qwen-VL" + ], + [ + "Question:Which model does Janus use as its base large language model (LLM)?\nAnswer:", + " GPT-Neo" + ], + [ + "Question:Which model does Janus use as its base large language model (LLM)?\nAnswer:", + " DeepSeek-LLM" + ], + [ + "Question:Which model does Janus use as its base large language model (LLM)?\nAnswer:", + " LLaMA 2" + ] + ], + "resps": [ + [ + [ + -9.289318084716797, + false + ] + ], + [ + [ + -8.6356201171875, + false + ] + ], + [ + [ + -12.277383804321289, + false + ] + ], + [ + [ + -10.94096851348877, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.289318084716797, + false + ], + [ + -8.6356201171875, + false + ], + [ + -12.277383804321289, + false + ], + [ + -10.94096851348877, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "09a112fdcf77b0745cd5d3b4f4d0133d042b7a76e9a26c7bb6113d69dbab6d4c", + "prompt_hash": "b6d3b003f088781b3137314937470a4a50e04c438b85653d0a77596b8529a21e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 57, + "doc": { + "question": "What is the FID score achieved by Janus on the MSCOCO-30K benchmark?", + "choices": [ + "10.39", + "9.24", + "8.53", + "7.32" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the FID score achieved by Janus on the MSCOCO-30K benchmark?\nAnswer:", + " 10.39" + ], + [ + "Question:What is the FID score achieved by Janus on the MSCOCO-30K benchmark?\nAnswer:", + " 9.24" + ], + [ + "Question:What is the FID score achieved by Janus on the MSCOCO-30K benchmark?\nAnswer:", + " 8.53" + ], + [ + "Question:What is the FID score achieved by Janus on the MSCOCO-30K benchmark?\nAnswer:", + " 7.32" + ] + ], + "resps": [ + [ + [ + -11.242629051208496, + false + ] + ], + [ + [ + -11.04965591430664, + false + ] + ], + [ + [ + -11.016393661499023, + false + ] + ], + [ + [ + -11.18939208984375, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.242629051208496, + false + ], + [ + -11.04965591430664, + false + ], + [ + -11.016393661499023, + false + ], + [ + -11.18939208984375, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6a08db5d8ac2bb1254553eb0da0ea2d3691559010c970be56df0c44c97c78ab4", + "prompt_hash": "a8e79602f14e078b6a6b23461115e4993d68ac4f95fe9896c64d905c2d2337be", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 58, + "doc": { + "question": "How many attention heads does the semantic decoder in the ablation study have?", + "choices": [ + "8", + "10", + "12", + "16" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many attention heads does the semantic decoder in the ablation study have?\nAnswer:", + " 8" + ], + [ + "Question:How many attention heads does the semantic decoder in the ablation study have?\nAnswer:", + " 10" + ], + [ + "Question:How many attention heads does the semantic decoder in the ablation study have?\nAnswer:", + " 12" + ], + [ + "Question:How many attention heads does the semantic decoder in the ablation study have?\nAnswer:", + " 16" + ] + ], + "resps": [ + [ + [ + -3.614625930786133, + false + ] + ], + [ + [ + -7.19932746887207, + false + ] + ], + [ + [ + -5.028623580932617, + false + ] + ], + [ + [ + -5.395282745361328, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.614625930786133, + false + ], + [ + -7.19932746887207, + false + ], + [ + -5.028623580932617, + false + ], + [ + -5.395282745361328, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "97eddd54ac52372dd9747fe6ede0c383a8a2aea0a608d85cdd1c94fa817d7c6d", + "prompt_hash": "a35a7ec64f3da7478f8048930b42444873cc40ec2b9963fea585728aaee40284", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 59, + "doc": { + "question": "How many nodes and GPUs were used in the full training process of Janus?", + "choices": [ + "8 nodes with 4 GPUs each", + "16 nodes with 8 GPUs each", + "32 nodes with 2 GPUs each", + "4 nodes with 16 GPUs each" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many nodes and GPUs were used in the full training process of Janus?\nAnswer:", + " 8 nodes with 4 GPUs each" + ], + [ + "Question:How many nodes and GPUs were used in the full training process of Janus?\nAnswer:", + " 16 nodes with 8 GPUs each" + ], + [ + "Question:How many nodes and GPUs were used in the full training process of Janus?\nAnswer:", + " 32 nodes with 2 GPUs each" + ], + [ + "Question:How many nodes and GPUs were used in the full training process of Janus?\nAnswer:", + " 4 nodes with 16 GPUs each" + ] + ], + "resps": [ + [ + [ + -16.610118865966797, + false + ] + ], + [ + [ + -17.693458557128906, + false + ] + ], + [ + [ + -16.693805694580078, + false + ] + ], + [ + [ + -17.05245590209961, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.610118865966797, + false + ], + [ + -17.693458557128906, + false + ], + [ + -16.693805694580078, + false + ], + [ + -17.05245590209961, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f674a2e6a1bb759c66f647721dda1ebcea5048921cce40120a2cf90e4695798f", + "prompt_hash": "a608a708c7e2c33240e761f0d73ad14f178f9cbf67b494a220b664d8f16ec08c", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 60, + "doc": { + "question": "What does the C3k2 block in YOLOv11 replace from previous versions?", + "choices": [ + "CSP block", + "CBS block", + "C2f block", + "C3 block" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does the C3k2 block in YOLOv11 replace from previous versions?\nAnswer:", + " CSP block" + ], + [ + "Question:What does the C3k2 block in YOLOv11 replace from previous versions?\nAnswer:", + " CBS block" + ], + [ + "Question:What does the C3k2 block in YOLOv11 replace from previous versions?\nAnswer:", + " C2f block" + ], + [ + "Question:What does the C3k2 block in YOLOv11 replace from previous versions?\nAnswer:", + " C3 block" + ] + ], + "resps": [ + [ + [ + -10.633159637451172, + false + ] + ], + [ + [ + -16.662315368652344, + false + ] + ], + [ + [ + -12.172492980957031, + false + ] + ], + [ + [ + -7.113786220550537, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.633159637451172, + false + ], + [ + -16.662315368652344, + false + ], + [ + -12.172492980957031, + false + ], + [ + -7.113786220550537, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5076961b809c42258c80aa58f12d0097ed42b99ed43ef50190f915a71c4b8416", + "prompt_hash": "b7135c8dbb3db09bcc7e1dd949c1bbdce2eb9be903a99350a8bc227231ed6b0b", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 61, + "doc": { + "question": "Which YOLOv11 component is responsible for generating the final object detection predictions?", + "choices": [ + "Backbone", + "Neck", + "Head", + "C2PSA" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which YOLOv11 component is responsible for generating the final object detection predictions?\nAnswer:", + " Backbone" + ], + [ + "Question:Which YOLOv11 component is responsible for generating the final object detection predictions?\nAnswer:", + " Neck" + ], + [ + "Question:Which YOLOv11 component is responsible for generating the final object detection predictions?\nAnswer:", + " Head" + ], + [ + "Question:Which YOLOv11 component is responsible for generating the final object detection predictions?\nAnswer:", + " C2PSA" + ] + ], + "resps": [ + [ + [ + -8.064074516296387, + false + ] + ], + [ + [ + -7.961186408996582, + false + ] + ], + [ + [ + -4.5736188888549805, + false + ] + ], + [ + [ + -28.57439422607422, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.064074516296387, + false + ], + [ + -7.961186408996582, + false + ], + [ + -4.5736188888549805, + false + ], + [ + -28.57439422607422, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "74c4acd48e33b31d2db4dcce90783d71007a901dd9550689dd6404b689183fa6", + "prompt_hash": "d561ea0b79eb2ffdb1d0fbbfa7120a55d977134eb76585649b70a53cffbffd61", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 62, + "doc": { + "question": "Which YOLO version introduced consistent dual assignments for NMS-free training?", + "choices": [ + "YOLOv8", + "YOLOv9", + "YOLOv10", + "YOLOv11" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which YOLO version introduced consistent dual assignments for NMS-free training?\nAnswer:", + " YOLOv8" + ], + [ + "Question:Which YOLO version introduced consistent dual assignments for NMS-free training?\nAnswer:", + " YOLOv9" + ], + [ + "Question:Which YOLO version introduced consistent dual assignments for NMS-free training?\nAnswer:", + " YOLOv10" + ], + [ + "Question:Which YOLO version introduced consistent dual assignments for NMS-free training?\nAnswer:", + " YOLOv11" + ] + ], + "resps": [ + [ + [ + -6.006664276123047, + false + ] + ], + [ + [ + -7.727287292480469, + false + ] + ], + [ + [ + -10.3282470703125, + false + ] + ], + [ + [ + -12.71410846710205, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.006664276123047, + false + ], + [ + -7.727287292480469, + false + ], + [ + -10.3282470703125, + false + ], + [ + -12.71410846710205, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "782bc04b40fdf646e23da0df3c4d456db2841d31c48d7f728473473a9d4eefd3", + "prompt_hash": "88b8f0a208713be8321190755c8a091359e89d20385829a8866f058617bd1246", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 63, + "doc": { + "question": "What task is the YOLOv11-obb variant specifically designed for?", + "choices": [ + "Pose Estimation", + "Oriented Object Detection", + "Instance Segmentation", + "Image Classification" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What task is the YOLOv11-obb variant specifically designed for?\nAnswer:", + " Pose Estimation" + ], + [ + "Question:What task is the YOLOv11-obb variant specifically designed for?\nAnswer:", + " Oriented Object Detection" + ], + [ + "Question:What task is the YOLOv11-obb variant specifically designed for?\nAnswer:", + " Instance Segmentation" + ], + [ + "Question:What task is the YOLOv11-obb variant specifically designed for?\nAnswer:", + " Image Classification" + ] + ], + "resps": [ + [ + [ + -10.559854507446289, + false + ] + ], + [ + [ + -12.302038192749023, + false + ] + ], + [ + [ + -9.71845531463623, + false + ] + ], + [ + [ + -10.004639625549316, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.559854507446289, + false + ], + [ + -12.302038192749023, + false + ], + [ + -9.71845531463623, + false + ], + [ + -10.004639625549316, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2362c80e349c52b80c72e0f5e575c2435ec9f5680e2fb7a386a2a533cc52b3ad", + "prompt_hash": "13ec62143491a215a0f7c97502a13fb9a5363536bd5328eae8b7d42f4bfb7250", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 64, + "doc": { + "question": "How much fewer parameters does YOLOv11m use compared to YOLOv8m while achieving superior mAP?", + "choices": [ + "10%", + "18%", + "22%", + "30%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How much fewer parameters does YOLOv11m use compared to YOLOv8m while achieving superior mAP?\nAnswer:", + " 10%" + ], + [ + "Question:How much fewer parameters does YOLOv11m use compared to YOLOv8m while achieving superior mAP?\nAnswer:", + " 18%" + ], + [ + "Question:How much fewer parameters does YOLOv11m use compared to YOLOv8m while achieving superior mAP?\nAnswer:", + " 22%" + ], + [ + "Question:How much fewer parameters does YOLOv11m use compared to YOLOv8m while achieving superior mAP?\nAnswer:", + " 30%" + ] + ], + "resps": [ + [ + [ + -13.356922149658203, + false + ] + ], + [ + [ + -13.9075927734375, + false + ] + ], + [ + [ + -14.149547576904297, + false + ] + ], + [ + [ + -13.93923568725586, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.356922149658203, + false + ], + [ + -13.9075927734375, + false + ], + [ + -14.149547576904297, + false + ], + [ + -13.93923568725586, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ed2455f30c263a79e1b14b6460eb0dc88c2f48ba69a9945217c4cd71079266ff", + "prompt_hash": "05f7d4beae3f3e039c97fe7b7862d7965444171656fd42e43210013ab83c0b5e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 65, + "doc": { + "question": "Which stage in the predictive model hierarchy is characterized by the ability to generate actionable videos?", + "choices": [ + "A. S0", + "B. S1", + "C. S2", + "D. S3" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which stage in the predictive model hierarchy is characterized by the ability to generate actionable videos?\nAnswer:", + " A. S0" + ], + [ + "Question:Which stage in the predictive model hierarchy is characterized by the ability to generate actionable videos?\nAnswer:", + " B. S1" + ], + [ + "Question:Which stage in the predictive model hierarchy is characterized by the ability to generate actionable videos?\nAnswer:", + " C. S2" + ], + [ + "Question:Which stage in the predictive model hierarchy is characterized by the ability to generate actionable videos?\nAnswer:", + " D. S3" + ] + ], + "resps": [ + [ + [ + -19.10530662536621, + false + ] + ], + [ + [ + -19.845218658447266, + false + ] + ], + [ + [ + -21.168285369873047, + false + ] + ], + [ + [ + -21.69548797607422, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.10530662536621, + false + ], + [ + -19.845218658447266, + false + ], + [ + -21.168285369873047, + false + ], + [ + -21.69548797607422, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1e0d10a547609c9ddd142bfcb86a301eb7d6bdfa579ede13d2b137378097532c", + "prompt_hash": "5d1e1c6ae3e6a61a17ea040689e1fe031915e198af7a4ab2d9e77076669f6ee4", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 66, + "doc": { + "question": "What simulation environment is used for the evaluation of predictive models in the \\u2018\\abarm\\u2019 scenario?", + "choices": [ + "A. CARLA", + "B. Calvin", + "C. Habitat", + "D. MineRL" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What simulation environment is used for the evaluation of predictive models in the \\u2018\\abarm\\u2019 scenario?\nAnswer:", + " A. CARLA" + ], + [ + "Question:What simulation environment is used for the evaluation of predictive models in the \\u2018\\abarm\\u2019 scenario?\nAnswer:", + " B. Calvin" + ], + [ + "Question:What simulation environment is used for the evaluation of predictive models in the \\u2018\\abarm\\u2019 scenario?\nAnswer:", + " C. Habitat" + ], + [ + "Question:What simulation environment is used for the evaluation of predictive models in the \\u2018\\abarm\\u2019 scenario?\nAnswer:", + " D. MineRL" + ] + ], + "resps": [ + [ + [ + -18.005578994750977, + false + ] + ], + [ + [ + -24.09957504272461, + false + ] + ], + [ + [ + -20.01169204711914, + false + ] + ], + [ + [ + -31.064380645751953, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.005578994750977, + false + ], + [ + -24.09957504272461, + false + ], + [ + -20.01169204711914, + false + ], + [ + -31.064380645751953, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f67a8549dcd2b8979f1de1e9580290429e1de35d629105905b3335e18f414831", + "prompt_hash": "ad95fec7874f6e14104bc5e0b9893ff429528f30f46d222389a546536e555e70", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 67, + "doc": { + "question": "Which benchmark uses 'Human Judgement' as its evaluation strategy and operates at Stage S0?", + "choices": [ + "A. EvalCrafter", + "B. VBench", + "C. AgentBench", + "D. LEGO" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which benchmark uses 'Human Judgement' as its evaluation strategy and operates at Stage S0?\nAnswer:", + " A. EvalCrafter" + ], + [ + "Question:Which benchmark uses 'Human Judgement' as its evaluation strategy and operates at Stage S0?\nAnswer:", + " B. VBench" + ], + [ + "Question:Which benchmark uses 'Human Judgement' as its evaluation strategy and operates at Stage S0?\nAnswer:", + " C. AgentBench" + ], + [ + "Question:Which benchmark uses 'Human Judgement' as its evaluation strategy and operates at Stage S0?\nAnswer:", + " D. LEGO" + ] + ], + "resps": [ + [ + [ + -24.232826232910156, + false + ] + ], + [ + [ + -22.290088653564453, + false + ] + ], + [ + [ + -18.042863845825195, + false + ] + ], + [ + [ + -19.464723587036133, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.232826232910156, + false + ], + [ + -22.290088653564453, + false + ], + [ + -18.042863845825195, + false + ], + [ + -19.464723587036133, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0106eeed7c37ae01edec7570e1ae1c221b7254621b32264510364b75a26e88a7", + "prompt_hash": "ee5726b35273b27cab1d8fdc2a7741030ee5426a2d8e9dcdfc70735de5142a86", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 68, + "doc": { + "question": "Which dimension is specifically used in the \\u2018\\abmc\\u2019 scenario to evaluate whether speed varies appropriately across environments?", + "choices": [ + "A. Trajectory", + "B. Embodied Interaction", + "C. Velocity", + "D. Aesthetics" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which dimension is specifically used in the \\u2018\\abmc\\u2019 scenario to evaluate whether speed varies appropriately across environments?\nAnswer:", + " A. Trajectory" + ], + [ + "Question:Which dimension is specifically used in the \\u2018\\abmc\\u2019 scenario to evaluate whether speed varies appropriately across environments?\nAnswer:", + " B. Embodied Interaction" + ], + [ + "Question:Which dimension is specifically used in the \\u2018\\abmc\\u2019 scenario to evaluate whether speed varies appropriately across environments?\nAnswer:", + " C. Velocity" + ], + [ + "Question:Which dimension is specifically used in the \\u2018\\abmc\\u2019 scenario to evaluate whether speed varies appropriately across environments?\nAnswer:", + " D. Aesthetics" + ] + ], + "resps": [ + [ + [ + -15.953933715820312, + false + ] + ], + [ + [ + -28.505935668945312, + false + ] + ], + [ + [ + -17.61598777770996, + false + ] + ], + [ + [ + -20.809507369995117, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.953933715820312, + false + ], + [ + -28.505935668945312, + false + ], + [ + -17.61598777770996, + false + ], + [ + -20.809507369995117, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3990cd538cbdf9365d531ac8430c2b207e2a1d05895d2a94e1f9bd888b363b82", + "prompt_hash": "2ce6d5a0bdaf02c26b50587735534e59d52e3cc879bc864e2be9113582787fa4", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 69, + "doc": { + "question": "What is the name of the video-level evaluator trained on HF-Embodied dataset?", + "choices": [ + "A. GPT-4o", + "B. FlashVStream", + "C. LoRA", + "D. evaluator" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the name of the video-level evaluator trained on HF-Embodied dataset?\nAnswer:", + " A. GPT-4o" + ], + [ + "Question:What is the name of the video-level evaluator trained on HF-Embodied dataset?\nAnswer:", + " B. FlashVStream" + ], + [ + "Question:What is the name of the video-level evaluator trained on HF-Embodied dataset?\nAnswer:", + " C. LoRA" + ], + [ + "Question:What is the name of the video-level evaluator trained on HF-Embodied dataset?\nAnswer:", + " D. evaluator" + ] + ], + "resps": [ + [ + [ + -14.393389701843262, + false + ] + ], + [ + [ + -43.895687103271484, + false + ] + ], + [ + [ + -19.041667938232422, + false + ] + ], + [ + [ + -21.171215057373047, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.393389701843262, + false + ], + [ + -43.895687103271484, + false + ], + [ + -19.041667938232422, + false + ], + [ + -21.171215057373047, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c4c0a3f71245d5aef21fb670f8d69a2ea0c6b9ff284ac337ce92c1639acbbc2b", + "prompt_hash": "b32bc05ee2c22f1d58ed12d7e2fca4906e868fc81e57716c4a023579e4ead278", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 70, + "doc": { + "question": "What was the performance improvement of journey learning over conventional supervised learning on the MATH dataset using only 327 training samples?", + "choices": [ + "Over 5%", + "Over 6%", + "Over 7%", + "Over 8%" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What was the performance improvement of journey learning over conventional supervised learning on the MATH dataset using only 327 training samples?\nAnswer:", + " Over 5%" + ], + [ + "Question:What was the performance improvement of journey learning over conventional supervised learning on the MATH dataset using only 327 training samples?\nAnswer:", + " Over 6%" + ], + [ + "Question:What was the performance improvement of journey learning over conventional supervised learning on the MATH dataset using only 327 training samples?\nAnswer:", + " Over 7%" + ], + [ + "Question:What was the performance improvement of journey learning over conventional supervised learning on the MATH dataset using only 327 training samples?\nAnswer:", + " Over 8%" + ] + ], + "resps": [ + [ + [ + -20.16574478149414, + false + ] + ], + [ + [ + -21.092784881591797, + false + ] + ], + [ + [ + -20.957870483398438, + false + ] + ], + [ + [ + -20.980009078979492, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.16574478149414, + false + ], + [ + -21.092784881591797, + false + ], + [ + -20.957870483398438, + false + ], + [ + -20.980009078979492, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0b38a2a56a0912b086eab973f3de7d206f452885ad75f309e1afecfe669a85ac", + "prompt_hash": "fe48919a07c4d400cd52f3a2516b1427fb5a6a6a0b9a6f25d08171a43a36e00b", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 71, + "doc": { + "question": "Which keyword is highlighted as an indicator of the model's ability to engage in reflection and self-correction?", + "choices": [ + "and", + "if", + "wait", + "consider" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which keyword is highlighted as an indicator of the model's ability to engage in reflection and self-correction?\nAnswer:", + " and" + ], + [ + "Question:Which keyword is highlighted as an indicator of the model's ability to engage in reflection and self-correction?\nAnswer:", + " if" + ], + [ + "Question:Which keyword is highlighted as an indicator of the model's ability to engage in reflection and self-correction?\nAnswer:", + " wait" + ], + [ + "Question:Which keyword is highlighted as an indicator of the model's ability to engage in reflection and self-correction?\nAnswer:", + " consider" + ] + ], + "resps": [ + [ + [ + -8.678184509277344, + false + ] + ], + [ + [ + -11.576055526733398, + false + ] + ], + [ + [ + -11.51467514038086, + false + ] + ], + [ + [ + -11.106267929077148, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.678184509277344, + false + ], + [ + -11.576055526733398, + false + ], + [ + -11.51467514038086, + false + ], + [ + -11.106267929077148, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4a96dbe55bfbdcc9ba72bc147dab79b3cc54ddf4e5750603408adcd6726a8c30", + "prompt_hash": "03917cb385f93db817b7541a7b7f5839f60a7932c74753f7df8af44ccb7ab021", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 72, + "doc": { + "question": "Which model achieved the highest F1 score on the subset of the PRM800K dataset?", + "choices": [ + "GPT-4o-mini", + "o1-preview", + "o1-mini", + "Math-shepherd" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model achieved the highest F1 score on the subset of the PRM800K dataset?\nAnswer:", + " GPT-4o-mini" + ], + [ + "Question:Which model achieved the highest F1 score on the subset of the PRM800K dataset?\nAnswer:", + " o1-preview" + ], + [ + "Question:Which model achieved the highest F1 score on the subset of the PRM800K dataset?\nAnswer:", + " o1-mini" + ], + [ + "Question:Which model achieved the highest F1 score on the subset of the PRM800K dataset?\nAnswer:", + " Math-shepherd" + ] + ], + "resps": [ + [ + [ + -16.167285919189453, + false + ] + ], + [ + [ + -29.22384262084961, + false + ] + ], + [ + [ + -26.068363189697266, + false + ] + ], + [ + [ + -29.02634048461914, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.167285919189453, + false + ], + [ + -29.22384262084961, + false + ], + [ + -26.068363189697266, + false + ], + [ + -29.02634048461914, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a1d92fb539e2211d60d6ca6220ce9941130781647199c626073e3ba9680df7e1", + "prompt_hash": "589e4789166075cb696d865c767e1dde0771e39b0b72cf0af4f8e4eaffb0a658", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 73, + "doc": { + "question": "Which phase in the training pipeline involved training on 327 examples of long thoughts including reflections and corrections?", + "choices": [ + "Supervised Fine-Tuning Phase 1", + "Supervised Fine-Tuning Phase 2: Shortcut Learning", + "Supervised Fine-Tuning Phase 2: Journey Learning", + "Direct Preference Optimization" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which phase in the training pipeline involved training on 327 examples of long thoughts including reflections and corrections?\nAnswer:", + " Supervised Fine-Tuning Phase 1" + ], + [ + "Question:Which phase in the training pipeline involved training on 327 examples of long thoughts including reflections and corrections?\nAnswer:", + " Supervised Fine-Tuning Phase 2: Shortcut Learning" + ], + [ + "Question:Which phase in the training pipeline involved training on 327 examples of long thoughts including reflections and corrections?\nAnswer:", + " Supervised Fine-Tuning Phase 2: Journey Learning" + ], + [ + "Question:Which phase in the training pipeline involved training on 327 examples of long thoughts including reflections and corrections?\nAnswer:", + " Direct Preference Optimization" + ] + ], + "resps": [ + [ + [ + -18.895179748535156, + false + ] + ], + [ + [ + -35.28461456298828, + false + ] + ], + [ + [ + -37.04241943359375, + false + ] + ], + [ + [ + -9.66794204711914, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.895179748535156, + false + ], + [ + -35.28461456298828, + false + ], + [ + -37.04241943359375, + false + ], + [ + -9.66794204711914, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "02febbe76e396536a06aa7130637a3193a3899e6913f805ce3b17ccd9c62f61a", + "prompt_hash": "db9356601bfe5d31416c62fc62bc9ff729c695dd8c9a625807dcb9db5b37ea44", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 74, + "doc": { + "question": "Which reasoning paradigm emphasizes learning through trial-and-error, reflection, and backtracking?", + "choices": [ + "Shortcut learning", + "Zero-shot learning", + "Journey learning", + "Chain-of-thought learning" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which reasoning paradigm emphasizes learning through trial-and-error, reflection, and backtracking?\nAnswer:", + " Shortcut learning" + ], + [ + "Question:Which reasoning paradigm emphasizes learning through trial-and-error, reflection, and backtracking?\nAnswer:", + " Zero-shot learning" + ], + [ + "Question:Which reasoning paradigm emphasizes learning through trial-and-error, reflection, and backtracking?\nAnswer:", + " Journey learning" + ], + [ + "Question:Which reasoning paradigm emphasizes learning through trial-and-error, reflection, and backtracking?\nAnswer:", + " Chain-of-thought learning" + ] + ], + "resps": [ + [ + [ + -16.114988327026367, + false + ] + ], + [ + [ + -11.982396125793457, + false + ] + ], + [ + [ + -21.8007869720459, + false + ] + ], + [ + [ + -18.297088623046875, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.114988327026367, + false + ], + [ + -11.982396125793457, + false + ], + [ + -21.8007869720459, + false + ], + [ + -18.297088623046875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b2c6e5fd2e897e762c53be4d164a0b0084990c633d5674b313b910d9317c6223", + "prompt_hash": "a59d7ff020b71e354b01ae1fcf7bb63b83a17c02d02a02ea95235ff1304fb392", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 75, + "doc": { + "question": "What types of data were included in the pre-training of GPT-4o?", + "choices": [ + "Only publicly available datasets", + "Only user-contributed content", + "Web data, code and math, multimodal data", + "Data exclusively from government archives" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What types of data were included in the pre-training of GPT-4o?\nAnswer:", + " Only publicly available datasets" + ], + [ + "Question:What types of data were included in the pre-training of GPT-4o?\nAnswer:", + " Only user-contributed content" + ], + [ + "Question:What types of data were included in the pre-training of GPT-4o?\nAnswer:", + " Web data, code and math, multimodal data" + ], + [ + "Question:What types of data were included in the pre-training of GPT-4o?\nAnswer:", + " Data exclusively from government archives" + ] + ], + "resps": [ + [ + [ + -16.46111297607422, + false + ] + ], + [ + [ + -21.123538970947266, + false + ] + ], + [ + [ + -33.048213958740234, + false + ] + ], + [ + [ + -30.38631248474121, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.46111297607422, + false + ], + [ + -21.123538970947266, + false + ], + [ + -33.048213958740234, + false + ], + [ + -30.38631248474121, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a372985e07f7abbe54f600d63ee6d5f7fc6bffc04505cbae54da4d2bf27521bc", + "prompt_hash": "0104587737bedb6bdb520872d906a3d9ad0822941ea4075fb38f27cf3e75c6ef", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 76, + "doc": { + "question": "What was the final risk classification for GPT-4o under OpenAI’s Preparedness Framework?", + "choices": [ + "High", + "Medium", + "Low", + "Critical" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What was the final risk classification for GPT-4o under OpenAI’s Preparedness Framework?\nAnswer:", + " High" + ], + [ + "Question:What was the final risk classification for GPT-4o under OpenAI’s Preparedness Framework?\nAnswer:", + " Medium" + ], + [ + "Question:What was the final risk classification for GPT-4o under OpenAI’s Preparedness Framework?\nAnswer:", + " Low" + ], + [ + "Question:What was the final risk classification for GPT-4o under OpenAI’s Preparedness Framework?\nAnswer:", + " Critical" + ] + ], + "resps": [ + [ + [ + -4.722260475158691, + false + ] + ], + [ + [ + -6.896150588989258, + false + ] + ], + [ + [ + -5.012272834777832, + false + ] + ], + [ + [ + -5.753252983093262, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.722260475158691, + false + ], + [ + -6.896150588989258, + false + ], + [ + -5.012272834777832, + false + ], + [ + -5.753252983093262, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4e0758f5d5174342e37f294d716b53e741c0f4bb1a12c6b51960807a7c821ac3", + "prompt_hash": "3256e306cb57e76dae5c1a79329caf9ed10dd3825feaaadb0e6415d3d8358cc7", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 77, + "doc": { + "question": "How many languages did the GPT-4o external red teamers collectively speak?", + "choices": [ + "15", + "29", + "45", + "60" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many languages did the GPT-4o external red teamers collectively speak?\nAnswer:", + " 15" + ], + [ + "Question:How many languages did the GPT-4o external red teamers collectively speak?\nAnswer:", + " 29" + ], + [ + "Question:How many languages did the GPT-4o external red teamers collectively speak?\nAnswer:", + " 45" + ], + [ + "Question:How many languages did the GPT-4o external red teamers collectively speak?\nAnswer:", + " 60" + ] + ], + "resps": [ + [ + [ + -5.321237087249756, + false + ] + ], + [ + [ + -6.68866491317749, + false + ] + ], + [ + [ + -6.993188858032227, + false + ] + ], + [ + [ + -6.728579521179199, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.321237087249756, + false + ], + [ + -6.68866491317749, + false + ], + [ + -6.993188858032227, + false + ], + [ + -6.728579521179199, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6339a9a7bef3a9eea574a239c5b1e0b022e7a614bde294b7614d603313d2ec5e", + "prompt_hash": "73b492ab37fb4cc79613d41e9e60760c03659310f5c6a8a2c8407621a38a4c6f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 78, + "doc": { + "question": "Which system was used to convert text to audio for evaluation purposes?", + "choices": [ + "DALL-E", + "Voice Engine", + "AudioLM", + "Whisper" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which system was used to convert text to audio for evaluation purposes?\nAnswer:", + " DALL-E" + ], + [ + "Question:Which system was used to convert text to audio for evaluation purposes?\nAnswer:", + " Voice Engine" + ], + [ + "Question:Which system was used to convert text to audio for evaluation purposes?\nAnswer:", + " AudioLM" + ], + [ + "Question:Which system was used to convert text to audio for evaluation purposes?\nAnswer:", + " Whisper" + ] + ], + "resps": [ + [ + [ + -10.943501472473145, + false + ] + ], + [ + [ + -16.534181594848633, + false + ] + ], + [ + [ + -8.04608154296875, + false + ] + ], + [ + [ + -6.90962553024292, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.943501472473145, + false + ], + [ + -16.534181594848633, + false + ], + [ + -8.04608154296875, + false + ], + [ + -6.90962553024292, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d28ce463101a24c6e1288b67a0efd3244c77b55169d76730f01b4b7e6177f1df", + "prompt_hash": "ca824eae6c8a5d78e8ea1f9a0dacb4e6f2ceebbd524d7b8fea9502455b9729d0", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 79, + "doc": { + "question": "In voice output classification, what recall was achieved for non-English languages?", + "choices": [ + "0.85", + "0.90", + "1.0", + "0.95" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In voice output classification, what recall was achieved for non-English languages?\nAnswer:", + " 0.85" + ], + [ + "Question:In voice output classification, what recall was achieved for non-English languages?\nAnswer:", + " 0.90" + ], + [ + "Question:In voice output classification, what recall was achieved for non-English languages?\nAnswer:", + " 1.0" + ], + [ + "Question:In voice output classification, what recall was achieved for non-English languages?\nAnswer:", + " 0.95" + ] + ], + "resps": [ + [ + [ + -9.070045471191406, + false + ] + ], + [ + [ + -9.178909301757812, + false + ] + ], + [ + [ + -8.597393035888672, + false + ] + ], + [ + [ + -9.124435424804688, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.070045471191406, + false + ], + [ + -9.178909301757812, + false + ], + [ + -8.597393035888672, + false + ], + [ + -9.124435424804688, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "32c94fe4437b01c5ef72e2aa2c37b59ea35d0fee9fbd37285420512012a356e6", + "prompt_hash": "436600b9a53a410a7a5ddc2c85b9f999f6477738cbcbdd3545f1dd649528a24f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 80, + "doc": { + "question": "What is the screen width tolerance used to evaluate click-based actions as correct?", + "choices": [ + "10% of screen width", + "12% of screen width", + "14% of screen width", + "16% of screen width" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the screen width tolerance used to evaluate click-based actions as correct?\nAnswer:", + " 10% of screen width" + ], + [ + "Question:What is the screen width tolerance used to evaluate click-based actions as correct?\nAnswer:", + " 12% of screen width" + ], + [ + "Question:What is the screen width tolerance used to evaluate click-based actions as correct?\nAnswer:", + " 14% of screen width" + ], + [ + "Question:What is the screen width tolerance used to evaluate click-based actions as correct?\nAnswer:", + " 16% of screen width" + ] + ], + "resps": [ + [ + [ + -17.385103225708008, + false + ] + ], + [ + [ + -22.00188446044922, + false + ] + ], + [ + [ + -22.98278045654297, + false + ] + ], + [ + [ + -22.503746032714844, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.385103225708008, + false + ], + [ + -22.00188446044922, + false + ], + [ + -22.98278045654297, + false + ], + [ + -22.503746032714844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9f8b5be0a3b9d267dde452ed7eac20b48ec86f51bbdf962ee9a7d9816042876c", + "prompt_hash": "310393acf8b5193cd43714a633b1db6ddcf4973879f52ae045a8c7e741bce8a2", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 81, + "doc": { + "question": "Which benchmark had 1,410 test samples and included 3 basic plus 2 custom actions?", + "choices": [ + "GUI-Act-Web", + "OmniAct-Desktop", + "GUI-Odyssey-App", + "AndroidControl-High" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which benchmark had 1,410 test samples and included 3 basic plus 2 custom actions?\nAnswer:", + " GUI-Act-Web" + ], + [ + "Question:Which benchmark had 1,410 test samples and included 3 basic plus 2 custom actions?\nAnswer:", + " OmniAct-Desktop" + ], + [ + "Question:Which benchmark had 1,410 test samples and included 3 basic plus 2 custom actions?\nAnswer:", + " GUI-Odyssey-App" + ], + [ + "Question:Which benchmark had 1,410 test samples and included 3 basic plus 2 custom actions?\nAnswer:", + " AndroidControl-High" + ] + ], + "resps": [ + [ + [ + -27.018978118896484, + false + ] + ], + [ + [ + -35.97882843017578, + false + ] + ], + [ + [ + -38.291839599609375, + false + ] + ], + [ + [ + -32.71967697143555, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.018978118896484, + false + ], + [ + -35.97882843017578, + false + ], + [ + -38.291839599609375, + false + ], + [ + -32.71967697143555, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2fa1819279e963427cc8e97b7fb0cb586778c2437dd702a77a7cfdc56d8a5af1", + "prompt_hash": "ad0ee4079fd39fe7f78fe4045eb915029c0d600e0584f0f0ad542099cf9cdd08", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 82, + "doc": { + "question": "Which dataset had the highest number of screenshots in the GUI grounding pretraining corpus?", + "choices": [ + "FineWeb-filtered", + "SeeClick", + "Windows-desktop", + "Mind2Web-annotated" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which dataset had the highest number of screenshots in the GUI grounding pretraining corpus?\nAnswer:", + " FineWeb-filtered" + ], + [ + "Question:Which dataset had the highest number of screenshots in the GUI grounding pretraining corpus?\nAnswer:", + " SeeClick" + ], + [ + "Question:Which dataset had the highest number of screenshots in the GUI grounding pretraining corpus?\nAnswer:", + " Windows-desktop" + ], + [ + "Question:Which dataset had the highest number of screenshots in the GUI grounding pretraining corpus?\nAnswer:", + " Mind2Web-annotated" + ] + ], + "resps": [ + [ + [ + -25.71546173095703, + false + ] + ], + [ + [ + -24.613677978515625, + false + ] + ], + [ + [ + -16.762664794921875, + false + ] + ], + [ + [ + -30.07876205444336, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.71546173095703, + false + ], + [ + -24.613677978515625, + false + ], + [ + -16.762664794921875, + false + ], + [ + -30.07876205444336, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "929fea3025f2a9b1affc7e2ab173b8b92436b91ff4daf4678c3664bd0fe130dc", + "prompt_hash": "d037808124da7a33a08b66e11ea5af7d9090610ebf90d71d8a3bbcc12d1bc1c4", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 83, + "doc": { + "question": "What method was used to synthesize sub-instructions for instruction grounding data?", + "choices": [ + "Chain-of-Thought prompting", + "Set-of-Mark prompting", + "Few-shot prompting", + "Auto-regressive prompting" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What method was used to synthesize sub-instructions for instruction grounding data?\nAnswer:", + " Chain-of-Thought prompting" + ], + [ + "Question:What method was used to synthesize sub-instructions for instruction grounding data?\nAnswer:", + " Set-of-Mark prompting" + ], + [ + "Question:What method was used to synthesize sub-instructions for instruction grounding data?\nAnswer:", + " Few-shot prompting" + ], + [ + "Question:What method was used to synthesize sub-instructions for instruction grounding data?\nAnswer:", + " Auto-regressive prompting" + ] + ], + "resps": [ + [ + [ + -10.850055694580078, + false + ] + ], + [ + [ + -22.5295352935791, + false + ] + ], + [ + [ + -11.125772476196289, + false + ] + ], + [ + [ + -13.05416488647461, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.850055694580078, + false + ], + [ + -22.5295352935791, + false + ], + [ + -11.125772476196289, + false + ], + [ + -13.05416488647461, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a859d38b3a1df340c92657008d4464c3f55dc09d6d2d59c847fae83a73b231af", + "prompt_hash": "33f05316acc2217230f633db435d256e9a6ab4886eaa408aa2850e362322267e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 84, + "doc": { + "question": "What unified action reduces naming conflicts such as 'tap' and 'click'?", + "choices": [ + "Dynamic Action Mapping", + "Standardized Action Encoding", + "Unified Action Space", + "Cross-Domain Action Bridge" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What unified action reduces naming conflicts such as 'tap' and 'click'?\nAnswer:", + " Dynamic Action Mapping" + ], + [ + "Question:What unified action reduces naming conflicts such as 'tap' and 'click'?\nAnswer:", + " Standardized Action Encoding" + ], + [ + "Question:What unified action reduces naming conflicts such as 'tap' and 'click'?\nAnswer:", + " Unified Action Space" + ], + [ + "Question:What unified action reduces naming conflicts such as 'tap' and 'click'?\nAnswer:", + " Cross-Domain Action Bridge" + ] + ], + "resps": [ + [ + [ + -22.388999938964844, + false + ] + ], + [ + [ + -23.831134796142578, + false + ] + ], + [ + [ + -21.0826358795166, + false + ] + ], + [ + [ + -29.27459716796875, + false + ] + ] + ], + "filtered_resps": [ + [ + -22.388999938964844, + false + ], + [ + -23.831134796142578, + false + ], + [ + -21.0826358795166, + false + ], + [ + -29.27459716796875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "279699b89160ac7cc040c8217e39a7c78bab1d462634e0e1716881353083fe4b", + "prompt_hash": "a43e32d200b3b898278c02ca262be5772b92ae6d2c2e741e1163da5b3ebcb957", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 85, + "doc": { + "question": "What is the codebook size used by the MAGVIT-v2 quantizer for image tokenization in Show-o?", + "choices": [ + "4,096", + "8,192", + "16,384", + "2,048" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the codebook size used by the MAGVIT-v2 quantizer for image tokenization in Show-o?\nAnswer:", + " 4,096" + ], + [ + "Question:What is the codebook size used by the MAGVIT-v2 quantizer for image tokenization in Show-o?\nAnswer:", + " 8,192" + ], + [ + "Question:What is the codebook size used by the MAGVIT-v2 quantizer for image tokenization in Show-o?\nAnswer:", + " 16,384" + ], + [ + "Question:What is the codebook size used by the MAGVIT-v2 quantizer for image tokenization in Show-o?\nAnswer:", + " 2,048" + ] + ], + "resps": [ + [ + [ + -9.725521087646484, + false + ] + ], + [ + [ + -11.152690887451172, + false + ] + ], + [ + [ + -9.869206428527832, + false + ] + ], + [ + [ + -9.555603981018066, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.725521087646484, + false + ], + [ + -11.152690887451172, + false + ], + [ + -9.869206428527832, + false + ], + [ + -9.555603981018066, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "537f8e13a8575c99b1c95067d90cf9ea8a3221d78b0b3c2bef9ba1ed2bc4f74f", + "prompt_hash": "51f7d65b8b8b0b214b4856c8f48f9c5ef119b17e890df6f58283de6b776c09d6", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 86, + "doc": { + "question": "Which operation is prepended to each attention layer in Show-o's architecture?", + "choices": [ + "Layer Normalization", + "Residual Connection", + "QK-Norm", + "Batch Normalization" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which operation is prepended to each attention layer in Show-o's architecture?\nAnswer:", + " Layer Normalization" + ], + [ + "Question:Which operation is prepended to each attention layer in Show-o's architecture?\nAnswer:", + " Residual Connection" + ], + [ + "Question:Which operation is prepended to each attention layer in Show-o's architecture?\nAnswer:", + " QK-Norm" + ], + [ + "Question:Which operation is prepended to each attention layer in Show-o's architecture?\nAnswer:", + " Batch Normalization" + ] + ], + "resps": [ + [ + [ + -8.532977104187012, + false + ] + ], + [ + [ + -10.449203491210938, + false + ] + ], + [ + [ + -17.17314910888672, + false + ] + ], + [ + [ + -10.140875816345215, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.532977104187012, + false + ], + [ + -10.449203491210938, + false + ], + [ + -17.17314910888672, + false + ], + [ + -10.140875816345215, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e82398508a7698a85b0ca7fce40294bedf9db1c8fa02e12de0642116958caea0", + "prompt_hash": "7a36bb0baac4bc975b5211cea95fc4cd933e9fa4a9ab24630ca9b841e27d2697", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 87, + "doc": { + "question": "Which dataset is used in the first stage of training Show-o to learn pixel dependency for image generation?", + "choices": [ + "COCO", + "ImageNet-1K", + "RefinedWeb", + "LAION-aesthetics-12M" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset is used in the first stage of training Show-o to learn pixel dependency for image generation?\nAnswer:", + " COCO" + ], + [ + "Question:Which dataset is used in the first stage of training Show-o to learn pixel dependency for image generation?\nAnswer:", + " ImageNet-1K" + ], + [ + "Question:Which dataset is used in the first stage of training Show-o to learn pixel dependency for image generation?\nAnswer:", + " RefinedWeb" + ], + [ + "Question:Which dataset is used in the first stage of training Show-o to learn pixel dependency for image generation?\nAnswer:", + " LAION-aesthetics-12M" + ] + ], + "resps": [ + [ + [ + -8.915375709533691, + false + ] + ], + [ + [ + -8.84010124206543, + false + ] + ], + [ + [ + -16.313142776489258, + false + ] + ], + [ + [ + -28.85309600830078, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.915375709533691, + false + ], + [ + -8.84010124206543, + false + ], + [ + -16.313142776489258, + false + ], + [ + -28.85309600830078, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a2fbac91fa5a58fbaad42c9fd06e66ed9324b31f1a17f89ec6f6fbe662e28996", + "prompt_hash": "9eeb735d357ee7d91d5a5551029508dc08da295606cb72b157d6cfa3a55b60d9", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 88, + "doc": { + "question": "What is the overall training loss of Show-o composed of?", + "choices": [ + "Only Mask Token Prediction loss", + "Only Next Token Prediction loss", + "Combination of Mask Token Prediction and Next Token Prediction losses", + "Reconstruction loss only" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the overall training loss of Show-o composed of?\nAnswer:", + " Only Mask Token Prediction loss" + ], + [ + "Question:What is the overall training loss of Show-o composed of?\nAnswer:", + " Only Next Token Prediction loss" + ], + [ + "Question:What is the overall training loss of Show-o composed of?\nAnswer:", + " Combination of Mask Token Prediction and Next Token Prediction losses" + ], + [ + "Question:What is the overall training loss of Show-o composed of?\nAnswer:", + " Reconstruction loss only" + ] + ], + "resps": [ + [ + [ + -32.088958740234375, + false + ] + ], + [ + [ + -26.067264556884766, + false + ] + ], + [ + [ + -45.478004455566406, + false + ] + ], + [ + [ + -15.250179290771484, + false + ] + ] + ], + "filtered_resps": [ + [ + -32.088958740234375, + false + ], + [ + -26.067264556884766, + false + ], + [ + -45.478004455566406, + false + ], + [ + -15.250179290771484, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ad4111d75628f6b33d86e9e55bd4403451877b988a90f06f7e5995caa76f78eb", + "prompt_hash": "3738b86b77deb7b607d86bd29382528e9215d70aa7fc71ea7d06be174226ec65", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 89, + "doc": { + "question": "During inference, how many steps does Show-o take to generate an image?", + "choices": [ + "1 step", + "T steps", + "10 steps", + "512 steps" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:During inference, how many steps does Show-o take to generate an image?\nAnswer:", + " 1 step" + ], + [ + "Question:During inference, how many steps does Show-o take to generate an image?\nAnswer:", + " T steps" + ], + [ + "Question:During inference, how many steps does Show-o take to generate an image?\nAnswer:", + " 10 steps" + ], + [ + "Question:During inference, how many steps does Show-o take to generate an image?\nAnswer:", + " 512 steps" + ] + ], + "resps": [ + [ + [ + -6.212777137756348, + false + ] + ], + [ + [ + -14.03068733215332, + false + ] + ], + [ + [ + -7.745494365692139, + false + ] + ], + [ + [ + -10.251701354980469, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.212777137756348, + false + ], + [ + -14.03068733215332, + false + ], + [ + -7.745494365692139, + false + ], + [ + -10.251701354980469, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "97590d342be3769a9363b2f9240f40242666e6744ec386dc8e6e036b5db150f9", + "prompt_hash": "769bd6096ad2dc7e8d1263e5bfe602f8928265bfdb9024d078d113b3a888c9f0", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 90, + "doc": { + "question": "How many total tokens did Qwen2-VL process during pre-training stages?", + "choices": [ + "600 billion", + "800 billion", + "1.4 trillion", + "2 trillion" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many total tokens did Qwen2-VL process during pre-training stages?\nAnswer:", + " 600 billion" + ], + [ + "Question:How many total tokens did Qwen2-VL process during pre-training stages?\nAnswer:", + " 800 billion" + ], + [ + "Question:How many total tokens did Qwen2-VL process during pre-training stages?\nAnswer:", + " 1.4 trillion" + ], + [ + "Question:How many total tokens did Qwen2-VL process during pre-training stages?\nAnswer:", + " 2 trillion" + ] + ], + "resps": [ + [ + [ + -9.786516189575195, + false + ] + ], + [ + [ + -10.318535804748535, + false + ] + ], + [ + [ + -10.681632995605469, + false + ] + ], + [ + [ + -9.810903549194336, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.786516189575195, + false + ], + [ + -10.318535804748535, + false + ], + [ + -10.681632995605469, + false + ], + [ + -9.810903549194336, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a343616497f6eb027f084ca72bf2bea7781fcffb6f13b70186848784fa662566", + "prompt_hash": "17881d277e9e366a35093991cd6021414d8f960bc8c022bc075c47e96590f8b4", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 91, + "doc": { + "question": "What is the parameter size of the Vision Transformer (ViT) used in Qwen2-VL?", + "choices": [ + "224M", + "675M", + "1.5B", + "2B" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the parameter size of the Vision Transformer (ViT) used in Qwen2-VL?\nAnswer:", + " 224M" + ], + [ + "Question:What is the parameter size of the Vision Transformer (ViT) used in Qwen2-VL?\nAnswer:", + " 675M" + ], + [ + "Question:What is the parameter size of the Vision Transformer (ViT) used in Qwen2-VL?\nAnswer:", + " 1.5B" + ], + [ + "Question:What is the parameter size of the Vision Transformer (ViT) used in Qwen2-VL?\nAnswer:", + " 2B" + ] + ], + "resps": [ + [ + [ + -11.379950523376465, + false + ] + ], + [ + [ + -14.706459999084473, + false + ] + ], + [ + [ + -10.820328712463379, + false + ] + ], + [ + [ + -12.057304382324219, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.379950523376465, + false + ], + [ + -14.706459999084473, + false + ], + [ + -10.820328712463379, + false + ], + [ + -12.057304382324219, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "832be3f29de56d53c7952bcdc890db00f6ccdb0ea4a060b4c54a74a9b415bf26", + "prompt_hash": "88dd86e33d060dccd865e0939d7661e72dd9e27790014246c7e6eb55efb53507", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 92, + "doc": { + "question": "Which technique was introduced in Qwen2-VL to replace absolute position embeddings in the Vision Transformer?", + "choices": [ + "1D-RoPE", + "Naive Dynamic Resolution", + "2D-RoPE", + "M-RoPE" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which technique was introduced in Qwen2-VL to replace absolute position embeddings in the Vision Transformer?\nAnswer:", + " 1D-RoPE" + ], + [ + "Question:Which technique was introduced in Qwen2-VL to replace absolute position embeddings in the Vision Transformer?\nAnswer:", + " Naive Dynamic Resolution" + ], + [ + "Question:Which technique was introduced in Qwen2-VL to replace absolute position embeddings in the Vision Transformer?\nAnswer:", + " 2D-RoPE" + ], + [ + "Question:Which technique was introduced in Qwen2-VL to replace absolute position embeddings in the Vision Transformer?\nAnswer:", + " M-RoPE" + ] + ], + "resps": [ + [ + [ + -23.201679229736328, + false + ] + ], + [ + [ + -26.624053955078125, + false + ] + ], + [ + [ + -19.48124122619629, + false + ] + ], + [ + [ + -22.730064392089844, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.201679229736328, + false + ], + [ + -26.624053955078125, + false + ], + [ + -19.48124122619629, + false + ], + [ + -22.730064392089844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "bf5e204e913102aebfc35f62da4c9110e1a76382a80de783093c72686f54e765", + "prompt_hash": "213aa4463b3b448efc1f02e8044c4e03a1d866b068bee8662d49dd9d80e21fbb", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 93, + "doc": { + "question": "What is the length limit of videos that Qwen2-VL is capable of understanding?", + "choices": [ + "5 minutes", + "10 minutes", + "15 minutes", + "20+ minutes" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the length limit of videos that Qwen2-VL is capable of understanding?\nAnswer:", + " 5 minutes" + ], + [ + "Question:What is the length limit of videos that Qwen2-VL is capable of understanding?\nAnswer:", + " 10 minutes" + ], + [ + "Question:What is the length limit of videos that Qwen2-VL is capable of understanding?\nAnswer:", + " 15 minutes" + ], + [ + "Question:What is the length limit of videos that Qwen2-VL is capable of understanding?\nAnswer:", + " 20+ minutes" + ] + ], + "resps": [ + [ + [ + -6.9861159324646, + false + ] + ], + [ + [ + -7.793810844421387, + false + ] + ], + [ + [ + -7.572601318359375, + false + ] + ], + [ + [ + -15.895883560180664, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.9861159324646, + false + ], + [ + -7.793810844421387, + false + ], + [ + -7.572601318359375, + false + ], + [ + -15.895883560180664, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "359a8f69c0acbb9228cd774f5931ed894339dfbcfe1c35d8f04728dd117dec03", + "prompt_hash": "2d70ef5aa8991e86d7ea0ed86261a6c7835cc0dc4fe835dfd948c2ad23fb407f", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 94, + "doc": { + "question": "Which benchmark did Qwen2-VL-72B achieve a score of 77.8, surpassing GPT-4o?", + "choices": [ + "MathVista", + "RealWorldQA", + "MMStar", + "DocVQA" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which benchmark did Qwen2-VL-72B achieve a score of 77.8, surpassing GPT-4o?\nAnswer:", + " MathVista" + ], + [ + "Question:Which benchmark did Qwen2-VL-72B achieve a score of 77.8, surpassing GPT-4o?\nAnswer:", + " RealWorldQA" + ], + [ + "Question:Which benchmark did Qwen2-VL-72B achieve a score of 77.8, surpassing GPT-4o?\nAnswer:", + " MMStar" + ], + [ + "Question:Which benchmark did Qwen2-VL-72B achieve a score of 77.8, surpassing GPT-4o?\nAnswer:", + " DocVQA" + ] + ], + "resps": [ + [ + [ + -13.382601737976074, + false + ] + ], + [ + [ + -14.511748313903809, + false + ] + ], + [ + [ + -16.00177764892578, + false + ] + ], + [ + [ + -12.91325569152832, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.382601737976074, + false + ], + [ + -14.511748313903809, + false + ], + [ + -16.00177764892578, + false + ], + [ + -12.91325569152832, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "80a3f392f925b83d72b0de9ae7ffab91162ac793f757151d7b8ac89c4b222093", + "prompt_hash": "9ec2afa7905d405eda6affd8a9c929f80e8bea537aca5568db62a5cb9eb23dc1", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 95, + "doc": { + "question": "Which dataset includes prompts from ShareGPT and is annotated with five specific attributes?", + "choices": [ + "Magpie Ultra", + "HelpSteer2", + "OffsetBias", + "WildGuardMix" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset includes prompts from ShareGPT and is annotated with five specific attributes?\nAnswer:", + " Magpie Ultra" + ], + [ + "Question:Which dataset includes prompts from ShareGPT and is annotated with five specific attributes?\nAnswer:", + " HelpSteer2" + ], + [ + "Question:Which dataset includes prompts from ShareGPT and is annotated with five specific attributes?\nAnswer:", + " OffsetBias" + ], + [ + "Question:Which dataset includes prompts from ShareGPT and is annotated with five specific attributes?\nAnswer:", + " WildGuardMix" + ] + ], + "resps": [ + [ + [ + -28.369426727294922, + false + ] + ], + [ + [ + -23.477996826171875, + false + ] + ], + [ + [ + -24.307876586914062, + false + ] + ], + [ + [ + -30.440494537353516, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.369426727294922, + false + ], + [ + -23.477996826171875, + false + ], + [ + -24.307876586914062, + false + ], + [ + -30.440494537353516, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "816acbde7d3d995f0878ae52bd5249e7d9dc11fb5c22c5485554cc5ab645ee47", + "prompt_hash": "a6b409831786f78a1950fc5a187ca30dbaec3320e3fb5b0d850d4bbeb5704c7e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 96, + "doc": { + "question": "Which loss function consistently outperformed others in the study for reward modeling?", + "choices": [ + "Hinge Loss", + "Cross-Entropy", + "Focal Loss", + "Bradley-Terry" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which loss function consistently outperformed others in the study for reward modeling?\nAnswer:", + " Hinge Loss" + ], + [ + "Question:Which loss function consistently outperformed others in the study for reward modeling?\nAnswer:", + " Cross-Entropy" + ], + [ + "Question:Which loss function consistently outperformed others in the study for reward modeling?\nAnswer:", + " Focal Loss" + ], + [ + "Question:Which loss function consistently outperformed others in the study for reward modeling?\nAnswer:", + " Bradley-Terry" + ] + ], + "resps": [ + [ + [ + -10.29498291015625, + false + ] + ], + [ + [ + -6.921933174133301, + false + ] + ], + [ + [ + -13.43044662475586, + false + ] + ], + [ + [ + -9.149613380432129, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.29498291015625, + false + ], + [ + -6.921933174133301, + false + ], + [ + -13.43044662475586, + false + ], + [ + -9.149613380432129, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "541e9fe2894b0f49b4709cdd760aa1d58ddff2d3a5250e5c3f03403e1b5b0519", + "prompt_hash": "a793f665a42e67ed3cb092ec91c1e13f07b0aa4e5db79a1c48d55eb0ed8ce46f", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 97, + "doc": { + "question": "What is the size of the Skywork Reward Preference 80K dataset in terms of total preference pairs?", + "choices": [ + "378,000", + "80,000", + "59,539", + "700,000" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the size of the Skywork Reward Preference 80K dataset in terms of total preference pairs?\nAnswer:", + " 378,000" + ], + [ + "Question:What is the size of the Skywork Reward Preference 80K dataset in terms of total preference pairs?\nAnswer:", + " 80,000" + ], + [ + "Question:What is the size of the Skywork Reward Preference 80K dataset in terms of total preference pairs?\nAnswer:", + " 59,539" + ], + [ + "Question:What is the size of the Skywork Reward Preference 80K dataset in terms of total preference pairs?\nAnswer:", + " 700,000" + ] + ], + "resps": [ + [ + [ + -13.991911888122559, + false + ] + ], + [ + [ + -6.92512321472168, + false + ] + ], + [ + [ + -14.930144309997559, + false + ] + ], + [ + [ + -10.193219184875488, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.991911888122559, + false + ], + [ + -6.92512321472168, + false + ], + [ + -14.930144309997559, + false + ], + [ + -10.193219184875488, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9c7304f1822614dc658230f603a51ccef62ce2fd7fa526a20780f2926cc80131", + "prompt_hash": "1b67194f16ca8ff42ec2c50785c1ba0fae603196263f8e7245d190520fbbc403", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 98, + "doc": { + "question": "Which preference dataset was created to address spurious signals such as length bias in reward modeling?", + "choices": [ + "OffsetBias", + "HelpSteer2", + "Magpie Air", + "WildGuardMix" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which preference dataset was created to address spurious signals such as length bias in reward modeling?\nAnswer:", + " OffsetBias" + ], + [ + "Question:Which preference dataset was created to address spurious signals such as length bias in reward modeling?\nAnswer:", + " HelpSteer2" + ], + [ + "Question:Which preference dataset was created to address spurious signals such as length bias in reward modeling?\nAnswer:", + " Magpie Air" + ], + [ + "Question:Which preference dataset was created to address spurious signals such as length bias in reward modeling?\nAnswer:", + " WildGuardMix" + ] + ], + "resps": [ + [ + [ + -19.423641204833984, + false + ] + ], + [ + [ + -24.912960052490234, + false + ] + ], + [ + [ + -27.164596557617188, + false + ] + ], + [ + [ + -30.299091339111328, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.423641204833984, + false + ], + [ + -24.912960052490234, + false + ], + [ + -27.164596557617188, + false + ], + [ + -30.299091339111328, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f1c3744150badebebd1a5bd3004cdcdfc0be7644e56224b379dadd9bf6037680", + "prompt_hash": "a1478f9ea022bc9f3bcfcdbfbcf56dbbf3309ce1bbb4645058515d3b593c35f4", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 99, + "doc": { + "question": "Which model achieved the highest average score on RewardBench among the models listed?", + "choices": [ + "Nemotron-4-340B-Reward", + "Skywork-Reward-Gemma-2-27B", + "InternLM-20B-Reward", + "SFR-nemo-12B-Judge-r" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which model achieved the highest average score on RewardBench among the models listed?\nAnswer:", + " Nemotron-4-340B-Reward" + ], + [ + "Question:Which model achieved the highest average score on RewardBench among the models listed?\nAnswer:", + " Skywork-Reward-Gemma-2-27B" + ], + [ + "Question:Which model achieved the highest average score on RewardBench among the models listed?\nAnswer:", + " InternLM-20B-Reward" + ], + [ + "Question:Which model achieved the highest average score on RewardBench among the models listed?\nAnswer:", + " SFR-nemo-12B-Judge-r" + ] + ], + "resps": [ + [ + [ + -35.623653411865234, + false + ] + ], + [ + [ + -42.98133087158203, + false + ] + ], + [ + [ + -24.74456787109375, + false + ] + ], + [ + [ + -56.343814849853516, + false + ] + ] + ], + "filtered_resps": [ + [ + -35.623653411865234, + false + ], + [ + -42.98133087158203, + false + ], + [ + -24.74456787109375, + false + ], + [ + -56.343814849853516, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fdf78a4f218d6f44cd8777913beb60fcf1903a7e3353c3f7e9b03097b4626ff9", + "prompt_hash": "c5c9fbc6a6ffa1795b7bac2ff3cea81de793c3fa936b2b6e10e3a5408a773fff", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + } + ], + "arxiv_mc_2411": [ + { + "doc_id": 0, + "doc": { + "question": "What is the peak data rate target for 6G networks?", + "choices": [ + "10 Gbps", + "100 Gbps", + "1 Tbps", + "10 Tbps" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the peak data rate target for 6G networks?\nAnswer:", + " 10 Gbps" + ], + [ + "Question:What is the peak data rate target for 6G networks?\nAnswer:", + " 100 Gbps" + ], + [ + "Question:What is the peak data rate target for 6G networks?\nAnswer:", + " 1 Tbps" + ], + [ + "Question:What is the peak data rate target for 6G networks?\nAnswer:", + " 10 Tbps" + ] + ], + "resps": [ + [ + [ + -5.873969078063965, + false + ] + ], + [ + [ + -4.819973468780518, + false + ] + ], + [ + [ + -6.002620220184326, + false + ] + ], + [ + [ + -7.384844779968262, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.873969078063965, + false + ], + [ + -4.819973468780518, + false + ], + [ + -6.002620220184326, + false + ], + [ + -7.384844779968262, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "eec87ea7bd2734619ac06ae996519d9ebd74b6278d70c4af453c19857d10fa27", + "prompt_hash": "da9989854daa6aaca25ea247e4f40df164f29efbef51497fed3846972dfd2a20", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 1, + "doc": { + "question": "Which usage scenario in 6G focuses on enabling sensing capabilities like movement tracking and environmental monitoring?", + "choices": [ + "Immersive communication", + "Hyper reliable, low-latency communication", + "Integrated sensing and communication", + "Massive communication" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which usage scenario in 6G focuses on enabling sensing capabilities like movement tracking and environmental monitoring?\nAnswer:", + " Immersive communication" + ], + [ + "Question:Which usage scenario in 6G focuses on enabling sensing capabilities like movement tracking and environmental monitoring?\nAnswer:", + " Hyper reliable, low-latency communication" + ], + [ + "Question:Which usage scenario in 6G focuses on enabling sensing capabilities like movement tracking and environmental monitoring?\nAnswer:", + " Integrated sensing and communication" + ], + [ + "Question:Which usage scenario in 6G focuses on enabling sensing capabilities like movement tracking and environmental monitoring?\nAnswer:", + " Massive communication" + ] + ], + "resps": [ + [ + [ + -15.115461349487305, + false + ] + ], + [ + [ + -36.961116790771484, + false + ] + ], + [ + [ + -11.438902854919434, + false + ] + ], + [ + [ + -17.085294723510742, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.115461349487305, + false + ], + [ + -36.961116790771484, + false + ], + [ + -11.438902854919434, + false + ], + [ + -17.085294723510742, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b07c19a35114e750f431c9afff3917842b3eeca2335694d93e1b08fa508ab18c", + "prompt_hash": "3d93644794afb74c70f8e5e45b917cce4bef90e92d768a37c92d452ff74abdb9", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 2, + "doc": { + "question": "What type of materials are mentioned as enabling technologies for Fluid Antenna Systems (FAS)?", + "choices": [ + "Rigid conductors and ceramic insulators", + "Liquid materials, RF pixels, stepper motors, and metamaterials", + "Photonic crystals and superconductors", + "Fiber optics and quantum dots" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What type of materials are mentioned as enabling technologies for Fluid Antenna Systems (FAS)?\nAnswer:", + " Rigid conductors and ceramic insulators" + ], + [ + "Question:What type of materials are mentioned as enabling technologies for Fluid Antenna Systems (FAS)?\nAnswer:", + " Liquid materials, RF pixels, stepper motors, and metamaterials" + ], + [ + "Question:What type of materials are mentioned as enabling technologies for Fluid Antenna Systems (FAS)?\nAnswer:", + " Photonic crystals and superconductors" + ], + [ + "Question:What type of materials are mentioned as enabling technologies for Fluid Antenna Systems (FAS)?\nAnswer:", + " Fiber optics and quantum dots" + ] + ], + "resps": [ + [ + [ + -27.662734985351562, + false + ] + ], + [ + [ + -51.00507736206055, + false + ] + ], + [ + [ + -18.791711807250977, + false + ] + ], + [ + [ + -20.55025863647461, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.662734985351562, + false + ], + [ + -51.00507736206055, + false + ], + [ + -18.791711807250977, + false + ], + [ + -20.55025863647461, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6017248b4020a1bd9d1de89edc12d80bcccabe21f64eefb720c28c9ad7a63547", + "prompt_hash": "a3007907bbb72246ec5f8411bf73de8fa59f03a60c35a8d7808979f807b79cae", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 3, + "doc": { + "question": "Which technology was inspired by FAS and focuses only on position reconfiguration?", + "choices": [ + "Reconfigurable Intelligent Surfaces", + "Movable Antenna System", + "Non-Orthogonal Multiple Access", + "Continuous Aperture MIMO" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which technology was inspired by FAS and focuses only on position reconfiguration?\nAnswer:", + " Reconfigurable Intelligent Surfaces" + ], + [ + "Question:Which technology was inspired by FAS and focuses only on position reconfiguration?\nAnswer:", + " Movable Antenna System" + ], + [ + "Question:Which technology was inspired by FAS and focuses only on position reconfiguration?\nAnswer:", + " Non-Orthogonal Multiple Access" + ], + [ + "Question:Which technology was inspired by FAS and focuses only on position reconfiguration?\nAnswer:", + " Continuous Aperture MIMO" + ] + ], + "resps": [ + [ + [ + -9.414812088012695, + false + ] + ], + [ + [ + -19.87340545654297, + false + ] + ], + [ + [ + -18.9980411529541, + false + ] + ], + [ + [ + -27.375492095947266, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.414812088012695, + false + ], + [ + -19.87340545654297, + false + ], + [ + -18.9980411529541, + false + ], + [ + -27.375492095947266, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6b4441be7a721e85bf01c62181a759643acb5ef0cd1f1cc54ca828a513129e43", + "prompt_hash": "ec50620dcf3e101926f2c83ab625fbb096569d82df72f8e2e50fcf87d4266425", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 4, + "doc": { + "question": "According to the philosophy quoted by Bruce Lee, what is the core inspiration behind Fluid Antenna Systems?", + "choices": [ + "Rigidity and strength", + "Fixed structure and control", + "Formless adaptability like water", + "Speed and precision like lightning" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:According to the philosophy quoted by Bruce Lee, what is the core inspiration behind Fluid Antenna Systems?\nAnswer:", + " Rigidity and strength" + ], + [ + "Question:According to the philosophy quoted by Bruce Lee, what is the core inspiration behind Fluid Antenna Systems?\nAnswer:", + " Fixed structure and control" + ], + [ + "Question:According to the philosophy quoted by Bruce Lee, what is the core inspiration behind Fluid Antenna Systems?\nAnswer:", + " Formless adaptability like water" + ], + [ + "Question:According to the philosophy quoted by Bruce Lee, what is the core inspiration behind Fluid Antenna Systems?\nAnswer:", + " Speed and precision like lightning" + ] + ], + "resps": [ + [ + [ + -21.77372169494629, + false + ] + ], + [ + [ + -27.31595230102539, + false + ] + ], + [ + [ + -39.78109359741211, + false + ] + ], + [ + [ + -33.8327751159668, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.77372169494629, + false + ], + [ + -27.31595230102539, + false + ], + [ + -39.78109359741211, + false + ], + [ + -33.8327751159668, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "abb6f7b232a71ac5a3da433396f253d80a2ca0f59fef83d8656c6178f446c8ba", + "prompt_hash": "f109106228e3a9f23fe802fcf92d0018fd3135689f42b4af5faa4d5a81340df8", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 5, + "doc": { + "question": "What is the main reason the paper proposes an activation-based importance estimation strategy for pruning?", + "choices": [ + "To improve training efficiency by using backward passes", + "To reduce memory and computation cost by avoiding gradient computation", + "To ensure compatibility with reinforcement learning models", + "To enhance the convergence rate of the optimizer" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main reason the paper proposes an activation-based importance estimation strategy for pruning?\nAnswer:", + " To improve training efficiency by using backward passes" + ], + [ + "Question:What is the main reason the paper proposes an activation-based importance estimation strategy for pruning?\nAnswer:", + " To reduce memory and computation cost by avoiding gradient computation" + ], + [ + "Question:What is the main reason the paper proposes an activation-based importance estimation strategy for pruning?\nAnswer:", + " To ensure compatibility with reinforcement learning models" + ], + [ + "Question:What is the main reason the paper proposes an activation-based importance estimation strategy for pruning?\nAnswer:", + " To enhance the convergence rate of the optimizer" + ] + ], + "resps": [ + [ + [ + -29.56960678100586, + false + ] + ], + [ + [ + -29.971778869628906, + false + ] + ], + [ + [ + -26.012845993041992, + false + ] + ], + [ + [ + -21.306804656982422, + false + ] + ] + ], + "filtered_resps": [ + [ + -29.56960678100586, + false + ], + [ + -29.971778869628906, + false + ], + [ + -26.012845993041992, + false + ], + [ + -21.306804656982422, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9014bad4efb87b06a349240f8654f967c495a9730c9f52539dda10a679c02965", + "prompt_hash": "ed3aa775dea305a6c6700f5b0d5a921615db4c56bcc14fd274a6e9af802b1a19", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 6, + "doc": { + "question": "Which technique is used to estimate the importance of depth (layers) using cosine distance?", + "choices": [ + "Logit matching", + "Perplexity evaluation", + "Block Importance (BI)", + "LayerNorm scaling" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which technique is used to estimate the importance of depth (layers) using cosine distance?\nAnswer:", + " Logit matching" + ], + [ + "Question:Which technique is used to estimate the importance of depth (layers) using cosine distance?\nAnswer:", + " Perplexity evaluation" + ], + [ + "Question:Which technique is used to estimate the importance of depth (layers) using cosine distance?\nAnswer:", + " Block Importance (BI)" + ], + [ + "Question:Which technique is used to estimate the importance of depth (layers) using cosine distance?\nAnswer:", + " LayerNorm scaling" + ] + ], + "resps": [ + [ + [ + -18.129497528076172, + false + ] + ], + [ + [ + -19.97736358642578, + false + ] + ], + [ + [ + -22.130346298217773, + false + ] + ], + [ + [ + -20.870519638061523, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.129497528076172, + false + ], + [ + -19.97736358642578, + false + ], + [ + -22.130346298217773, + false + ], + [ + -20.870519638061523, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d2f0d78afdaa7596ecc6fdf6ae7488274a6dea3cd6a83d5eccfd065e4f6a5a54", + "prompt_hash": "9bd727639df70abc9f26f83569d56b87f3cbf59a45bb8e8a898532a74f7323c3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 7, + "doc": { + "question": "In the process of head pruning, what is added back to the remaining heads to preserve knowledge?", + "choices": [ + "Randomly initialized heads", + "Residual information from pruned heads", + "Gradients of the original heads", + "Normalized attention weights" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:In the process of head pruning, what is added back to the remaining heads to preserve knowledge?\nAnswer:", + " Randomly initialized heads" + ], + [ + "Question:In the process of head pruning, what is added back to the remaining heads to preserve knowledge?\nAnswer:", + " Residual information from pruned heads" + ], + [ + "Question:In the process of head pruning, what is added back to the remaining heads to preserve knowledge?\nAnswer:", + " Gradients of the original heads" + ], + [ + "Question:In the process of head pruning, what is added back to the remaining heads to preserve knowledge?\nAnswer:", + " Normalized attention weights" + ] + ], + "resps": [ + [ + [ + -16.398324966430664, + false + ] + ], + [ + [ + -20.017580032348633, + false + ] + ], + [ + [ + -17.852195739746094, + false + ] + ], + [ + [ + -19.09755516052246, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.398324966430664, + false + ], + [ + -20.017580032348633, + false + ], + [ + -17.852195739746094, + false + ], + [ + -19.09755516052246, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0d2983d11675001038546a0bcafe2426794fc9c33d4ffae94ed328d7c0b23971", + "prompt_hash": "9d567fc6e697e92f05881bf33f910659dee9ea08852f64601862c068d6ef4421", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 8, + "doc": { + "question": "How many samples are used in the calibration dataset for activation-based importance estimation?", + "choices": [ + "256", + "512", + "1024", + "2048" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many samples are used in the calibration dataset for activation-based importance estimation?\nAnswer:", + " 256" + ], + [ + "Question:How many samples are used in the calibration dataset for activation-based importance estimation?\nAnswer:", + " 512" + ], + [ + "Question:How many samples are used in the calibration dataset for activation-based importance estimation?\nAnswer:", + " 1024" + ], + [ + "Question:How many samples are used in the calibration dataset for activation-based importance estimation?\nAnswer:", + " 2048" + ] + ], + "resps": [ + [ + [ + -6.168458938598633, + false + ] + ], + [ + [ + -6.1954240798950195, + false + ] + ], + [ + [ + -6.884011268615723, + false + ] + ], + [ + [ + -7.902861595153809, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.168458938598633, + false + ], + [ + -6.1954240798950195, + false + ], + [ + -6.884011268615723, + false + ], + [ + -7.902861595153809, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e6b013c5031193ef5a41ed4462d5c0de4a10b762a1051b60ea7c32e67437ceea", + "prompt_hash": "5f52b44600d809d763b29b4bef6ad0ca70bcc2189ba974d2ef17ec03d844078c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 9, + "doc": { + "question": "Which loss function is found to be most effective for logit-based knowledge distillation in this work?", + "choices": [ + "Mean Squared Error (MSE)", + "Reverse KL Divergence (R-KLD)", + "Cosine Similarity", + "Kullback-Leibler Divergence (KLD)" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which loss function is found to be most effective for logit-based knowledge distillation in this work?\nAnswer:", + " Mean Squared Error (MSE)" + ], + [ + "Question:Which loss function is found to be most effective for logit-based knowledge distillation in this work?\nAnswer:", + " Reverse KL Divergence (R-KLD)" + ], + [ + "Question:Which loss function is found to be most effective for logit-based knowledge distillation in this work?\nAnswer:", + " Cosine Similarity" + ], + [ + "Question:Which loss function is found to be most effective for logit-based knowledge distillation in this work?\nAnswer:", + " Kullback-Leibler Divergence (KLD)" + ] + ], + "resps": [ + [ + [ + -8.091793060302734, + false + ] + ], + [ + [ + -20.987077713012695, + false + ] + ], + [ + [ + -12.251108169555664, + false + ] + ], + [ + [ + -10.422445297241211, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.091793060302734, + false + ], + [ + -20.987077713012695, + false + ], + [ + -12.251108169555664, + false + ], + [ + -10.422445297241211, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b146330c14a52f1e5eb3ad27d75a0d6aa712be86ed927041fc598cbca1e50c78", + "prompt_hash": "291f73cd7fd6e150e50b0088ff0ed3c461b12aaaf85b6b5e22ba43ccfdd17b68", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 10, + "doc": { + "question": "What is the generating velocity formula for discrete Flow Matching using denoiser parameterization?", + "choices": [ + "u^i_t(x^i,z) = \\dot{\\kappa}_t(x^i - z^i)", + "u^i_t(x^i,z) = \\frac{\\dot{\\kappa}_t}{\\kappa_t}(\\delta_{z}(x^i) - p_{1|t}(x^i|z))", + "u^i_t(x^i,z) = \\frac{\\dot{\\kappa}_t}{1-\\kappa_t}(p_{1|t}(x^i|z) - \\delta_{z}(x^i))", + "u^i_t(x^i,z) = \\dot{\\kappa}_t(x^i + z^i)" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the generating velocity formula for discrete Flow Matching using denoiser parameterization?\nAnswer:", + " u^i_t(x^i,z) = \\dot{\\kappa}_t(x^i - z^i)" + ], + [ + "Question:What is the generating velocity formula for discrete Flow Matching using denoiser parameterization?\nAnswer:", + " u^i_t(x^i,z) = \\frac{\\dot{\\kappa}_t}{\\kappa_t}(\\delta_{z}(x^i) - p_{1|t}(x^i|z))" + ], + [ + "Question:What is the generating velocity formula for discrete Flow Matching using denoiser parameterization?\nAnswer:", + " u^i_t(x^i,z) = \\frac{\\dot{\\kappa}_t}{1-\\kappa_t}(p_{1|t}(x^i|z) - \\delta_{z}(x^i))" + ], + [ + "Question:What is the generating velocity formula for discrete Flow Matching using denoiser parameterization?\nAnswer:", + " u^i_t(x^i,z) = \\dot{\\kappa}_t(x^i + z^i)" + ] + ], + "resps": [ + [ + [ + -68.01116943359375, + false + ] + ], + [ + [ + -107.26423645019531, + false + ] + ], + [ + [ + -109.30007934570312, + false + ] + ], + [ + [ + -70.12492370605469, + false + ] + ] + ], + "filtered_resps": [ + [ + -68.01116943359375, + false + ], + [ + -107.26423645019531, + false + ], + [ + -109.30007934570312, + false + ], + [ + -70.12492370605469, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3e572071690cd40ec828547c492d433d58578af3dcab3e6d254a6a1e417c21d4", + "prompt_hash": "542d77d6830a44a1fe70f95d36a61a73a17813946366d9c754e76dbb5894e066", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 11, + "doc": { + "question": "Which of the following is a key condition for the discrete probability velocity to define a proper PMF?", + "choices": [ + "u_t^i(x^i,z) must be negative for all i and x^i ≠ z^i", + "The sum over i of u_t^i(x^i,z) must be zero", + "u_t^i(x^i,z) must be symmetric in x^i and z", + "The sum over x^i of u_t^i(x^i,z) must be zero and u_t^i(x^i,z) ≥ 0 for x^i ≠ z^i" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which of the following is a key condition for the discrete probability velocity to define a proper PMF?\nAnswer:", + " u_t^i(x^i,z) must be negative for all i and x^i ≠ z^i" + ], + [ + "Question:Which of the following is a key condition for the discrete probability velocity to define a proper PMF?\nAnswer:", + " The sum over i of u_t^i(x^i,z) must be zero" + ], + [ + "Question:Which of the following is a key condition for the discrete probability velocity to define a proper PMF?\nAnswer:", + " u_t^i(x^i,z) must be symmetric in x^i and z" + ], + [ + "Question:Which of the following is a key condition for the discrete probability velocity to define a proper PMF?\nAnswer:", + " The sum over x^i of u_t^i(x^i,z) must be zero and u_t^i(x^i,z) ≥ 0 for x^i ≠ z^i" + ] + ], + "resps": [ + [ + [ + -91.10096740722656, + false + ] + ], + [ + [ + -65.13204956054688, + false + ] + ], + [ + [ + -68.94792175292969, + false + ] + ], + [ + [ + -109.533935546875, + false + ] + ] + ], + "filtered_resps": [ + [ + -91.10096740722656, + false + ], + [ + -65.13204956054688, + false + ], + [ + -68.94792175292969, + false + ], + [ + -109.533935546875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0c766ba70e3c44cffef5e20b395c9e59dc2a99656fe7a6996a7f569c868c52b6", + "prompt_hash": "b21dc9d9aa57995066b718add6746175c197930d2f46fb39ae3d599309277b9f", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 12, + "doc": { + "question": "In the experiments, which source distribution is primarily used for training?", + "choices": [ + "Uniform over all tokens", + "Gaussian noise", + "All-mask sequences \\(\\delta_\\dummy(x)\\)", + "Random binary sequences" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In the experiments, which source distribution is primarily used for training?\nAnswer:", + " Uniform over all tokens" + ], + [ + "Question:In the experiments, which source distribution is primarily used for training?\nAnswer:", + " Gaussian noise" + ], + [ + "Question:In the experiments, which source distribution is primarily used for training?\nAnswer:", + " All-mask sequences \\(\\delta_\\dummy(x)\\)" + ], + [ + "Question:In the experiments, which source distribution is primarily used for training?\nAnswer:", + " Random binary sequences" + ] + ], + "resps": [ + [ + [ + -28.035892486572266, + false + ] + ], + [ + [ + -14.957950592041016, + false + ] + ], + [ + [ + -76.81466674804688, + false + ] + ], + [ + [ + -23.489944458007812, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.035892486572266, + false + ], + [ + -14.957950592041016, + false + ], + [ + -76.81466674804688, + false + ], + [ + -23.489944458007812, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6a8ff128da3cb5032b99962c3ace2ce14de5c804eb4bf8537a559ca3760ddc3b", + "prompt_hash": "e5faa1d4061d60bea510fb45889de76a5d0e821c8966fa860c43c14ddb8bd409", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 13, + "doc": { + "question": "What is the entropy score achieved by the 1.7B Discrete Flow Matching model with 1024 NFE on unconditional text generation?", + "choices": [ + "7.1", + "7.6", + "8.3", + "6.7" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the entropy score achieved by the 1.7B Discrete Flow Matching model with 1024 NFE on unconditional text generation?\nAnswer:", + " 7.1" + ], + [ + "Question:What is the entropy score achieved by the 1.7B Discrete Flow Matching model with 1024 NFE on unconditional text generation?\nAnswer:", + " 7.6" + ], + [ + "Question:What is the entropy score achieved by the 1.7B Discrete Flow Matching model with 1024 NFE on unconditional text generation?\nAnswer:", + " 8.3" + ], + [ + "Question:What is the entropy score achieved by the 1.7B Discrete Flow Matching model with 1024 NFE on unconditional text generation?\nAnswer:", + " 6.7" + ] + ], + "resps": [ + [ + [ + -8.013721466064453, + false + ] + ], + [ + [ + -7.946964263916016, + false + ] + ], + [ + [ + -7.716164588928223, + false + ] + ], + [ + [ + -7.653365135192871, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.013721466064453, + false + ], + [ + -7.946964263916016, + false + ], + [ + -7.716164588928223, + false + ], + [ + -7.653365135192871, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f104cfef0729719cf261c81dd1aae450ed57ebe3dac02c6d8f0dfa014afde700", + "prompt_hash": "ddcad478860bed220703544578375a51f2ee47123b98690d6ebd8e8b610d1433", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 14, + "doc": { + "question": "What does the conditional probability path p_t(x^i|x_0,x_1) become in the simplest convex interpolant case?", + "choices": [ + "A uniform distribution over [d]", + "A Gaussian centered between x_0 and x_1", + "(1-\\kappa_t)\\delta_{x_0}(x^i) + \\kappa_t \\delta_{x_1}(x^i)", + "An average of multiple learned posteriors" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does the conditional probability path p_t(x^i|x_0,x_1) become in the simplest convex interpolant case?\nAnswer:", + " A uniform distribution over [d]" + ], + [ + "Question:What does the conditional probability path p_t(x^i|x_0,x_1) become in the simplest convex interpolant case?\nAnswer:", + " A Gaussian centered between x_0 and x_1" + ], + [ + "Question:What does the conditional probability path p_t(x^i|x_0,x_1) become in the simplest convex interpolant case?\nAnswer:", + " (1-\\kappa_t)\\delta_{x_0}(x^i) + \\kappa_t \\delta_{x_1}(x^i)" + ], + [ + "Question:What does the conditional probability path p_t(x^i|x_0,x_1) become in the simplest convex interpolant case?\nAnswer:", + " An average of multiple learned posteriors" + ] + ], + "resps": [ + [ + [ + -28.31566619873047, + false + ] + ], + [ + [ + -26.569395065307617, + false + ] + ], + [ + [ + -34.784271240234375, + false + ] + ], + [ + [ + -38.008331298828125, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.31566619873047, + false + ], + [ + -26.569395065307617, + false + ], + [ + -34.784271240234375, + false + ], + [ + -38.008331298828125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "32da34d1bd40cfa82315f3867a08cad8b4592f25737cf0108a9bd7364027f0ca", + "prompt_hash": "5ddd5f5bd1e88d6f2e9385a4f2dd16cc85ed382780f3729120582bd6d1d6dfb3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 15, + "doc": { + "question": "Which library is described as a minimalistic implementation focused on transparency and simplicity in reinforcement learning algorithms?", + "choices": [ + "A. Stable Baselines3", + "B. CleanRL", + "C. Ray Rllib", + "D. Dopamine" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which library is described as a minimalistic implementation focused on transparency and simplicity in reinforcement learning algorithms?\nAnswer:", + " A. Stable Baselines3" + ], + [ + "Question:Which library is described as a minimalistic implementation focused on transparency and simplicity in reinforcement learning algorithms?\nAnswer:", + " B. CleanRL" + ], + [ + "Question:Which library is described as a minimalistic implementation focused on transparency and simplicity in reinforcement learning algorithms?\nAnswer:", + " C. Ray Rllib" + ], + [ + "Question:Which library is described as a minimalistic implementation focused on transparency and simplicity in reinforcement learning algorithms?\nAnswer:", + " D. Dopamine" + ] + ], + "resps": [ + [ + [ + -18.61655616760254, + false + ] + ], + [ + [ + -22.445934295654297, + false + ] + ], + [ + [ + -30.368227005004883, + false + ] + ], + [ + [ + -24.299724578857422, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.61655616760254, + false + ], + [ + -22.445934295654297, + false + ], + [ + -30.368227005004883, + false + ], + [ + -24.299724578857422, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "904ecd6be1774e67177b9cf34468c95060f6fd6625165af4d0adb6d7c2452bfa", + "prompt_hash": "42a2094e37c0ea21296696b3a2983f73a082205c23e7bd7636ce07eac786bce3", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 16, + "doc": { + "question": "What is the purpose of the 'TimeLimit' wrapper in Gymnasium?", + "choices": [ + "A. To handle multi-agent environments", + "B. To limit the number of parallel environments", + "C. To enforce episode truncation after a set number of steps", + "D. To restrict the size of the action space" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the purpose of the 'TimeLimit' wrapper in Gymnasium?\nAnswer:", + " A. To handle multi-agent environments" + ], + [ + "Question:What is the purpose of the 'TimeLimit' wrapper in Gymnasium?\nAnswer:", + " B. To limit the number of parallel environments" + ], + [ + "Question:What is the purpose of the 'TimeLimit' wrapper in Gymnasium?\nAnswer:", + " C. To enforce episode truncation after a set number of steps" + ], + [ + "Question:What is the purpose of the 'TimeLimit' wrapper in Gymnasium?\nAnswer:", + " D. To restrict the size of the action space" + ] + ], + "resps": [ + [ + [ + -26.94652557373047, + false + ] + ], + [ + [ + -30.293476104736328, + false + ] + ], + [ + [ + -44.030799865722656, + false + ] + ], + [ + [ + -26.48984146118164, + false + ] + ] + ], + "filtered_resps": [ + [ + -26.94652557373047, + false + ], + [ + -30.293476104736328, + false + ], + [ + -44.030799865722656, + false + ], + [ + -26.48984146118164, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "85dc42ac9b74820219e7e7e66f1c3d98d606738371118a8b8472288595de8966", + "prompt_hash": "fe682fd0f2002613a9b1ca97e74c720df39b12300931b0dac7f60a025774e8ed", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 17, + "doc": { + "question": "Which composite space in Gymnasium is equivalent to a sum type in algebraic data types?", + "choices": [ + "A. Tuple", + "B. Dict", + "C. Graph", + "D. OneOf" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which composite space in Gymnasium is equivalent to a sum type in algebraic data types?\nAnswer:", + " A. Tuple" + ], + [ + "Question:Which composite space in Gymnasium is equivalent to a sum type in algebraic data types?\nAnswer:", + " B. Dict" + ], + [ + "Question:Which composite space in Gymnasium is equivalent to a sum type in algebraic data types?\nAnswer:", + " C. Graph" + ], + [ + "Question:Which composite space in Gymnasium is equivalent to a sum type in algebraic data types?\nAnswer:", + " D. OneOf" + ] + ], + "resps": [ + [ + [ + -14.58735466003418, + false + ] + ], + [ + [ + -19.29799461364746, + false + ] + ], + [ + [ + -15.71833610534668, + false + ] + ], + [ + [ + -24.84959602355957, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.58735466003418, + false + ], + [ + -19.29799461364746, + false + ], + [ + -15.71833610534668, + false + ], + [ + -24.84959602355957, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0d19232b2ba1f0667f11f8fbd617daf13ecfc4f70863a1719cab8bb40964130d", + "prompt_hash": "81d03b60cac6606bfc09a29327254fd491995841dfafbe33e0fd2b9d640bc77f", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 18, + "doc": { + "question": "What advantage does FuncEnv offer over the standard Env in Gymnasium?", + "choices": [ + "A. It supports more games", + "B. It enables higher resolution rendering", + "C. It allows easy hardware acceleration and theoretical alignment with POMDPs", + "D. It reduces the number of required libraries" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What advantage does FuncEnv offer over the standard Env in Gymnasium?\nAnswer:", + " A. It supports more games" + ], + [ + "Question:What advantage does FuncEnv offer over the standard Env in Gymnasium?\nAnswer:", + " B. It enables higher resolution rendering" + ], + [ + "Question:What advantage does FuncEnv offer over the standard Env in Gymnasium?\nAnswer:", + " C. It allows easy hardware acceleration and theoretical alignment with POMDPs" + ], + [ + "Question:What advantage does FuncEnv offer over the standard Env in Gymnasium?\nAnswer:", + " D. It reduces the number of required libraries" + ] + ], + "resps": [ + [ + [ + -25.4164981842041, + false + ] + ], + [ + [ + -31.929250717163086, + false + ] + ], + [ + [ + -67.37519836425781, + false + ] + ], + [ + [ + -33.48160934448242, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.4164981842041, + false + ], + [ + -31.929250717163086, + false + ], + [ + -67.37519836425781, + false + ], + [ + -33.48160934448242, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9d21d8c4990fa96ac7bb357064d95a6aaa9aa93bda23c77b06d141f55f93b696", + "prompt_hash": "575fd9b7ee03779f48f5051f52088719a892c99cb7a8dc3824fb9254d66e992c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 19, + "doc": { + "question": "Which environment category in Gymnasium includes simulations using the Box2D physics engine?", + "choices": [ + "A. Toy text", + "B. Classic control", + "C. Mujoco", + "D. Box2D" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which environment category in Gymnasium includes simulations using the Box2D physics engine?\nAnswer:", + " A. Toy text" + ], + [ + "Question:Which environment category in Gymnasium includes simulations using the Box2D physics engine?\nAnswer:", + " B. Classic control" + ], + [ + "Question:Which environment category in Gymnasium includes simulations using the Box2D physics engine?\nAnswer:", + " C. Mujoco" + ], + [ + "Question:Which environment category in Gymnasium includes simulations using the Box2D physics engine?\nAnswer:", + " D. Box2D" + ] + ], + "resps": [ + [ + [ + -26.217872619628906, + false + ] + ], + [ + [ + -25.281362533569336, + false + ] + ], + [ + [ + -17.874332427978516, + false + ] + ], + [ + [ + -16.03414535522461, + false + ] + ] + ], + "filtered_resps": [ + [ + -26.217872619628906, + false + ], + [ + -25.281362533569336, + false + ], + [ + -17.874332427978516, + false + ], + [ + -16.03414535522461, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b663e101e05e8daa81280fdbf27f747086424f897b25e4522d58408114cda7e5", + "prompt_hash": "bc815e7a9b1fbdfda309fdbb37ff1f4e46139a48cbe3f87b0b52013093017bfa", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 20, + "doc": { + "question": "What is the main function of the KV Cache in a Transformer model?", + "choices": [ + "To store the attention scores for future layers", + "To reduce the time complexity by storing keys and values from past tokens", + "To improve model accuracy by storing gradients", + "To keep track of the vocabulary index for each token" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main function of the KV Cache in a Transformer model?\nAnswer:", + " To store the attention scores for future layers" + ], + [ + "Question:What is the main function of the KV Cache in a Transformer model?\nAnswer:", + " To reduce the time complexity by storing keys and values from past tokens" + ], + [ + "Question:What is the main function of the KV Cache in a Transformer model?\nAnswer:", + " To improve model accuracy by storing gradients" + ], + [ + "Question:What is the main function of the KV Cache in a Transformer model?\nAnswer:", + " To keep track of the vocabulary index for each token" + ] + ], + "resps": [ + [ + [ + -25.331092834472656, + false + ] + ], + [ + [ + -33.660465240478516, + false + ] + ], + [ + [ + -26.1849308013916, + false + ] + ], + [ + [ + -31.01213836669922, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.331092834472656, + false + ], + [ + -33.660465240478516, + false + ], + [ + -26.1849308013916, + false + ], + [ + -31.01213836669922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3463744ea89b2f25f6539f138a7e4232ea0c8fad55aa5d2f6d1d9c1a0f621c31", + "prompt_hash": "6c634aebb08ad3543ecc982fee051e3bfae803b63d3e7dd0236990c4e853f2ce", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 21, + "doc": { + "question": "What happens to the size of the KV Cache as the input sequence grows?", + "choices": [ + "It remains constant regardless of sequence length", + "It decreases exponentially with sequence length", + "It increases linearly with sequence length", + "It increases quadratically with sequence length" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What happens to the size of the KV Cache as the input sequence grows?\nAnswer:", + " It remains constant regardless of sequence length" + ], + [ + "Question:What happens to the size of the KV Cache as the input sequence grows?\nAnswer:", + " It decreases exponentially with sequence length" + ], + [ + "Question:What happens to the size of the KV Cache as the input sequence grows?\nAnswer:", + " It increases linearly with sequence length" + ], + [ + "Question:What happens to the size of the KV Cache as the input sequence grows?\nAnswer:", + " It increases quadratically with sequence length" + ] + ], + "resps": [ + [ + [ + -15.243741035461426, + false + ] + ], + [ + [ + -13.885777473449707, + false + ] + ], + [ + [ + -11.21221923828125, + false + ] + ], + [ + [ + -15.224578857421875, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.243741035461426, + false + ], + [ + -13.885777473449707, + false + ], + [ + -11.21221923828125, + false + ], + [ + -15.224578857421875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "118ddfec11dcb0fe65b26ed3a40af3d40dadba30da3beb6bd65e7c2fb508ef47", + "prompt_hash": "a91b09a7e82745c745c163096df4fdbbbc90dd0a172550d69395461d43a5b5ba", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 22, + "doc": { + "question": "Which method divides query heads into groups to share key heads for improved KV Cache efficiency?", + "choices": [ + "Multi-Query Attention (MQA)", + "Grouped Query Attention (GQA)", + "Dynamic Memory Compression (DMC)", + "Multi-Head Latent Attention (MLA)" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which method divides query heads into groups to share key heads for improved KV Cache efficiency?\nAnswer:", + " Multi-Query Attention (MQA)" + ], + [ + "Question:Which method divides query heads into groups to share key heads for improved KV Cache efficiency?\nAnswer:", + " Grouped Query Attention (GQA)" + ], + [ + "Question:Which method divides query heads into groups to share key heads for improved KV Cache efficiency?\nAnswer:", + " Dynamic Memory Compression (DMC)" + ], + [ + "Question:Which method divides query heads into groups to share key heads for improved KV Cache efficiency?\nAnswer:", + " Multi-Head Latent Attention (MLA)" + ] + ], + "resps": [ + [ + [ + -14.856151580810547, + false + ] + ], + [ + [ + -8.729805946350098, + false + ] + ], + [ + [ + -22.34358787536621, + false + ] + ], + [ + [ + -27.207630157470703, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.856151580810547, + false + ], + [ + -8.729805946350098, + false + ], + [ + -22.34358787536621, + false + ], + [ + -27.207630157470703, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "90a08f669aa49d769ceac60983fd76066b90276bee096b610e9a9996f44c3983", + "prompt_hash": "1d846ff11639e33d06c3be8b2e5f27d7eb7d5328d7836aba456472da5e3efcee", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 23, + "doc": { + "question": "What is the key idea behind Token Omission Via Attention (TOVA)?", + "choices": [ + "To discard tokens based on entropy scores", + "To merge token representations based on Gaussian weighting", + "To evict tokens with minimal attention weights layer-wise", + "To compress keys using low-rank factorization" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the key idea behind Token Omission Via Attention (TOVA)?\nAnswer:", + " To discard tokens based on entropy scores" + ], + [ + "Question:What is the key idea behind Token Omission Via Attention (TOVA)?\nAnswer:", + " To merge token representations based on Gaussian weighting" + ], + [ + "Question:What is the key idea behind Token Omission Via Attention (TOVA)?\nAnswer:", + " To evict tokens with minimal attention weights layer-wise" + ], + [ + "Question:What is the key idea behind Token Omission Via Attention (TOVA)?\nAnswer:", + " To compress keys using low-rank factorization" + ] + ], + "resps": [ + [ + [ + -30.414817810058594, + false + ] + ], + [ + [ + -41.22924041748047, + false + ] + ], + [ + [ + -42.73115921020508, + false + ] + ], + [ + [ + -37.3741455078125, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.414817810058594, + false + ], + [ + -41.22924041748047, + false + ], + [ + -42.73115921020508, + false + ], + [ + -37.3741455078125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "aecb2b4b9d87e1ec1445221a5e45e21be8e240abab840cef2d7f455ac9e5d3b6", + "prompt_hash": "e853258d32fcf11b3960428130f3102bd970cf4ea7b6fc8680e79fee800d0b38", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 24, + "doc": { + "question": "Which quantization method applies different strategies to key and value caches and uses an attention window to guide quantization?", + "choices": [ + "FlexGen", + "WKVQuant", + "QAQ", + "GEAR" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which quantization method applies different strategies to key and value caches and uses an attention window to guide quantization?\nAnswer:", + " FlexGen" + ], + [ + "Question:Which quantization method applies different strategies to key and value caches and uses an attention window to guide quantization?\nAnswer:", + " WKVQuant" + ], + [ + "Question:Which quantization method applies different strategies to key and value caches and uses an attention window to guide quantization?\nAnswer:", + " QAQ" + ], + [ + "Question:Which quantization method applies different strategies to key and value caches and uses an attention window to guide quantization?\nAnswer:", + " GEAR" + ] + ], + "resps": [ + [ + [ + -15.81790542602539, + false + ] + ], + [ + [ + -21.988216400146484, + false + ] + ], + [ + [ + -12.134068489074707, + false + ] + ], + [ + [ + -15.453916549682617, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.81790542602539, + false + ], + [ + -21.988216400146484, + false + ], + [ + -12.134068489074707, + false + ], + [ + -15.453916549682617, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9b9ec799f0644347f00ccdc559a2de9b2950ecdd38505a1eea0fa43b6e678612", + "prompt_hash": "7bd19b063a0a24820d40409d88ce869b3d85a9e2382085f4f1e8098cef48240c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 25, + "doc": { + "question": "What evaluation metric is used for the HumanEval benchmark?", + "choices": [ + "Accuracy", + "F1 Score", + "Pass@1", + "BLEU" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What evaluation metric is used for the HumanEval benchmark?\nAnswer:", + " Accuracy" + ], + [ + "Question:What evaluation metric is used for the HumanEval benchmark?\nAnswer:", + " F1 Score" + ], + [ + "Question:What evaluation metric is used for the HumanEval benchmark?\nAnswer:", + " Pass@1" + ], + [ + "Question:What evaluation metric is used for the HumanEval benchmark?\nAnswer:", + " BLEU" + ] + ], + "resps": [ + [ + [ + -5.508859157562256, + false + ] + ], + [ + [ + -7.320566177368164, + false + ] + ], + [ + [ + -6.618783473968506, + false + ] + ], + [ + [ + -6.3945088386535645, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.508859157562256, + false + ], + [ + -7.320566177368164, + false + ], + [ + -6.618783473968506, + false + ], + [ + -6.3945088386535645, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "40b7639a11a2d6401329ccab248911a0b42df0e8dc28672aae354083c26f73e8", + "prompt_hash": "36ddf66b2a490ca68d938f5ebc940a8b85fcd8dc65d94c9a32f6e976909f94cf", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 26, + "doc": { + "question": "How many samples are in the MMLU benchmark?", + "choices": [ + "5000", + "14042", + "1954", + "87048" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many samples are in the MMLU benchmark?\nAnswer:", + " 5000" + ], + [ + "Question:How many samples are in the MMLU benchmark?\nAnswer:", + " 14042" + ], + [ + "Question:How many samples are in the MMLU benchmark?\nAnswer:", + " 1954" + ], + [ + "Question:How many samples are in the MMLU benchmark?\nAnswer:", + " 87048" + ] + ], + "resps": [ + [ + [ + -5.450283527374268, + false + ] + ], + [ + [ + -13.787731170654297, + false + ] + ], + [ + [ + -11.449134826660156, + false + ] + ], + [ + [ + -16.09764862060547, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.450283527374268, + false + ], + [ + -13.787731170654297, + false + ], + [ + -11.449134826660156, + false + ], + [ + -16.09764862060547, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ba828c25f156de6325888a9da9951b77820404aaadb21d08e34211803a90f4b2", + "prompt_hash": "8210b467843557958d435189a7ded66d735a1f9a89ee4a09fd5f1ba9c2fa1cf1", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 27, + "doc": { + "question": "Which benchmark uses the 'MC2_score' metric?", + "choices": [ + "TruthfulQA", + "AGIEval", + "GSM8k", + "SiQA" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which benchmark uses the 'MC2_score' metric?\nAnswer:", + " TruthfulQA" + ], + [ + "Question:Which benchmark uses the 'MC2_score' metric?\nAnswer:", + " AGIEval" + ], + [ + "Question:Which benchmark uses the 'MC2_score' metric?\nAnswer:", + " GSM8k" + ], + [ + "Question:Which benchmark uses the 'MC2_score' metric?\nAnswer:", + " SiQA" + ] + ], + "resps": [ + [ + [ + -12.017866134643555, + false + ] + ], + [ + [ + -11.988103866577148, + false + ] + ], + [ + [ + -9.616325378417969, + false + ] + ], + [ + [ + -19.542951583862305, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.017866134643555, + false + ], + [ + -11.988103866577148, + false + ], + [ + -9.616325378417969, + false + ], + [ + -19.542951583862305, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3fcc58cfc3351602d8c12fcc817be59779b123a9aefe8e79e34d8584309fa707", + "prompt_hash": "1d5a9234f36bec2a977788235df16dc92ff6c526203551a7286ba6e227ea9586", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 28, + "doc": { + "question": "What is the maximum number of few-shot examples used in GSM8k?", + "choices": [ + "3", + "5", + "8", + "0" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the maximum number of few-shot examples used in GSM8k?\nAnswer:", + " 3" + ], + [ + "Question:What is the maximum number of few-shot examples used in GSM8k?\nAnswer:", + " 5" + ], + [ + "Question:What is the maximum number of few-shot examples used in GSM8k?\nAnswer:", + " 8" + ], + [ + "Question:What is the maximum number of few-shot examples used in GSM8k?\nAnswer:", + " 0" + ] + ], + "resps": [ + [ + [ + -3.252291202545166, + false + ] + ], + [ + [ + -2.7015891075134277, + false + ] + ], + [ + [ + -3.2675089836120605, + false + ] + ], + [ + [ + -6.1405487060546875, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.252291202545166, + false + ], + [ + -2.7015891075134277, + false + ], + [ + -3.2675089836120605, + false + ], + [ + -6.1405487060546875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6402abb438cc24d2f202a81faf00c17046e16338291e66758b64a58fae4ab0eb", + "prompt_hash": "3327f679ce2b69e15b832d37f8c391899a7a09786e0265e26494977d0bf82a18", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 29, + "doc": { + "question": "Which dataset includes 5000 samples and uses 'EM maj1@1' as its metric?", + "choices": [ + "MATH", + "MBPP", + "PiQA", + "OpenBookQA" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which dataset includes 5000 samples and uses 'EM maj1@1' as its metric?\nAnswer:", + " MATH" + ], + [ + "Question:Which dataset includes 5000 samples and uses 'EM maj1@1' as its metric?\nAnswer:", + " MBPP" + ], + [ + "Question:Which dataset includes 5000 samples and uses 'EM maj1@1' as its metric?\nAnswer:", + " PiQA" + ], + [ + "Question:Which dataset includes 5000 samples and uses 'EM maj1@1' as its metric?\nAnswer:", + " OpenBookQA" + ] + ], + "resps": [ + [ + [ + -12.23568058013916, + false + ] + ], + [ + [ + -12.814970016479492, + false + ] + ], + [ + [ + -16.764488220214844, + false + ] + ], + [ + [ + -18.32901954650879, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.23568058013916, + false + ], + [ + -12.814970016479492, + false + ], + [ + -16.764488220214844, + false + ], + [ + -18.32901954650879, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "22fc097dd2c97659d37f74fb56bb16765fb7a8e35b418b058170082a53c3f80f", + "prompt_hash": "f2fc571d9e327979e9e02ac55b91d28937cf446a2c28ceedc0b42b5e9713ec5d", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 30, + "doc": { + "question": "What is Mini-Omni described as?", + "choices": [ + "A closed-source multimodal model", + "The first open-source end-to-end multimodal model with real-time speech capabilities", + "A text-only language model", + "A dataset for speech assistants" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is Mini-Omni described as?\nAnswer:", + " A closed-source multimodal model" + ], + [ + "Question:What is Mini-Omni described as?\nAnswer:", + " The first open-source end-to-end multimodal model with real-time speech capabilities" + ], + [ + "Question:What is Mini-Omni described as?\nAnswer:", + " A text-only language model" + ], + [ + "Question:What is Mini-Omni described as?\nAnswer:", + " A dataset for speech assistants" + ] + ], + "resps": [ + [ + [ + -20.025798797607422, + false + ] + ], + [ + [ + -49.17539978027344, + false + ] + ], + [ + [ + -21.924789428710938, + false + ] + ], + [ + [ + -28.139144897460938, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.025798797607422, + false + ], + [ + -49.17539978027344, + false + ], + [ + -21.924789428710938, + false + ], + [ + -28.139144897460938, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a2b690276c4d1ed0d8b85440c4630674bd06ec2ca989bb2c0cb93a9d1352d366", + "prompt_hash": "41e98cf04516b6ad4fae7b0d5180fc01b236a5fb254cd03e7589e27a34380887", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 31, + "doc": { + "question": "Which audio encoder was selected to ensure high-quality audio in Mini-Omni?", + "choices": [ + "HuBERT", + "WavLM", + "SNAC", + "SpearTTS" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which audio encoder was selected to ensure high-quality audio in Mini-Omni?\nAnswer:", + " HuBERT" + ], + [ + "Question:Which audio encoder was selected to ensure high-quality audio in Mini-Omni?\nAnswer:", + " WavLM" + ], + [ + "Question:Which audio encoder was selected to ensure high-quality audio in Mini-Omni?\nAnswer:", + " SNAC" + ], + [ + "Question:Which audio encoder was selected to ensure high-quality audio in Mini-Omni?\nAnswer:", + " SpearTTS" + ] + ], + "resps": [ + [ + [ + -6.846920490264893, + false + ] + ], + [ + [ + -4.520740509033203, + false + ] + ], + [ + [ + -16.42870330810547, + false + ] + ], + [ + [ + -18.480022430419922, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.846920490264893, + false + ], + [ + -4.520740509033203, + false + ], + [ + -16.42870330810547, + false + ], + [ + -18.480022430419922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "500b61420e80d825b160ea80c086d69b09dbf803be743dc13dac871587d6b136", + "prompt_hash": "1e0185e543741408a3054af6df183379519be7f31028489dae0044fe6abe0e5a", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 32, + "doc": { + "question": "What is the size of the base language model used in Mini-Omni?", + "choices": [ + "7 billion parameters", + "13 billion parameters", + "0.5 billion parameters", + "2 billion parameters" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the size of the base language model used in Mini-Omni?\nAnswer:", + " 7 billion parameters" + ], + [ + "Question:What is the size of the base language model used in Mini-Omni?\nAnswer:", + " 13 billion parameters" + ], + [ + "Question:What is the size of the base language model used in Mini-Omni?\nAnswer:", + " 0.5 billion parameters" + ], + [ + "Question:What is the size of the base language model used in Mini-Omni?\nAnswer:", + " 2 billion parameters" + ] + ], + "resps": [ + [ + [ + -7.209259510040283, + false + ] + ], + [ + [ + -8.974302291870117, + false + ] + ], + [ + [ + -11.013339042663574, + false + ] + ], + [ + [ + -8.469942092895508, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.209259510040283, + false + ], + [ + -8.974302291870117, + false + ], + [ + -11.013339042663574, + false + ], + [ + -8.469942092895508, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d4000b0545a7feb4021e6fc2823d64ab08bbedd1417b1397cde5ec9df741909d", + "prompt_hash": "0de74d67eaa0ed7054024f50cdde5a742db9019e2f5bf1aea009ad7989472de0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 33, + "doc": { + "question": "What is the name of the dataset introduced to fine-tune models into a voice assistant style?", + "choices": [ + "LibriSpeech", + "VoiceAssistant-400K", + "Open-Orca", + "Moss-SFT" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the name of the dataset introduced to fine-tune models into a voice assistant style?\nAnswer:", + " LibriSpeech" + ], + [ + "Question:What is the name of the dataset introduced to fine-tune models into a voice assistant style?\nAnswer:", + " VoiceAssistant-400K" + ], + [ + "Question:What is the name of the dataset introduced to fine-tune models into a voice assistant style?\nAnswer:", + " Open-Orca" + ], + [ + "Question:What is the name of the dataset introduced to fine-tune models into a voice assistant style?\nAnswer:", + " Moss-SFT" + ] + ], + "resps": [ + [ + [ + -6.687619209289551, + false + ] + ], + [ + [ + -20.196495056152344, + false + ] + ], + [ + [ + -17.251708984375, + false + ] + ], + [ + [ + -27.790363311767578, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.687619209289551, + false + ], + [ + -20.196495056152344, + false + ], + [ + -17.251708984375, + false + ], + [ + -27.790363311767578, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ce713f869e9e226e4e6243d90e96d1d8e130a8a761f36585e5f3a48ea1295069", + "prompt_hash": "4e377807034ef44a7a8d59253a6432b3ac6d9464956a15954af4d154870aedff", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 34, + "doc": { + "question": "Which training stage of Mini-Omni involves unfreezing all model weights for holistic fine-tuning?", + "choices": [ + "Stage 1: Modality Alignment", + "Stage 2: Adaptation Training", + "Stage 3: Multi-modal Finetuning", + "Stage 4: Evaluation" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which training stage of Mini-Omni involves unfreezing all model weights for holistic fine-tuning?\nAnswer:", + " Stage 1: Modality Alignment" + ], + [ + "Question:Which training stage of Mini-Omni involves unfreezing all model weights for holistic fine-tuning?\nAnswer:", + " Stage 2: Adaptation Training" + ], + [ + "Question:Which training stage of Mini-Omni involves unfreezing all model weights for holistic fine-tuning?\nAnswer:", + " Stage 3: Multi-modal Finetuning" + ], + [ + "Question:Which training stage of Mini-Omni involves unfreezing all model weights for holistic fine-tuning?\nAnswer:", + " Stage 4: Evaluation" + ] + ], + "resps": [ + [ + [ + -21.08725357055664, + false + ] + ], + [ + [ + -18.947019577026367, + false + ] + ], + [ + [ + -22.240943908691406, + false + ] + ], + [ + [ + -14.598235130310059, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.08725357055664, + false + ], + [ + -18.947019577026367, + false + ], + [ + -22.240943908691406, + false + ], + [ + -14.598235130310059, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d07e134beb38224450f014439b85131c773e0959af7b93b2e4c921492cf859bd", + "prompt_hash": "7031d671c2508328142687f9266f37ca2135e8ac581b5e9b2f79f1cae7ac811d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 35, + "doc": { + "question": "What does a single instance of ReKep map an array of keypoints to?", + "choices": [ + "A probability distribution", + "A bounded energy value", + "An unbounded cost", + "A sequence of discrete poses" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does a single instance of ReKep map an array of keypoints to?\nAnswer:", + " A probability distribution" + ], + [ + "Question:What does a single instance of ReKep map an array of keypoints to?\nAnswer:", + " A bounded energy value" + ], + [ + "Question:What does a single instance of ReKep map an array of keypoints to?\nAnswer:", + " An unbounded cost" + ], + [ + "Question:What does a single instance of ReKep map an array of keypoints to?\nAnswer:", + " A sequence of discrete poses" + ] + ], + "resps": [ + [ + [ + -11.133968353271484, + false + ] + ], + [ + [ + -26.93883514404297, + false + ] + ], + [ + [ + -22.620363235473633, + false + ] + ], + [ + [ + -19.65053939819336, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.133968353271484, + false + ], + [ + -26.93883514404297, + false + ], + [ + -22.620363235473633, + false + ], + [ + -19.65053939819336, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4db36d9910572980f07baf6fe692acf7552b15a3f17912dc3c9dd5ce464447ec", + "prompt_hash": "4cede1e94eded219e5a81eb153a41f1a84cc1f2cdf064b84363143b3049e11c4", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 36, + "doc": { + "question": "Which model is used in the system to propose fine-grained and semantically meaningful keypoints in the scene?", + "choices": [ + "BLIP-2", + "SAM", + "DINOv2", + "AnyGrasp" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model is used in the system to propose fine-grained and semantically meaningful keypoints in the scene?\nAnswer:", + " BLIP-2" + ], + [ + "Question:Which model is used in the system to propose fine-grained and semantically meaningful keypoints in the scene?\nAnswer:", + " SAM" + ], + [ + "Question:Which model is used in the system to propose fine-grained and semantically meaningful keypoints in the scene?\nAnswer:", + " DINOv2" + ], + [ + "Question:Which model is used in the system to propose fine-grained and semantically meaningful keypoints in the scene?\nAnswer:", + " AnyGrasp" + ] + ], + "resps": [ + [ + [ + -9.148873329162598, + false + ] + ], + [ + [ + -8.042795181274414, + false + ] + ], + [ + [ + -8.681550025939941, + false + ] + ], + [ + [ + -19.21855926513672, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.148873329162598, + false + ], + [ + -8.042795181274414, + false + ], + [ + -8.681550025939941, + false + ], + [ + -19.21855926513672, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "74a656301051bfff3335d6544f70a32cc1ade176e1c8177b21c267b58dcc33b3", + "prompt_hash": "43718c78e0c8a34db04835b5ba255717edd73cb30afa6de56e38a7f3ff8bc1c1", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 37, + "doc": { + "question": "What is the optimization framework’s approximate solving frequency for the tasks considered in the work?", + "choices": [ + "1 Hz", + "5 Hz", + "10 Hz", + "20 Hz" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the optimization framework’s approximate solving frequency for the tasks considered in the work?\nAnswer:", + " 1 Hz" + ], + [ + "Question:What is the optimization framework’s approximate solving frequency for the tasks considered in the work?\nAnswer:", + " 5 Hz" + ], + [ + "Question:What is the optimization framework’s approximate solving frequency for the tasks considered in the work?\nAnswer:", + " 10 Hz" + ], + [ + "Question:What is the optimization framework’s approximate solving frequency for the tasks considered in the work?\nAnswer:", + " 20 Hz" + ] + ], + "resps": [ + [ + [ + -11.280828475952148, + false + ] + ], + [ + [ + -11.805965423583984, + false + ] + ], + [ + [ + -10.422658920288086, + false + ] + ], + [ + [ + -11.62740707397461, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.280828475952148, + false + ], + [ + -11.805965423583984, + false + ], + [ + -10.422658920288086, + false + ], + [ + -11.62740707397461, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "bf78edf81140ee8a9beda3a69bd1fe95a2866cca0ccf89f5d91a1553787204a7", + "prompt_hash": "4bece240a600ba03e92ec71087ccf5b03ae8d9393b28dca5645c8b707fccc7e6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 38, + "doc": { + "question": "What auxiliary cost is specifically included in the sub-goal problem when a stage is concerned with grasping?", + "choices": [ + "Path length minimization", + "A grasp metric", + "Self-collision avoidance", + "Consistency cost" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What auxiliary cost is specifically included in the sub-goal problem when a stage is concerned with grasping?\nAnswer:", + " Path length minimization" + ], + [ + "Question:What auxiliary cost is specifically included in the sub-goal problem when a stage is concerned with grasping?\nAnswer:", + " A grasp metric" + ], + [ + "Question:What auxiliary cost is specifically included in the sub-goal problem when a stage is concerned with grasping?\nAnswer:", + " Self-collision avoidance" + ], + [ + "Question:What auxiliary cost is specifically included in the sub-goal problem when a stage is concerned with grasping?\nAnswer:", + " Consistency cost" + ] + ], + "resps": [ + [ + [ + -18.370542526245117, + false + ] + ], + [ + [ + -19.641986846923828, + false + ] + ], + [ + [ + -14.385573387145996, + false + ] + ], + [ + [ + -15.92432975769043, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.370542526245117, + false + ], + [ + -19.641986846923828, + false + ], + [ + -14.385573387145996, + false + ], + [ + -15.92432975769043, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c06195ea7e94ff57f5dd396d3fffa0f6546f0629988a77b82a5234c7579a21bc", + "prompt_hash": "330d51adcd40a4f69715390ef1b0223c4b230bab7b8bb500d1fe3ba2f348f4e6", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 39, + "doc": { + "question": "Which component contributed the largest portion of system errors according to the error breakdown analysis?", + "choices": [ + "Keypoint proposal", + "Vision-language model", + "Optimization module", + "Point tracker" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which component contributed the largest portion of system errors according to the error breakdown analysis?\nAnswer:", + " Keypoint proposal" + ], + [ + "Question:Which component contributed the largest portion of system errors according to the error breakdown analysis?\nAnswer:", + " Vision-language model" + ], + [ + "Question:Which component contributed the largest portion of system errors according to the error breakdown analysis?\nAnswer:", + " Optimization module" + ], + [ + "Question:Which component contributed the largest portion of system errors according to the error breakdown analysis?\nAnswer:", + " Point tracker" + ] + ], + "resps": [ + [ + [ + -23.096324920654297, + false + ] + ], + [ + [ + -21.57367706298828, + false + ] + ], + [ + [ + -17.320581436157227, + false + ] + ], + [ + [ + -19.568580627441406, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.096324920654297, + false + ], + [ + -21.57367706298828, + false + ], + [ + -17.320581436157227, + false + ], + [ + -19.568580627441406, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "04a4dfcb73d83e71a147cfcba69814390ffe6d2a77c6e0cdada07d010423dba8", + "prompt_hash": "6792587cdddb5ff84a1438cd151eaf7860da1b4fb2100f0e33b5cca813c0c760", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 40, + "doc": { + "question": "What is the primary goal of monocular depth estimation?", + "choices": [ + "To predict the depth map from a single image", + "To generate high-resolution RGB images", + "To classify objects in 3D scenes", + "To segment moving objects in videos" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the primary goal of monocular depth estimation?\nAnswer:", + " To predict the depth map from a single image" + ], + [ + "Question:What is the primary goal of monocular depth estimation?\nAnswer:", + " To generate high-resolution RGB images" + ], + [ + "Question:What is the primary goal of monocular depth estimation?\nAnswer:", + " To classify objects in 3D scenes" + ], + [ + "Question:What is the primary goal of monocular depth estimation?\nAnswer:", + " To segment moving objects in videos" + ] + ], + "resps": [ + [ + [ + -11.69544792175293, + false + ] + ], + [ + [ + -20.34946060180664, + false + ] + ], + [ + [ + -17.6307315826416, + false + ] + ], + [ + [ + -21.26361846923828, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.69544792175293, + false + ], + [ + -20.34946060180664, + false + ], + [ + -17.6307315826416, + false + ], + [ + -21.26361846923828, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fd6ac965fdc8b359488adc36e6b635fc9198c2aba8774119dc641286ade6e6e2", + "prompt_hash": "69901e85b64035444f79c66e1412b12263a82428b19daa57a4134cff44e3f0e7", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 41, + "doc": { + "question": "Which dataset was combined with DynamicReplica to form the synthetic dataset for DepthCrafter training?", + "choices": [ + "MatrixCity", + "ScanNet v2", + "KITTI", + "Sintel" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which dataset was combined with DynamicReplica to form the synthetic dataset for DepthCrafter training?\nAnswer:", + " MatrixCity" + ], + [ + "Question:Which dataset was combined with DynamicReplica to form the synthetic dataset for DepthCrafter training?\nAnswer:", + " ScanNet v2" + ], + [ + "Question:Which dataset was combined with DynamicReplica to form the synthetic dataset for DepthCrafter training?\nAnswer:", + " KITTI" + ], + [ + "Question:Which dataset was combined with DynamicReplica to form the synthetic dataset for DepthCrafter training?\nAnswer:", + " Sintel" + ] + ], + "resps": [ + [ + [ + -18.05780792236328, + false + ] + ], + [ + [ + -16.11438751220703, + false + ] + ], + [ + [ + -7.35317850112915, + false + ] + ], + [ + [ + -15.332847595214844, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.05780792236328, + false + ], + [ + -16.11438751220703, + false + ], + [ + -7.35317850112915, + false + ], + [ + -15.332847595214844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3e85574950d0bcb2dc31b42eca31b8559b154e0263855ed93eb652cdd64d3518", + "prompt_hash": "33448e03dee4e4b13849c0651c6ad9786e0c84071ce59ca2db90133b397de274", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 42, + "doc": { + "question": "How many training stages are involved in DepthCrafter's training strategy?", + "choices": [ + "One", + "Two", + "Three", + "Four" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many training stages are involved in DepthCrafter's training strategy?\nAnswer:", + " One" + ], + [ + "Question:How many training stages are involved in DepthCrafter's training strategy?\nAnswer:", + " Two" + ], + [ + "Question:How many training stages are involved in DepthCrafter's training strategy?\nAnswer:", + " Three" + ], + [ + "Question:How many training stages are involved in DepthCrafter's training strategy?\nAnswer:", + " Four" + ] + ], + "resps": [ + [ + [ + -5.532721042633057, + false + ] + ], + [ + [ + -5.27767276763916, + false + ] + ], + [ + [ + -5.411354064941406, + false + ] + ], + [ + [ + -6.417080402374268, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.532721042633057, + false + ], + [ + -5.27767276763916, + false + ], + [ + -5.411354064941406, + false + ], + [ + -6.417080402374268, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b44adef1648b27a5a5dd13a349c17593441f704701a56c8ee62bd6626e2f6944", + "prompt_hash": "55709d651755d15c0f99595b4611603adc6430a5f68194a99ad0638bf094d428", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 43, + "doc": { + "question": "What is the maximum number of frames DepthCrafter can generate in a single sequence after training?", + "choices": [ + "25", + "50", + "90", + "110" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the maximum number of frames DepthCrafter can generate in a single sequence after training?\nAnswer:", + " 25" + ], + [ + "Question:What is the maximum number of frames DepthCrafter can generate in a single sequence after training?\nAnswer:", + " 50" + ], + [ + "Question:What is the maximum number of frames DepthCrafter can generate in a single sequence after training?\nAnswer:", + " 90" + ], + [ + "Question:What is the maximum number of frames DepthCrafter can generate in a single sequence after training?\nAnswer:", + " 110" + ] + ], + "resps": [ + [ + [ + -4.397420406341553, + false + ] + ], + [ + [ + -4.530661106109619, + false + ] + ], + [ + [ + -5.898233413696289, + false + ] + ], + [ + [ + -7.255423069000244, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.397420406341553, + false + ], + [ + -4.530661106109619, + false + ], + [ + -5.898233413696289, + false + ], + [ + -7.255423069000244, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "32ef88ba8b3db37223d2380545cfb74b1dbe2219fe962aff8a8ec1ca9136f19e", + "prompt_hash": "a5b25ded009375961e1d26e28deed8d0d580dbde9bfe97316270ca89d3fdc6f3", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 44, + "doc": { + "question": "Which baseline method does ChronoDepth support in terms of temporal context length?", + "choices": [ + "10 frames", + "25 frames", + "50 frames", + "110 frames" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which baseline method does ChronoDepth support in terms of temporal context length?\nAnswer:", + " 10 frames" + ], + [ + "Question:Which baseline method does ChronoDepth support in terms of temporal context length?\nAnswer:", + " 25 frames" + ], + [ + "Question:Which baseline method does ChronoDepth support in terms of temporal context length?\nAnswer:", + " 50 frames" + ], + [ + "Question:Which baseline method does ChronoDepth support in terms of temporal context length?\nAnswer:", + " 110 frames" + ] + ], + "resps": [ + [ + [ + -9.895761489868164, + false + ] + ], + [ + [ + -10.400951385498047, + false + ] + ], + [ + [ + -11.934900283813477, + false + ] + ], + [ + [ + -14.357010841369629, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.895761489868164, + false + ], + [ + -10.400951385498047, + false + ], + [ + -11.934900283813477, + false + ], + [ + -14.357010841369629, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "51c5df2258ee6bbda62c0c66f6152a967f8434e2f9764045f20dae32a9327957", + "prompt_hash": "2060a60d6854a6f73fc75059f57f1ed3d482de4aae258a212b1bcca7ea590006", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 45, + "doc": { + "question": "Which base model is Qwen2.5-Coder derived from?", + "choices": [ + "GPT-4o", + "Qwen2.5", + "StarCoder2", + "CodeLlama" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which base model is Qwen2.5-Coder derived from?\nAnswer:", + " GPT-4o" + ], + [ + "Question:Which base model is Qwen2.5-Coder derived from?\nAnswer:", + " Qwen2.5" + ], + [ + "Question:Which base model is Qwen2.5-Coder derived from?\nAnswer:", + " StarCoder2" + ], + [ + "Question:Which base model is Qwen2.5-Coder derived from?\nAnswer:", + " CodeLlama" + ] + ], + "resps": [ + [ + [ + -9.98691177368164, + false + ] + ], + [ + [ + -2.9870645999908447, + false + ] + ], + [ + [ + -14.584648132324219, + false + ] + ], + [ + [ + -9.186256408691406, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.98691177368164, + false + ], + [ + -2.9870645999908447, + false + ], + [ + -14.584648132324219, + false + ], + [ + -9.186256408691406, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9184d47659d765f50610d9b95090031ec3c8dc62d330f69af8ec5deaea34f03c", + "prompt_hash": "ebf9dac2363b8a5ca7ba4947530d007bf8719cb7420805d3116a28df8bc57253", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 46, + "doc": { + "question": "How many different model sizes are included in the Qwen2.5-Coder series?", + "choices": [ + "Four", + "Six", + "Eight", + "Ten" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many different model sizes are included in the Qwen2.5-Coder series?\nAnswer:", + " Four" + ], + [ + "Question:How many different model sizes are included in the Qwen2.5-Coder series?\nAnswer:", + " Six" + ], + [ + "Question:How many different model sizes are included in the Qwen2.5-Coder series?\nAnswer:", + " Eight" + ], + [ + "Question:How many different model sizes are included in the Qwen2.5-Coder series?\nAnswer:", + " Ten" + ] + ], + "resps": [ + [ + [ + -5.960505962371826, + false + ] + ], + [ + [ + -6.804236888885498, + false + ] + ], + [ + [ + -8.083930015563965, + false + ] + ], + [ + [ + -9.204055786132812, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.960505962371826, + false + ], + [ + -6.804236888885498, + false + ], + [ + -8.083930015563965, + false + ], + [ + -9.204055786132812, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "61511e1ed9fbd67b26fc1d3a2a743382894c8370ced0cc61d26a2bab76d28318", + "prompt_hash": "f0c03dd966a50b77ea4090116c3fbb3b5e4851c743c0eca6bbff3bace1c5cca7", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 47, + "doc": { + "question": "What is the total number of tokens in the Qwen2.5-Coder pretraining dataset?", + "choices": [ + "3.2 trillion", + "4.1 trillion", + "5.5 trillion", + "6.7 trillion" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the total number of tokens in the Qwen2.5-Coder pretraining dataset?\nAnswer:", + " 3.2 trillion" + ], + [ + "Question:What is the total number of tokens in the Qwen2.5-Coder pretraining dataset?\nAnswer:", + " 4.1 trillion" + ], + [ + "Question:What is the total number of tokens in the Qwen2.5-Coder pretraining dataset?\nAnswer:", + " 5.5 trillion" + ], + [ + "Question:What is the total number of tokens in the Qwen2.5-Coder pretraining dataset?\nAnswer:", + " 6.7 trillion" + ] + ], + "resps": [ + [ + [ + -13.193757057189941, + false + ] + ], + [ + [ + -14.902152061462402, + false + ] + ], + [ + [ + -14.214559555053711, + false + ] + ], + [ + [ + -14.328497886657715, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.193757057189941, + false + ], + [ + -14.902152061462402, + false + ], + [ + -14.214559555053711, + false + ], + [ + -14.328497886657715, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "519136bb6844023e6bc3559c19b3c52f48c6f04ed430e4109e45413fee24a586", + "prompt_hash": "5e46854b005aabfaefe305fc51bd6413e5335ef12525ff420d44a53890dc5eb8", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 48, + "doc": { + "question": "Which token is used in Qwen2.5-Coder to mark the end of a text or sequence?", + "choices": [ + "<|fim_suffix|>", + "<|file_sep|>", + "<|endoftext|>", + "<|fim_middle|>" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which token is used in Qwen2.5-Coder to mark the end of a text or sequence?\nAnswer:", + " <|fim_suffix|>" + ], + [ + "Question:Which token is used in Qwen2.5-Coder to mark the end of a text or sequence?\nAnswer:", + " <|file_sep|>" + ], + [ + "Question:Which token is used in Qwen2.5-Coder to mark the end of a text or sequence?\nAnswer:", + " <|endoftext|>" + ], + [ + "Question:Which token is used in Qwen2.5-Coder to mark the end of a text or sequence?\nAnswer:", + " <|fim_middle|>" + ] + ], + "resps": [ + [ + [ + -40.7396125793457, + false + ] + ], + [ + [ + -31.887859344482422, + false + ] + ], + [ + [ + -15.501543045043945, + false + ] + ], + [ + [ + -38.91889572143555, + false + ] + ] + ], + "filtered_resps": [ + [ + -40.7396125793457, + false + ], + [ + -31.887859344482422, + false + ], + [ + -15.501543045043945, + false + ], + [ + -38.91889572143555, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "32688a8ff3a6312b9cf1f4d13435c86284be27e68f112c1ea7c044b55f9876f5", + "prompt_hash": "4a280c533e45a6c1f706ed603bc9f59e96708374dfc6fc0bf823638c32ffe8ad", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 49, + "doc": { + "question": "What was the final mixture ratio of Code, Text, and Math data chosen for Qwen2.5-Coder?", + "choices": [ + "100:0:0", + "85:15:5", + "70:20:10", + "60:30:10" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What was the final mixture ratio of Code, Text, and Math data chosen for Qwen2.5-Coder?\nAnswer:", + " 100:0:0" + ], + [ + "Question:What was the final mixture ratio of Code, Text, and Math data chosen for Qwen2.5-Coder?\nAnswer:", + " 85:15:5" + ], + [ + "Question:What was the final mixture ratio of Code, Text, and Math data chosen for Qwen2.5-Coder?\nAnswer:", + " 70:20:10" + ], + [ + "Question:What was the final mixture ratio of Code, Text, and Math data chosen for Qwen2.5-Coder?\nAnswer:", + " 60:30:10" + ] + ], + "resps": [ + [ + [ + -12.652618408203125, + false + ] + ], + [ + [ + -11.911613464355469, + false + ] + ], + [ + [ + -10.841472625732422, + false + ] + ], + [ + [ + -11.832803726196289, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.652618408203125, + false + ], + [ + -11.911613464355469, + false + ], + [ + -10.841472625732422, + false + ], + [ + -11.832803726196289, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8686768174e3565d0100e217dbda69483a98ee1b6994f8f3201ee0ca5e0a49cb", + "prompt_hash": "aac1d2cd9a36d901ef435efee1953e49f52c134a3315544ae637814f935077c2", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 50, + "doc": { + "question": "Which classical equation is widely used to predict the oscillation behavior of spherical cavitation bubbles?", + "choices": [ + "Gilmore equation", + "Keller-Miksis equation", + "Rayleigh-Plesset equation", + "Doubly Asymptotic Approximation equation" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which classical equation is widely used to predict the oscillation behavior of spherical cavitation bubbles?\nAnswer:", + " Gilmore equation" + ], + [ + "Question:Which classical equation is widely used to predict the oscillation behavior of spherical cavitation bubbles?\nAnswer:", + " Keller-Miksis equation" + ], + [ + "Question:Which classical equation is widely used to predict the oscillation behavior of spherical cavitation bubbles?\nAnswer:", + " Rayleigh-Plesset equation" + ], + [ + "Question:Which classical equation is widely used to predict the oscillation behavior of spherical cavitation bubbles?\nAnswer:", + " Doubly Asymptotic Approximation equation" + ] + ], + "resps": [ + [ + [ + -13.879878044128418, + false + ] + ], + [ + [ + -31.828353881835938, + false + ] + ], + [ + [ + -14.75823974609375, + false + ] + ], + [ + [ + -32.73979187011719, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.879878044128418, + false + ], + [ + -31.828353881835938, + false + ], + [ + -14.75823974609375, + false + ], + [ + -32.73979187011719, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "86560bbde1abd696430083a87c3d3ed80ef43f5eb498566253bb3713a6b16b2e", + "prompt_hash": "9888374170c0feb7fc1506a6e95683c4d811614684fd0bb3fc707931241f05ee", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 51, + "doc": { + "question": "What additional factors, besides compressibility and viscosity, have been shown to significantly affect the dynamics of laser-induced and spark bubbles?", + "choices": [ + "Electrical conductivity and surface roughness", + "Condensation and evaporation processes", + "Magnetic field fluctuations", + "Bubble shape irregularities" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What additional factors, besides compressibility and viscosity, have been shown to significantly affect the dynamics of laser-induced and spark bubbles?\nAnswer:", + " Electrical conductivity and surface roughness" + ], + [ + "Question:What additional factors, besides compressibility and viscosity, have been shown to significantly affect the dynamics of laser-induced and spark bubbles?\nAnswer:", + " Condensation and evaporation processes" + ], + [ + "Question:What additional factors, besides compressibility and viscosity, have been shown to significantly affect the dynamics of laser-induced and spark bubbles?\nAnswer:", + " Magnetic field fluctuations" + ], + [ + "Question:What additional factors, besides compressibility and viscosity, have been shown to significantly affect the dynamics of laser-induced and spark bubbles?\nAnswer:", + " Bubble shape irregularities" + ] + ], + "resps": [ + [ + [ + -19.15823745727539, + false + ] + ], + [ + [ + -22.268741607666016, + false + ] + ], + [ + [ + -17.82883644104004, + false + ] + ], + [ + [ + -25.504554748535156, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.15823745727539, + false + ], + [ + -22.268741607666016, + false + ], + [ + -17.82883644104004, + false + ], + [ + -25.504554748535156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "579da9a1d1f14a4f7b9696f859194c3f6a831c5cba3710a01dc5548241c45e6e", + "prompt_hash": "b7184e53f62c00f868964d623ed2924be0ff3e37482871c7dbb7600dc757c8af", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 52, + "doc": { + "question": "What equation of state is employed to model the uniform inner pressure of a bubble under violent compression?", + "choices": [ + "Ideal gas equation", + "Boyle’s law", + "Van der Waals equation", + "Clapeyron equation" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What equation of state is employed to model the uniform inner pressure of a bubble under violent compression?\nAnswer:", + " Ideal gas equation" + ], + [ + "Question:What equation of state is employed to model the uniform inner pressure of a bubble under violent compression?\nAnswer:", + " Boyle’s law" + ], + [ + "Question:What equation of state is employed to model the uniform inner pressure of a bubble under violent compression?\nAnswer:", + " Van der Waals equation" + ], + [ + "Question:What equation of state is employed to model the uniform inner pressure of a bubble under violent compression?\nAnswer:", + " Clapeyron equation" + ] + ], + "resps": [ + [ + [ + -12.687660217285156, + false + ] + ], + [ + [ + -15.20535659790039, + false + ] + ], + [ + [ + -10.195505142211914, + false + ] + ], + [ + [ + -13.616377830505371, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.687660217285156, + false + ], + [ + -15.20535659790039, + false + ], + [ + -10.195505142211914, + false + ], + [ + -13.616377830505371, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3cb46df70cd1fd88e71c4fb2d06e77ca266e27965aac6d41767d56bcf1122835", + "prompt_hash": "f2b917857232c039bc388235d8cdb0eaadf3dbdbe16624f6c6e91f2d9d335317", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 53, + "doc": { + "question": "What is the typical value of the surface tension coefficient at room temperature (293 K) provided in the model parameters?", + "choices": [ + "0.001 Pa·s", + "0.075 N/m", + "0.1402 J·m³·mol⁻²", + "0.55 W·m⁻¹·K⁻¹" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the typical value of the surface tension coefficient at room temperature (293 K) provided in the model parameters?\nAnswer:", + " 0.001 Pa·s" + ], + [ + "Question:What is the typical value of the surface tension coefficient at room temperature (293 K) provided in the model parameters?\nAnswer:", + " 0.075 N/m" + ], + [ + "Question:What is the typical value of the surface tension coefficient at room temperature (293 K) provided in the model parameters?\nAnswer:", + " 0.1402 J·m³·mol⁻²" + ], + [ + "Question:What is the typical value of the surface tension coefficient at room temperature (293 K) provided in the model parameters?\nAnswer:", + " 0.55 W·m⁻¹·K⁻¹" + ] + ], + "resps": [ + [ + [ + -14.709671974182129, + false + ] + ], + [ + [ + -12.447090148925781, + false + ] + ], + [ + [ + -48.58300018310547, + false + ] + ], + [ + [ + -41.24347686767578, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.709671974182129, + false + ], + [ + -12.447090148925781, + false + ], + [ + -48.58300018310547, + false + ], + [ + -41.24347686767578, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "60c2ea3472dcbc20b38df041644ae8bce015f14b1fd11b1ca65a21e057fbe2f5", + "prompt_hash": "2ca5dd46a48399f1a5aa66a281c6a0986d24dfdd7319e23ebdd5d0df8cbe3d80", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 54, + "doc": { + "question": "In the spark-generated bubble experiment, what was the maximum bubble radius reached?", + "choices": [ + "1.01 mm", + "12 mm", + "18.1 mm", + "30 mm" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In the spark-generated bubble experiment, what was the maximum bubble radius reached?\nAnswer:", + " 1.01 mm" + ], + [ + "Question:In the spark-generated bubble experiment, what was the maximum bubble radius reached?\nAnswer:", + " 12 mm" + ], + [ + "Question:In the spark-generated bubble experiment, what was the maximum bubble radius reached?\nAnswer:", + " 18.1 mm" + ], + [ + "Question:In the spark-generated bubble experiment, what was the maximum bubble radius reached?\nAnswer:", + " 30 mm" + ] + ], + "resps": [ + [ + [ + -12.215747833251953, + false + ] + ], + [ + [ + -9.254501342773438, + false + ] + ], + [ + [ + -13.88467025756836, + false + ] + ], + [ + [ + -9.194122314453125, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.215747833251953, + false + ], + [ + -9.254501342773438, + false + ], + [ + -13.88467025756836, + false + ], + [ + -9.194122314453125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0e8a2ce0886b9e422a63b1299efe89deaacb5858a6ea609eaff8ac381b44b4ab", + "prompt_hash": "315a87f41522f6690ece93bd824296a44a8f5d3acece53a5ecfb557a47647875", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 55, + "doc": { + "question": "Which vision-language model serves as the backbone for the \\u201cModelSymbol\\u201d architecture?", + "choices": [ + "PaliGemma", + "Flamingo", + "Gemini", + "Octo" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which vision-language model serves as the backbone for the \\u201cModelSymbol\\u201d architecture?\nAnswer:", + " PaliGemma" + ], + [ + "Question:Which vision-language model serves as the backbone for the \\u201cModelSymbol\\u201d architecture?\nAnswer:", + " Flamingo" + ], + [ + "Question:Which vision-language model serves as the backbone for the \\u201cModelSymbol\\u201d architecture?\nAnswer:", + " Gemini" + ], + [ + "Question:Which vision-language model serves as the backbone for the \\u201cModelSymbol\\u201d architecture?\nAnswer:", + " Octo" + ] + ], + "resps": [ + [ + [ + -23.2396240234375, + false + ] + ], + [ + [ + -8.813681602478027, + false + ] + ], + [ + [ + -6.939650535583496, + false + ] + ], + [ + [ + -16.771682739257812, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.2396240234375, + false + ], + [ + -8.813681602478027, + false + ], + [ + -6.939650535583496, + false + ], + [ + -16.771682739257812, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "44cfab75afc74c3f9040b647b44c8a6cd66f888103de6a4e7fc72665cca5ad9f", + "prompt_hash": "e760c9d7b72d0fa964da31194d5bb2ddc1361f8ebc0a9893eb3d7be1f75cf88c", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 56, + "doc": { + "question": "What technique does the \\u201cModelSymbol\\u201d model use to generate continuous action distributions?", + "choices": [ + "Autoregressive discretization", + "Conditional flow matching", + "Reinforcement learning", + "Cross-entropy loss" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What technique does the \\u201cModelSymbol\\u201d model use to generate continuous action distributions?\nAnswer:", + " Autoregressive discretization" + ], + [ + "Question:What technique does the \\u201cModelSymbol\\u201d model use to generate continuous action distributions?\nAnswer:", + " Conditional flow matching" + ], + [ + "Question:What technique does the \\u201cModelSymbol\\u201d model use to generate continuous action distributions?\nAnswer:", + " Reinforcement learning" + ], + [ + "Question:What technique does the \\u201cModelSymbol\\u201d model use to generate continuous action distributions?\nAnswer:", + " Cross-entropy loss" + ] + ], + "resps": [ + [ + [ + -19.554214477539062, + false + ] + ], + [ + [ + -19.457931518554688, + false + ] + ], + [ + [ + -8.617112159729004, + false + ] + ], + [ + [ + -12.441343307495117, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.554214477539062, + false + ], + [ + -19.457931518554688, + false + ], + [ + -8.617112159729004, + false + ], + [ + -12.441343307495117, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "34120c787c63f3cd993bcfdbf7c316dd59a741f2c53a6b9663c10007fc52909a", + "prompt_hash": "47bba7a5ba4c27f3890616b2747252222a0805931bab6c0bb030384c75608899", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 57, + "doc": { + "question": "How many parameters does the \\u201cModelSymbol\\u201d model have, including the action expert?", + "choices": [ + "470M", + "3B", + "3.3B", + "7B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many parameters does the \\u201cModelSymbol\\u201d model have, including the action expert?\nAnswer:", + " 470M" + ], + [ + "Question:How many parameters does the \\u201cModelSymbol\\u201d model have, including the action expert?\nAnswer:", + " 3B" + ], + [ + "Question:How many parameters does the \\u201cModelSymbol\\u201d model have, including the action expert?\nAnswer:", + " 3.3B" + ], + [ + "Question:How many parameters does the \\u201cModelSymbol\\u201d model have, including the action expert?\nAnswer:", + " 7B" + ] + ], + "resps": [ + [ + [ + -14.699792861938477, + false + ] + ], + [ + [ + -12.135706901550293, + false + ] + ], + [ + [ + -14.984294891357422, + false + ] + ], + [ + [ + -11.93716049194336, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.699792861938477, + false + ], + [ + -12.135706901550293, + false + ], + [ + -14.984294891357422, + false + ], + [ + -11.93716049194336, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "290acbf83831bbc31d777b5bc2f83c49ae3dec908c2a92ceac3e542f64403a1f", + "prompt_hash": "e9bb7024e9173b093e62fd7785fba67863ebe1682748bc459de2243494c6f04d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 58, + "doc": { + "question": "What is the purpose of weighting each task-robot combination by $n^{0.43}$ in the pre-training mixture?", + "choices": [ + "To give more weight to overrepresented tasks", + "To down-weight overrepresented combinations", + "To exclude rare tasks", + "To normalize action dimensions" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the purpose of weighting each task-robot combination by $n^{0.43}$ in the pre-training mixture?\nAnswer:", + " To give more weight to overrepresented tasks" + ], + [ + "Question:What is the purpose of weighting each task-robot combination by $n^{0.43}$ in the pre-training mixture?\nAnswer:", + " To down-weight overrepresented combinations" + ], + [ + "Question:What is the purpose of weighting each task-robot combination by $n^{0.43}$ in the pre-training mixture?\nAnswer:", + " To exclude rare tasks" + ], + [ + "Question:What is the purpose of weighting each task-robot combination by $n^{0.43}$ in the pre-training mixture?\nAnswer:", + " To normalize action dimensions" + ] + ], + "resps": [ + [ + [ + -24.754289627075195, + false + ] + ], + [ + [ + -28.06523895263672, + false + ] + ], + [ + [ + -18.77325439453125, + false + ] + ], + [ + [ + -24.841075897216797, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.754289627075195, + false + ], + [ + -28.06523895263672, + false + ], + [ + -18.77325439453125, + false + ], + [ + -24.841075897216797, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ae68a34690e4101ae12959abbabd3d053cfe34bc5cdc918da5cc0668c788706b", + "prompt_hash": "cbee554b6bf15797b1f6b7df3a5f97907a8c24a124c42fc767a8197d561e4d63", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 59, + "doc": { + "question": "What is the control frequency that the action chunking architecture enables for dexterous tasks such as laundry folding?", + "choices": [ + "2 Hz", + "10 Hz", + "25 Hz", + "50 Hz" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the control frequency that the action chunking architecture enables for dexterous tasks such as laundry folding?\nAnswer:", + " 2 Hz" + ], + [ + "Question:What is the control frequency that the action chunking architecture enables for dexterous tasks such as laundry folding?\nAnswer:", + " 10 Hz" + ], + [ + "Question:What is the control frequency that the action chunking architecture enables for dexterous tasks such as laundry folding?\nAnswer:", + " 25 Hz" + ], + [ + "Question:What is the control frequency that the action chunking architecture enables for dexterous tasks such as laundry folding?\nAnswer:", + " 50 Hz" + ] + ], + "resps": [ + [ + [ + -6.943455696105957, + false + ] + ], + [ + [ + -5.619145393371582, + false + ] + ], + [ + [ + -7.140289306640625, + false + ] + ], + [ + [ + -7.105968952178955, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.943455696105957, + false + ], + [ + -5.619145393371582, + false + ], + [ + -7.140289306640625, + false + ], + [ + -7.105968952178955, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0d1c836e5acfd601aca57b114c2e4b75c32aeb471d59df90b0932f0ad5ce0efd", + "prompt_hash": "699d91475e58382dad8ac2e319ff793e97e57a1799094f57e59be2c803407bde", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 60, + "doc": { + "question": "Which optimizer was used for the weight fitting procedure?", + "choices": [ + "SGD", + "Adam", + "RMSProp", + "Adagrad" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which optimizer was used for the weight fitting procedure?\nAnswer:", + " SGD" + ], + [ + "Question:Which optimizer was used for the weight fitting procedure?\nAnswer:", + " Adam" + ], + [ + "Question:Which optimizer was used for the weight fitting procedure?\nAnswer:", + " RMSProp" + ], + [ + "Question:Which optimizer was used for the weight fitting procedure?\nAnswer:", + " Adagrad" + ] + ], + "resps": [ + [ + [ + -5.789969444274902, + false + ] + ], + [ + [ + -2.538813591003418, + false + ] + ], + [ + [ + -6.673875331878662, + false + ] + ], + [ + [ + -5.803485870361328, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.789969444274902, + false + ], + [ + -2.538813591003418, + false + ], + [ + -6.673875331878662, + false + ], + [ + -5.803485870361328, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "090dde0ae1db7d78793c9d6912aa6afb8c8f6488d175d3c5ea5850dc65750634", + "prompt_hash": "7aee1cbf7079e009ad787cb4272dd65a51025f5d2c300668e30a7dd42218369a", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 61, + "doc": { + "question": "How many completions were manually labeled in the Gold set?", + "choices": [ + "268", + "132", + "518", + "118" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many completions were manually labeled in the Gold set?\nAnswer:", + " 268" + ], + [ + "Question:How many completions were manually labeled in the Gold set?\nAnswer:", + " 132" + ], + [ + "Question:How many completions were manually labeled in the Gold set?\nAnswer:", + " 518" + ], + [ + "Question:How many completions were manually labeled in the Gold set?\nAnswer:", + " 118" + ] + ], + "resps": [ + [ + [ + -7.861661911010742, + false + ] + ], + [ + [ + -6.840625762939453, + false + ] + ], + [ + [ + -8.422104835510254, + false + ] + ], + [ + [ + -6.897829055786133, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.861661911010742, + false + ], + [ + -6.840625762939453, + false + ], + [ + -8.422104835510254, + false + ], + [ + -6.897829055786133, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9de3ad02da421e52280aa159084cbd6ea988739691cdd2ffc4d2109cc4b80c81", + "prompt_hash": "9d41457bfedf823bd9b92b57415b78e62d7f5a14128f49243121a9c5d230afed", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 62, + "doc": { + "question": "What was the learning rate used during weight fitting?", + "choices": [ + "0.001", + "0.01", + "0.05", + "0.1" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What was the learning rate used during weight fitting?\nAnswer:", + " 0.001" + ], + [ + "Question:What was the learning rate used during weight fitting?\nAnswer:", + " 0.01" + ], + [ + "Question:What was the learning rate used during weight fitting?\nAnswer:", + " 0.05" + ], + [ + "Question:What was the learning rate used during weight fitting?\nAnswer:", + " 0.1" + ] + ], + "resps": [ + [ + [ + -4.894764423370361, + false + ] + ], + [ + [ + -4.790251731872559, + false + ] + ], + [ + [ + -6.303414344787598, + false + ] + ], + [ + [ + -5.2953596115112305, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.894764423370361, + false + ], + [ + -4.790251731872559, + false + ], + [ + -6.303414344787598, + false + ], + [ + -5.2953596115112305, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4324393b6c3286950e5766986366bc8e8568567b195d8a801d217c700e457dae", + "prompt_hash": "fad73082b060c33f0c83c2209196e99cc73ff6237e07ebb4a45d4124a70d6842", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 63, + "doc": { + "question": "Which content policy areas were considered in the experiments?", + "choices": [ + "Erotic Content, Hate Speech, Criminal Advice, Self-Harm", + "Political Speech, Hate Speech, Spam, Self-Harm", + "Fraud, Hate Speech, Misinformation, Self-Harm", + "Erotic Content, Terrorism, Criminal Advice, Bullying" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which content policy areas were considered in the experiments?\nAnswer:", + " Erotic Content, Hate Speech, Criminal Advice, Self-Harm" + ], + [ + "Question:Which content policy areas were considered in the experiments?\nAnswer:", + " Political Speech, Hate Speech, Spam, Self-Harm" + ], + [ + "Question:Which content policy areas were considered in the experiments?\nAnswer:", + " Fraud, Hate Speech, Misinformation, Self-Harm" + ], + [ + "Question:Which content policy areas were considered in the experiments?\nAnswer:", + " Erotic Content, Terrorism, Criminal Advice, Bullying" + ] + ], + "resps": [ + [ + [ + -43.40247344970703, + false + ] + ], + [ + [ + -37.416038513183594, + false + ] + ], + [ + [ + -29.05992889404297, + false + ] + ], + [ + [ + -49.14972686767578, + false + ] + ] + ], + "filtered_resps": [ + [ + -43.40247344970703, + false + ], + [ + -37.416038513183594, + false + ], + [ + -29.05992889404297, + false + ], + [ + -49.14972686767578, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ec1dc5d71815891ddedae21bbe7cc1a744059a5e451c5bcc2093a92d4e0a6118", + "prompt_hash": "c03957363758c5755716117db8f5e52b831f079592e8fbcb8ae08130ab55f79e", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 64, + "doc": { + "question": "What was the size of the synthetically generated RBR weight fitting comparison dataset ($\\\\DRBR$)?", + "choices": [ + "~6.7K", + "518", + "6.7K * 4", + "3K" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What was the size of the synthetically generated RBR weight fitting comparison dataset ($\\\\DRBR$)?\nAnswer:", + " ~6.7K" + ], + [ + "Question:What was the size of the synthetically generated RBR weight fitting comparison dataset ($\\\\DRBR$)?\nAnswer:", + " 518" + ], + [ + "Question:What was the size of the synthetically generated RBR weight fitting comparison dataset ($\\\\DRBR$)?\nAnswer:", + " 6.7K * 4" + ], + [ + "Question:What was the size of the synthetically generated RBR weight fitting comparison dataset ($\\\\DRBR$)?\nAnswer:", + " 3K" + ] + ], + "resps": [ + [ + [ + -19.00811004638672, + false + ] + ], + [ + [ + -9.766897201538086, + false + ] + ], + [ + [ + -26.137615203857422, + false + ] + ], + [ + [ + -9.225601196289062, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.00811004638672, + false + ], + [ + -9.766897201538086, + false + ], + [ + -26.137615203857422, + false + ], + [ + -9.225601196289062, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7ee279fed47dd700458e5d931b275ac58723b8d23378f36ec4753610edbda3ca", + "prompt_hash": "8698351b05ba5a7514fea881bafb225b56a1c0c8657ba5062889ef7098db1c3e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 65, + "doc": { + "question": "How many short, fact-seeking questions are included in the SimpleQA benchmark?", + "choices": [ + "3,500", + "4,326", + "5,200", + "6,000" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many short, fact-seeking questions are included in the SimpleQA benchmark?\nAnswer:", + " 3,500" + ], + [ + "Question:How many short, fact-seeking questions are included in the SimpleQA benchmark?\nAnswer:", + " 4,326" + ], + [ + "Question:How many short, fact-seeking questions are included in the SimpleQA benchmark?\nAnswer:", + " 5,200" + ], + [ + "Question:How many short, fact-seeking questions are included in the SimpleQA benchmark?\nAnswer:", + " 6,000" + ] + ], + "resps": [ + [ + [ + -11.849647521972656, + false + ] + ], + [ + [ + -15.20968246459961, + false + ] + ], + [ + [ + -13.093934059143066, + false + ] + ], + [ + [ + -9.874208450317383, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.849647521972656, + false + ], + [ + -15.20968246459961, + false + ], + [ + -13.093934059143066, + false + ], + [ + -9.874208450317383, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b99421e5a3ca71704d84295315f2f7d05abf1b053eea290e1385acdf3a8c58cd", + "prompt_hash": "73ea9f9187b6d5d8d6c3ca3d9892d2a486a91e14344cae62dc4d2ec031148a3b", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 66, + "doc": { + "question": "Which U.S. TV station did the Canadian reality series 'To Serve and Protect' debut on?", + "choices": [ + "CBS", + "NBC", + "KVOS-TV", + "FOX" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which U.S. TV station did the Canadian reality series 'To Serve and Protect' debut on?\nAnswer:", + " CBS" + ], + [ + "Question:Which U.S. TV station did the Canadian reality series 'To Serve and Protect' debut on?\nAnswer:", + " NBC" + ], + [ + "Question:Which U.S. TV station did the Canadian reality series 'To Serve and Protect' debut on?\nAnswer:", + " KVOS-TV" + ], + [ + "Question:Which U.S. TV station did the Canadian reality series 'To Serve and Protect' debut on?\nAnswer:", + " FOX" + ] + ], + "resps": [ + [ + [ + -5.485742568969727, + false + ] + ], + [ + [ + -4.589082717895508, + false + ] + ], + [ + [ + -20.12972640991211, + false + ] + ], + [ + [ + -7.1394171714782715, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.485742568969727, + false + ], + [ + -4.589082717895508, + false + ], + [ + -20.12972640991211, + false + ], + [ + -7.1394171714782715, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d47ff8aed5c4c8e891737fc835f92eaec4e7017a5027c46907f35c132f267484", + "prompt_hash": "c3c6b529c8d2cdd703d3ee5d1cb1fb9247e86a111d23697441045befcd7dedc1", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 67, + "doc": { + "question": "According to the dataset diversity analysis, which topic had the largest number of questions in SimpleQA?", + "choices": [ + "Art", + "Science & Technology", + "Politics", + "Sports" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:According to the dataset diversity analysis, which topic had the largest number of questions in SimpleQA?\nAnswer:", + " Art" + ], + [ + "Question:According to the dataset diversity analysis, which topic had the largest number of questions in SimpleQA?\nAnswer:", + " Science & Technology" + ], + [ + "Question:According to the dataset diversity analysis, which topic had the largest number of questions in SimpleQA?\nAnswer:", + " Politics" + ], + [ + "Question:According to the dataset diversity analysis, which topic had the largest number of questions in SimpleQA?\nAnswer:", + " Sports" + ] + ], + "resps": [ + [ + [ + -7.214799404144287, + false + ] + ], + [ + [ + -19.02574920654297, + false + ] + ], + [ + [ + -5.521854877471924, + false + ] + ], + [ + [ + -7.079794406890869, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.214799404144287, + false + ], + [ + -19.02574920654297, + false + ], + [ + -5.521854877471924, + false + ], + [ + -7.079794406890869, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2a14bf241290cd59c02e68a8a871233553960d2d6b92ecef98384c9d0295d751", + "prompt_hash": "457e06339f9efb9f7a4fe366fb0483f40a53b329c2841f568610f8dc85099a12", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 68, + "doc": { + "question": "What percentage of SimpleQA answers were classified as dates?", + "choices": [ + "15.3%", + "18.0%", + "24.1%", + "32.8%" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What percentage of SimpleQA answers were classified as dates?\nAnswer:", + " 15.3%" + ], + [ + "Question:What percentage of SimpleQA answers were classified as dates?\nAnswer:", + " 18.0%" + ], + [ + "Question:What percentage of SimpleQA answers were classified as dates?\nAnswer:", + " 24.1%" + ], + [ + "Question:What percentage of SimpleQA answers were classified as dates?\nAnswer:", + " 32.8%" + ] + ], + "resps": [ + [ + [ + -13.22205924987793, + false + ] + ], + [ + [ + -13.355619430541992, + false + ] + ], + [ + [ + -13.3021240234375, + false + ] + ], + [ + [ + -12.97262191772461, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.22205924987793, + false + ], + [ + -13.355619430541992, + false + ], + [ + -13.3021240234375, + false + ], + [ + -12.97262191772461, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "70fe071bc00888a605080b6809c984dd38af6f592b3cf8fac781c0cc5debfe9a", + "prompt_hash": "0d0f0a519437a8b378320b22a982baf60b834b7a212efccfd6cdbc0c8a76f27c", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 69, + "doc": { + "question": "Who received the IEEE Frank Rosenblatt Award in 2010?", + "choices": [ + "Geoffrey Hinton", + "Yoshua Bengio", + "Michio Sugeno", + "Yann LeCun" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Who received the IEEE Frank Rosenblatt Award in 2010?\nAnswer:", + " Geoffrey Hinton" + ], + [ + "Question:Who received the IEEE Frank Rosenblatt Award in 2010?\nAnswer:", + " Yoshua Bengio" + ], + [ + "Question:Who received the IEEE Frank Rosenblatt Award in 2010?\nAnswer:", + " Michio Sugeno" + ], + [ + "Question:Who received the IEEE Frank Rosenblatt Award in 2010?\nAnswer:", + " Yann LeCun" + ] + ], + "resps": [ + [ + [ + -6.306613445281982, + false + ] + ], + [ + [ + -8.654630661010742, + false + ] + ], + [ + [ + -18.297204971313477, + false + ] + ], + [ + [ + -9.778219223022461, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.306613445281982, + false + ], + [ + -8.654630661010742, + false + ], + [ + -18.297204971313477, + false + ], + [ + -9.778219223022461, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4856395e23d1b486e1805d9ddc475e84c1e94fdea8e608fec4a76c055406f2d1", + "prompt_hash": "84f9802f578f7253fd0d08d4ab097d2cca0f113c9ca1fa0bde7bd0d654aa68e9", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 70, + "doc": { + "question": "How many subjects are covered in the MMLU benchmark?", + "choices": [ + "57", + "45", + "63", + "52" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:How many subjects are covered in the MMLU benchmark?\nAnswer:", + " 57" + ], + [ + "Question:How many subjects are covered in the MMLU benchmark?\nAnswer:", + " 45" + ], + [ + "Question:How many subjects are covered in the MMLU benchmark?\nAnswer:", + " 63" + ], + [ + "Question:How many subjects are covered in the MMLU benchmark?\nAnswer:", + " 52" + ] + ], + "resps": [ + [ + [ + -6.606266021728516, + false + ] + ], + [ + [ + -6.3166961669921875, + false + ] + ], + [ + [ + -7.257956504821777, + false + ] + ], + [ + [ + -5.796786308288574, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.606266021728516, + false + ], + [ + -6.3166961669921875, + false + ], + [ + -7.257956504821777, + false + ], + [ + -5.796786308288574, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c1da0f228cb3f562ef18855bb58eb38974856f02b0c3d6b81cff1f66ca2a7330", + "prompt_hash": "f592d4ab3de93b612df4c290531f42796baea4eb2e26217eb2204b2d5797b70b", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 71, + "doc": { + "question": "What evaluation metric is used for the HumanEval dataset?", + "choices": [ + "BLEU score", + "Accuracy", + "pass@1", + "F1 score" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What evaluation metric is used for the HumanEval dataset?\nAnswer:", + " BLEU score" + ], + [ + "Question:What evaluation metric is used for the HumanEval dataset?\nAnswer:", + " Accuracy" + ], + [ + "Question:What evaluation metric is used for the HumanEval dataset?\nAnswer:", + " pass@1" + ], + [ + "Question:What evaluation metric is used for the HumanEval dataset?\nAnswer:", + " F1 score" + ] + ], + "resps": [ + [ + [ + -8.921050071716309, + false + ] + ], + [ + [ + -5.304507732391357, + false + ] + ], + [ + [ + -8.254301071166992, + false + ] + ], + [ + [ + -6.906661033630371, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.921050071716309, + false + ], + [ + -5.304507732391357, + false + ], + [ + -8.254301071166992, + false + ], + [ + -6.906661033630371, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6a4660cb3646307034774e93d85c6d72209b4ca41a5df7903df763e627d8573d", + "prompt_hash": "05e0f4eabe8be341cf59c011a598fcca173afacbaa69dd8923e7041dd37edd1d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 72, + "doc": { + "question": "Which programming languages are involved in the CODEXGLUE dataset for code translation?", + "choices": [ + "Python and Go", + "Java and C-Sharp", + "JavaScript and Ruby", + "C++ and Rust" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which programming languages are involved in the CODEXGLUE dataset for code translation?\nAnswer:", + " Python and Go" + ], + [ + "Question:Which programming languages are involved in the CODEXGLUE dataset for code translation?\nAnswer:", + " Java and C-Sharp" + ], + [ + "Question:Which programming languages are involved in the CODEXGLUE dataset for code translation?\nAnswer:", + " JavaScript and Ruby" + ], + [ + "Question:Which programming languages are involved in the CODEXGLUE dataset for code translation?\nAnswer:", + " C++ and Rust" + ] + ], + "resps": [ + [ + [ + -8.412498474121094, + false + ] + ], + [ + [ + -16.453357696533203, + false + ] + ], + [ + [ + -10.993216514587402, + false + ] + ], + [ + [ + -11.8782958984375, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.412498474121094, + false + ], + [ + -16.453357696533203, + false + ], + [ + -10.993216514587402, + false + ], + [ + -11.8782958984375, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6607f36aaafbab5635ca3d0516438a7bfb48e4f4b45e146a71ee003072753675", + "prompt_hash": "3e132e3c2fb334c40e1ba4facbfda3b456123737b67a644b42cd76489493a8eb", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 73, + "doc": { + "question": "How many total samples are included in the HumanEval dataset?", + "choices": [ + "164", + "200", + "500", + "1000" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:How many total samples are included in the HumanEval dataset?\nAnswer:", + " 164" + ], + [ + "Question:How many total samples are included in the HumanEval dataset?\nAnswer:", + " 200" + ], + [ + "Question:How many total samples are included in the HumanEval dataset?\nAnswer:", + " 500" + ], + [ + "Question:How many total samples are included in the HumanEval dataset?\nAnswer:", + " 1000" + ] + ], + "resps": [ + [ + [ + -7.63142204284668, + false + ] + ], + [ + [ + -4.791470527648926, + false + ] + ], + [ + [ + -4.627772331237793, + false + ] + ], + [ + [ + -4.572315692901611, + true + ] + ] + ], + "filtered_resps": [ + [ + -7.63142204284668, + false + ], + [ + -4.791470527648926, + false + ], + [ + -4.627772331237793, + false + ], + [ + -4.572315692901611, + true + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8a593a36da549ab39302960a4ce5b84f98c4ba2cf0ee9912a92b0161990963f4", + "prompt_hash": "aa0d9a994fe693ac1b36b62dce0fe1bcc5858e551a69b9932ce351a5b7c7c88e", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 74, + "doc": { + "question": "What is the size of the FIND benchmark subset used in evaluation?", + "choices": [ + "200 functions", + "500 functions", + "820 functions", + "1000 functions" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the size of the FIND benchmark subset used in evaluation?\nAnswer:", + " 200 functions" + ], + [ + "Question:What is the size of the FIND benchmark subset used in evaluation?\nAnswer:", + " 500 functions" + ], + [ + "Question:What is the size of the FIND benchmark subset used in evaluation?\nAnswer:", + " 820 functions" + ], + [ + "Question:What is the size of the FIND benchmark subset used in evaluation?\nAnswer:", + " 1000 functions" + ] + ], + "resps": [ + [ + [ + -14.669370651245117, + false + ] + ], + [ + [ + -14.397915840148926, + false + ] + ], + [ + [ + -19.312631607055664, + false + ] + ], + [ + [ + -14.173040390014648, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.669370651245117, + false + ], + [ + -14.397915840148926, + false + ], + [ + -19.312631607055664, + false + ], + [ + -14.173040390014648, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3a4b93b9301b5f196f596c1d88217f8055d17447b91145d6bdde5f118c870feb", + "prompt_hash": "c2457c947e0ab2dc92e423641a122338685f4821906af3091fce7438e74cd40d", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 75, + "doc": { + "question": "What was the total token count of the RedPajama-V1 dataset?", + "choices": [ + "1.2 Trillion", + "3.0 Trillion", + "878 Billion", + "2.8 Trillion" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What was the total token count of the RedPajama-V1 dataset?\nAnswer:", + " 1.2 Trillion" + ], + [ + "Question:What was the total token count of the RedPajama-V1 dataset?\nAnswer:", + " 3.0 Trillion" + ], + [ + "Question:What was the total token count of the RedPajama-V1 dataset?\nAnswer:", + " 878 Billion" + ], + [ + "Question:What was the total token count of the RedPajama-V1 dataset?\nAnswer:", + " 2.8 Trillion" + ] + ], + "resps": [ + [ + [ + -12.873974800109863, + false + ] + ], + [ + [ + -17.502328872680664, + false + ] + ], + [ + [ + -16.340862274169922, + false + ] + ], + [ + [ + -14.272342681884766, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.873974800109863, + false + ], + [ + -17.502328872680664, + false + ], + [ + -16.340862274169922, + false + ], + [ + -14.272342681884766, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "343df6d4a5c375184cbd53649f5a4ce41db0cefeec9ec624869da66e561e6f23", + "prompt_hash": "38bf56f206f6e40bed40f0873a1f9fd49f9026c986fa0c744a31b7edd92ed879", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 76, + "doc": { + "question": "Which architecture was used to train the RedPajama-INCITE models on the Summit supercomputer?", + "choices": [ + "Intel Xeon", + "IBM Power9", + "Apple M1", + "AMD EPYC" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which architecture was used to train the RedPajama-INCITE models on the Summit supercomputer?\nAnswer:", + " Intel Xeon" + ], + [ + "Question:Which architecture was used to train the RedPajama-INCITE models on the Summit supercomputer?\nAnswer:", + " IBM Power9" + ], + [ + "Question:Which architecture was used to train the RedPajama-INCITE models on the Summit supercomputer?\nAnswer:", + " Apple M1" + ], + [ + "Question:Which architecture was used to train the RedPajama-INCITE models on the Summit supercomputer?\nAnswer:", + " AMD EPYC" + ] + ], + "resps": [ + [ + [ + -10.685319900512695, + false + ] + ], + [ + [ + -16.690317153930664, + false + ] + ], + [ + [ + -14.072364807128906, + false + ] + ], + [ + [ + -12.965118408203125, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.685319900512695, + false + ], + [ + -16.690317153930664, + false + ], + [ + -14.072364807128906, + false + ], + [ + -12.965118408203125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c07272860702c841e10c51dec9727d2a1539d601ba161c4ed979bec80a7548db", + "prompt_hash": "27c2c137ee521266467ce7e2661c2de39e14ca9bd7ced8e64e883be01b78572d", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 77, + "doc": { + "question": "How many Common Crawl snapshots were used to create RedPajama-V2?", + "choices": [ + "5", + "12", + "84", + "46" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many Common Crawl snapshots were used to create RedPajama-V2?\nAnswer:", + " 5" + ], + [ + "Question:How many Common Crawl snapshots were used to create RedPajama-V2?\nAnswer:", + " 12" + ], + [ + "Question:How many Common Crawl snapshots were used to create RedPajama-V2?\nAnswer:", + " 84" + ], + [ + "Question:How many Common Crawl snapshots were used to create RedPajama-V2?\nAnswer:", + " 46" + ] + ], + "resps": [ + [ + [ + -3.648129940032959, + false + ] + ], + [ + [ + -4.735817909240723, + false + ] + ], + [ + [ + -7.701600074768066, + false + ] + ], + [ + [ + -6.892649173736572, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.648129940032959, + false + ], + [ + -4.735817909240723, + false + ], + [ + -7.701600074768066, + false + ], + [ + -6.892649173736572, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c98d0ea6b984a0c02763b82495d98ecafe07e6c8d1172f5b059883d48745bb45", + "prompt_hash": "df25701a85694241d855e290288603b96157d5f225d32e35399fc9937194fe76", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 78, + "doc": { + "question": "Which five languages are included in RedPajama-V2?", + "choices": [ + "English, German, French, Spanish, Italian", + "English, Chinese, Japanese, Korean, French", + "English, Russian, Portuguese, German, Spanish", + "English, Hindi, Arabic, French, Italian" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which five languages are included in RedPajama-V2?\nAnswer:", + " English, German, French, Spanish, Italian" + ], + [ + "Question:Which five languages are included in RedPajama-V2?\nAnswer:", + " English, Chinese, Japanese, Korean, French" + ], + [ + "Question:Which five languages are included in RedPajama-V2?\nAnswer:", + " English, Russian, Portuguese, German, Spanish" + ], + [ + "Question:Which five languages are included in RedPajama-V2?\nAnswer:", + " English, Hindi, Arabic, French, Italian" + ] + ], + "resps": [ + [ + [ + -10.228062629699707, + false + ] + ], + [ + [ + -14.981597900390625, + false + ] + ], + [ + [ + -18.74696922302246, + false + ] + ], + [ + [ + -16.782047271728516, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.228062629699707, + false + ], + [ + -14.981597900390625, + false + ], + [ + -18.74696922302246, + false + ], + [ + -16.782047271728516, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ffad02ec878dccdcdb3f69be4e6cb3c3ab7299eaf77297b1fd06c7fb4a14aaac", + "prompt_hash": "b02bb87f442385006941957c27541b2a25db5ba040cf7f8361e92ff644650c9e", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 79, + "doc": { + "question": "What was the total number of documents in the RedPajama-V2 dataset?", + "choices": [ + "32.8 Billion", + "113.3 Billion", + "87.5 Billion", + "50.7 Billion" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What was the total number of documents in the RedPajama-V2 dataset?\nAnswer:", + " 32.8 Billion" + ], + [ + "Question:What was the total number of documents in the RedPajama-V2 dataset?\nAnswer:", + " 113.3 Billion" + ], + [ + "Question:What was the total number of documents in the RedPajama-V2 dataset?\nAnswer:", + " 87.5 Billion" + ], + [ + "Question:What was the total number of documents in the RedPajama-V2 dataset?\nAnswer:", + " 50.7 Billion" + ] + ], + "resps": [ + [ + [ + -16.723188400268555, + false + ] + ], + [ + [ + -19.00351333618164, + false + ] + ], + [ + [ + -17.76470947265625, + false + ] + ], + [ + [ + -17.174829483032227, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.723188400268555, + false + ], + [ + -19.00351333618164, + false + ], + [ + -17.76470947265625, + false + ], + [ + -17.174829483032227, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "cd2499d878b3dbdfcbcc7b73b842f7265b17497e41e5ed6877c38f04e3c04df0", + "prompt_hash": "86ede63365518aa6a0c4469591e5d3a720087f05e09a59ead51ed51cdb9eabbc", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 80, + "doc": { + "question": "Which base model was fine-tuned to create Marco-o1-CoT?", + "choices": [ + "LLaMA-2-13B", + "Qwen2-7B-Instruct", + "GPT-3.5", + "Mistral-7B" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which base model was fine-tuned to create Marco-o1-CoT?\nAnswer:", + " LLaMA-2-13B" + ], + [ + "Question:Which base model was fine-tuned to create Marco-o1-CoT?\nAnswer:", + " Qwen2-7B-Instruct" + ], + [ + "Question:Which base model was fine-tuned to create Marco-o1-CoT?\nAnswer:", + " GPT-3.5" + ], + [ + "Question:Which base model was fine-tuned to create Marco-o1-CoT?\nAnswer:", + " Mistral-7B" + ] + ], + "resps": [ + [ + [ + -12.792409896850586, + false + ] + ], + [ + [ + -12.102909088134766, + false + ] + ], + [ + [ + -5.229546070098877, + false + ] + ], + [ + [ + -6.480302810668945, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.792409896850586, + false + ], + [ + -12.102909088134766, + false + ], + [ + -5.229546070098877, + false + ], + [ + -6.480302810668945, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "854e0f9653c3da0877c954e315bb63f134dabd432dad36ecd6086e36e77b6f4a", + "prompt_hash": "05924aef3a26920fc1c2502accd55e327b7cffc40a6128c2e9eb69d62435b3d8", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 81, + "doc": { + "question": "How many samples are included in the Marco-o1 CoT Dataset (Synthetic)?", + "choices": [ + "45,125", + "10,000", + "5,141", + "60,266" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many samples are included in the Marco-o1 CoT Dataset (Synthetic)?\nAnswer:", + " 45,125" + ], + [ + "Question:How many samples are included in the Marco-o1 CoT Dataset (Synthetic)?\nAnswer:", + " 10,000" + ], + [ + "Question:How many samples are included in the Marco-o1 CoT Dataset (Synthetic)?\nAnswer:", + " 5,141" + ], + [ + "Question:How many samples are included in the Marco-o1 CoT Dataset (Synthetic)?\nAnswer:", + " 60,266" + ] + ], + "resps": [ + [ + [ + -15.957291603088379, + false + ] + ], + [ + [ + -5.9744343757629395, + false + ] + ], + [ + [ + -13.689861297607422, + false + ] + ], + [ + [ + -16.420578002929688, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.957291603088379, + false + ], + [ + -5.9744343757629395, + false + ], + [ + -13.689861297607422, + false + ], + [ + -16.420578002929688, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "027cd7fa470750d66dabff86af0374caba6f6d4e90d8d1f3e1a558c7cfe74d9a", + "prompt_hash": "470ba892ff98e50aac9bef0ef10e99661216ce36d267acb86701c7b114cee2d1", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 82, + "doc": { + "question": "What improvement in accuracy did Marco-o1 achieve on the MGSM (English) dataset compared to its baseline?", + "choices": [ + "+4.10%", + "+6.17%", + "+5.60%", + "+7.25%" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What improvement in accuracy did Marco-o1 achieve on the MGSM (English) dataset compared to its baseline?\nAnswer:", + " +4.10%" + ], + [ + "Question:What improvement in accuracy did Marco-o1 achieve on the MGSM (English) dataset compared to its baseline?\nAnswer:", + " +6.17%" + ], + [ + "Question:What improvement in accuracy did Marco-o1 achieve on the MGSM (English) dataset compared to its baseline?\nAnswer:", + " +5.60%" + ], + [ + "Question:What improvement in accuracy did Marco-o1 achieve on the MGSM (English) dataset compared to its baseline?\nAnswer:", + " +7.25%" + ] + ], + "resps": [ + [ + [ + -20.366485595703125, + false + ] + ], + [ + [ + -20.923389434814453, + false + ] + ], + [ + [ + -20.500133514404297, + false + ] + ], + [ + [ + -20.437623977661133, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.366485595703125, + false + ], + [ + -20.923389434814453, + false + ], + [ + -20.500133514404297, + false + ], + [ + -20.437623977661133, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3b81c9861f63328c55eee8f9a171bc29ee04e18fc8c291667a8318746146b4d3", + "prompt_hash": "0ea6e0375ab9dec9aa8b8bf8d07007c702d0b13ff85345c2eeac5c15062d23ac", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 83, + "doc": { + "question": "In the MCTS framework, what does each node represent?", + "choices": [ + "A reasoning state", + "A dataset sample", + "A confidence score", + "A translation example" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:In the MCTS framework, what does each node represent?\nAnswer:", + " A reasoning state" + ], + [ + "Question:In the MCTS framework, what does each node represent?\nAnswer:", + " A dataset sample" + ], + [ + "Question:In the MCTS framework, what does each node represent?\nAnswer:", + " A confidence score" + ], + [ + "Question:In the MCTS framework, what does each node represent?\nAnswer:", + " A translation example" + ] + ], + "resps": [ + [ + [ + -17.532194137573242, + false + ] + ], + [ + [ + -17.33028793334961, + false + ] + ], + [ + [ + -15.44752311706543, + false + ] + ], + [ + [ + -23.827362060546875, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.532194137573242, + false + ], + [ + -17.33028793334961, + false + ], + [ + -15.44752311706543, + false + ], + [ + -23.827362060546875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "da0d104d4c28b6023871cd7c042e6100b946db17c1487fc3fdab43d0280f16ee", + "prompt_hash": "917998b7b8619495489d5230e2c452d9a10f4ea710ccf4b965290ffedde46aef", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 84, + "doc": { + "question": "What reflection prompt was added to encourage self-criticism in reasoning?", + "choices": [ + "Think again carefully before answering.", + "Wait! Maybe I made some mistakes! I need to rethink from scratch.", + "Check your reasoning one more time.", + "Be cautious, errors may occur." + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What reflection prompt was added to encourage self-criticism in reasoning?\nAnswer:", + " Think again carefully before answering." + ], + [ + "Question:What reflection prompt was added to encourage self-criticism in reasoning?\nAnswer:", + " Wait! Maybe I made some mistakes! I need to rethink from scratch." + ], + [ + "Question:What reflection prompt was added to encourage self-criticism in reasoning?\nAnswer:", + " Check your reasoning one more time." + ], + [ + "Question:What reflection prompt was added to encourage self-criticism in reasoning?\nAnswer:", + " Be cautious, errors may occur." + ] + ], + "resps": [ + [ + [ + -20.51762580871582, + false + ] + ], + [ + [ + -56.214149475097656, + false + ] + ], + [ + [ + -22.922752380371094, + false + ] + ], + [ + [ + -32.4448127746582, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.51762580871582, + false + ], + [ + -56.214149475097656, + false + ], + [ + -22.922752380371094, + false + ], + [ + -32.4448127746582, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c1db4d89e8536ae6c2c91d8b610e10f3f3c8e002eb44a61cf024ce7028741e67", + "prompt_hash": "ac16d04f80938514c60a3c23e4df6f1010417fdf9b3aeed1c3d69a4ba274c044", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 85, + "doc": { + "question": "Which base model was selected due to its strong foundational capability in mathematical reasoning?", + "choices": [ + "GPT-4o", + "LLaMA-2", + "Qwen2.5-Math-72B", + "WizardLM" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which base model was selected due to its strong foundational capability in mathematical reasoning?\nAnswer:", + " GPT-4o" + ], + [ + "Question:Which base model was selected due to its strong foundational capability in mathematical reasoning?\nAnswer:", + " LLaMA-2" + ], + [ + "Question:Which base model was selected due to its strong foundational capability in mathematical reasoning?\nAnswer:", + " Qwen2.5-Math-72B" + ], + [ + "Question:Which base model was selected due to its strong foundational capability in mathematical reasoning?\nAnswer:", + " WizardLM" + ] + ], + "resps": [ + [ + [ + -4.530148506164551, + false + ] + ], + [ + [ + -8.570853233337402, + false + ] + ], + [ + [ + -29.780563354492188, + false + ] + ], + [ + [ + -9.87548828125, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.530148506164551, + false + ], + [ + -8.570853233337402, + false + ], + [ + -29.780563354492188, + false + ], + [ + -9.87548828125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "642effc96f010f6d01ce3cc3d9e1e4da4cf9c0e2a95c61926efc5c2058a19767", + "prompt_hash": "eaba60c4953a3f517eb30c9042d6f00862ccf2dcf8c249eaba2737571adc36dd", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 86, + "doc": { + "question": "Which dataset subset was used to facilitate more extensive inference-time scaling experiments?", + "choices": [ + "AIME2024", + "MATH500", + "MATH2024", + "SimpleQA" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset subset was used to facilitate more extensive inference-time scaling experiments?\nAnswer:", + " AIME2024" + ], + [ + "Question:Which dataset subset was used to facilitate more extensive inference-time scaling experiments?\nAnswer:", + " MATH500" + ], + [ + "Question:Which dataset subset was used to facilitate more extensive inference-time scaling experiments?\nAnswer:", + " MATH2024" + ], + [ + "Question:Which dataset subset was used to facilitate more extensive inference-time scaling experiments?\nAnswer:", + " SimpleQA" + ] + ], + "resps": [ + [ + [ + -23.071943283081055, + false + ] + ], + [ + [ + -21.142967224121094, + false + ] + ], + [ + [ + -20.790672302246094, + false + ] + ], + [ + [ + -17.837215423583984, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.071943283081055, + false + ], + [ + -21.142967224121094, + false + ], + [ + -20.790672302246094, + false + ], + [ + -17.837215423583984, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "386f81490afe08849f4cdfe9479ae1093010c99a8d2b1e2c414fb2428d8ffb94", + "prompt_hash": "602a2d88d523c4d0f12f48973eb2892ed39d2caf4f98e09fdc10c6d983272287", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 87, + "doc": { + "question": "What is the total transparency score assigned in the proposed scoring framework?", + "choices": [ + "50", + "75", + "100", + "150" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the total transparency score assigned in the proposed scoring framework?\nAnswer:", + " 50" + ], + [ + "Question:What is the total transparency score assigned in the proposed scoring framework?\nAnswer:", + " 75" + ], + [ + "Question:What is the total transparency score assigned in the proposed scoring framework?\nAnswer:", + " 100" + ], + [ + "Question:What is the total transparency score assigned in the proposed scoring framework?\nAnswer:", + " 150" + ] + ], + "resps": [ + [ + [ + -6.404786109924316, + false + ] + ], + [ + [ + -7.529520511627197, + false + ] + ], + [ + [ + -5.6015729904174805, + false + ] + ], + [ + [ + -9.762228012084961, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.404786109924316, + false + ], + [ + -7.529520511627197, + false + ], + [ + -5.6015729904174805, + false + ], + [ + -9.762228012084961, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4c0d03e44a52a1380460a9e87840ed540aa0d5306090a4e992d9dffe62962bd6", + "prompt_hash": "3da4faa28aa643a53554e62566baaa9eb8c27bf624b650a137ef1d8d4a9596f3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 88, + "doc": { + "question": "From which competition were 30 additional problems curated to diversify evaluation benchmarks?", + "choices": [ + "American Invitational Mathematics Examination", + "China National High School Mathematics Competition", + "International Mathematical Olympiad", + "Putnam Competition" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:From which competition were 30 additional problems curated to diversify evaluation benchmarks?\nAnswer:", + " American Invitational Mathematics Examination" + ], + [ + "Question:From which competition were 30 additional problems curated to diversify evaluation benchmarks?\nAnswer:", + " China National High School Mathematics Competition" + ], + [ + "Question:From which competition were 30 additional problems curated to diversify evaluation benchmarks?\nAnswer:", + " International Mathematical Olympiad" + ], + [ + "Question:From which competition were 30 additional problems curated to diversify evaluation benchmarks?\nAnswer:", + " Putnam Competition" + ] + ], + "resps": [ + [ + [ + -23.391521453857422, + false + ] + ], + [ + [ + -23.65379524230957, + false + ] + ], + [ + [ + -7.464717388153076, + false + ] + ], + [ + [ + -15.607044219970703, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.391521453857422, + false + ], + [ + -23.65379524230957, + false + ], + [ + -7.464717388153076, + false + ], + [ + -15.607044219970703, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "adfa9e48b2843d6e20fbfc770b373e66f427d22f799d9f27d40b51348af6a198", + "prompt_hash": "9e3bbcb449ce9fabe9f85c2e2cc158442c7acd6cb4a610e329e729e4b9e2f200", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 89, + "doc": { + "question": "Which dataset demonstrated a drop in accuracy from 92% to 86.5% after fine-tuning?", + "choices": [ + "Flames", + "DiaSafety", + "WildSafety", + "SimpleQA" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which dataset demonstrated a drop in accuracy from 92% to 86.5% after fine-tuning?\nAnswer:", + " Flames" + ], + [ + "Question:Which dataset demonstrated a drop in accuracy from 92% to 86.5% after fine-tuning?\nAnswer:", + " DiaSafety" + ], + [ + "Question:Which dataset demonstrated a drop in accuracy from 92% to 86.5% after fine-tuning?\nAnswer:", + " WildSafety" + ], + [ + "Question:Which dataset demonstrated a drop in accuracy from 92% to 86.5% after fine-tuning?\nAnswer:", + " SimpleQA" + ] + ], + "resps": [ + [ + [ + -14.188251495361328, + false + ] + ], + [ + [ + -26.24142074584961, + false + ] + ], + [ + [ + -24.381790161132812, + false + ] + ], + [ + [ + -18.144813537597656, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.188251495361328, + false + ], + [ + -26.24142074584961, + false + ], + [ + -24.381790161132812, + false + ], + [ + -18.144813537597656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b3f4cd4668762b23da69af6238b379755d32b68bd1327203e306eaa09d42a916", + "prompt_hash": "003ceb9e9bd248a70d3c0a8df60b0e368e9458581e05b605cdd2f897ca650c2b", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 90, + "doc": { + "question": "How many screenshots were collected for web instruction-tuning data before filtering static text?", + "choices": [ + "10K", + "15K", + "22K", + "30K" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many screenshots were collected for web instruction-tuning data before filtering static text?\nAnswer:", + " 10K" + ], + [ + "Question:How many screenshots were collected for web instruction-tuning data before filtering static text?\nAnswer:", + " 15K" + ], + [ + "Question:How many screenshots were collected for web instruction-tuning data before filtering static text?\nAnswer:", + " 22K" + ], + [ + "Question:How many screenshots were collected for web instruction-tuning data before filtering static text?\nAnswer:", + " 30K" + ] + ], + "resps": [ + [ + [ + -9.913565635681152, + false + ] + ], + [ + [ + -11.398613929748535, + false + ] + ], + [ + [ + -11.466139793395996, + false + ] + ], + [ + [ + -11.128503799438477, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.913565635681152, + false + ], + [ + -11.398613929748535, + false + ], + [ + -11.466139793395996, + false + ], + [ + -11.128503799438477, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8a0701d6015b21c37484eb7e6d9e276cf922e61fa5f4dbb92dbf6161f1177900", + "prompt_hash": "818b0cb599a378908e016043ae4eb55a5c8a12bb882db7d6fdbf0191f7d4325e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 91, + "doc": { + "question": "Which dataset contains 30k instructions and 715K trajectories in an Android smartphone environment?", + "choices": [ + "Mind2Web", + "AITW", + "MiniWob", + "OmniAct" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset contains 30k instructions and 715K trajectories in an Android smartphone environment?\nAnswer:", + " Mind2Web" + ], + [ + "Question:Which dataset contains 30k instructions and 715K trajectories in an Android smartphone environment?\nAnswer:", + " AITW" + ], + [ + "Question:Which dataset contains 30k instructions and 715K trajectories in an Android smartphone environment?\nAnswer:", + " MiniWob" + ], + [ + "Question:Which dataset contains 30k instructions and 715K trajectories in an Android smartphone environment?\nAnswer:", + " OmniAct" + ] + ], + "resps": [ + [ + [ + -22.257719039916992, + false + ] + ], + [ + [ + -18.680747985839844, + false + ] + ], + [ + [ + -28.62395668029785, + false + ] + ], + [ + [ + -20.713193893432617, + false + ] + ] + ], + "filtered_resps": [ + [ + -22.257719039916992, + false + ], + [ + -18.680747985839844, + false + ], + [ + -28.62395668029785, + false + ], + [ + -20.713193893432617, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "81b2317defb712221e1a47661cf025da1057bc3cd60746839d59fcb0bb89bd4f", + "prompt_hash": "b7f9de5a9dea5182c107cef796da80fee91ca22ac47f5834e4b6adaf7faa5d31", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 92, + "doc": { + "question": "What was the learning rate configured during training?", + "choices": [ + "1e-3", + "1e-4", + "5e-5", + "2e-4" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What was the learning rate configured during training?\nAnswer:", + " 1e-3" + ], + [ + "Question:What was the learning rate configured during training?\nAnswer:", + " 1e-4" + ], + [ + "Question:What was the learning rate configured during training?\nAnswer:", + " 5e-5" + ], + [ + "Question:What was the learning rate configured during training?\nAnswer:", + " 2e-4" + ] + ], + "resps": [ + [ + [ + -6.130802631378174, + false + ] + ], + [ + [ + -5.951138019561768, + false + ] + ], + [ + [ + -7.6521711349487305, + false + ] + ], + [ + [ + -7.785113334655762, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.130802631378174, + false + ], + [ + -5.951138019561768, + false + ], + [ + -7.6521711349487305, + false + ], + [ + -7.785113334655762, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "982cb0f36b4a6afcf5533c6dd81994dbfdbc261d841b5381a135a56fae46e7f8", + "prompt_hash": "d5ad5afbf26b85078270057741fb62aedbd62983b6e895295bc67a40691eb237", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 93, + "doc": { + "question": "Which benchmark dataset for grounding is used to assess text and widget recognition separately?", + "choices": [ + "Screenspot", + "SeeClick", + "GUIAct", + "Mind2Web" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which benchmark dataset for grounding is used to assess text and widget recognition separately?\nAnswer:", + " Screenspot" + ], + [ + "Question:Which benchmark dataset for grounding is used to assess text and widget recognition separately?\nAnswer:", + " SeeClick" + ], + [ + "Question:Which benchmark dataset for grounding is used to assess text and widget recognition separately?\nAnswer:", + " GUIAct" + ], + [ + "Question:Which benchmark dataset for grounding is used to assess text and widget recognition separately?\nAnswer:", + " Mind2Web" + ] + ], + "resps": [ + [ + [ + -19.185638427734375, + false + ] + ], + [ + [ + -24.059486389160156, + false + ] + ], + [ + [ + -20.9055118560791, + false + ] + ], + [ + [ + -19.651004791259766, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.185638427734375, + false + ], + [ + -24.059486389160156, + false + ], + [ + -20.9055118560791, + false + ], + [ + -19.651004791259766, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "66ceed2da311e496a4d1ecc1d6d4a8d63337f396192d8124982e79d9abcb2364", + "prompt_hash": "5d6bf6fe135a7b5a7127295f52ca865dfc4822886083c430de092233a04d15f3", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 94, + "doc": { + "question": "On which layer was cross-layer insertion applied for the UI Connected Graph?", + "choices": [ + "Layer 10", + "Layer 12", + "Layer 14", + "Layer 16" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:On which layer was cross-layer insertion applied for the UI Connected Graph?\nAnswer:", + " Layer 10" + ], + [ + "Question:On which layer was cross-layer insertion applied for the UI Connected Graph?\nAnswer:", + " Layer 12" + ], + [ + "Question:On which layer was cross-layer insertion applied for the UI Connected Graph?\nAnswer:", + " Layer 14" + ], + [ + "Question:On which layer was cross-layer insertion applied for the UI Connected Graph?\nAnswer:", + " Layer 16" + ] + ], + "resps": [ + [ + [ + -10.43406867980957, + false + ] + ], + [ + [ + -11.22581672668457, + false + ] + ], + [ + [ + -12.237310409545898, + false + ] + ], + [ + [ + -12.027925491333008, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.43406867980957, + false + ], + [ + -11.22581672668457, + false + ], + [ + -12.237310409545898, + false + ], + [ + -12.027925491333008, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3759bdb6424860a4263ca8eb7807cd0b3cf957b42967ae7ccacb768de79db59a", + "prompt_hash": "268b4707c6a7004b4c5ad78090c712885d24a4d5fc1101ad378edc3a1324761e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 95, + "doc": { + "question": "What is the name of the proposed Vision-Language-Action model that decouples cognition and action capabilities?", + "choices": [ + "OpenVLA", + "CogACT", + "RT-2-X", + "Octo-Base" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the name of the proposed Vision-Language-Action model that decouples cognition and action capabilities?\nAnswer:", + " OpenVLA" + ], + [ + "Question:What is the name of the proposed Vision-Language-Action model that decouples cognition and action capabilities?\nAnswer:", + " CogACT" + ], + [ + "Question:What is the name of the proposed Vision-Language-Action model that decouples cognition and action capabilities?\nAnswer:", + " RT-2-X" + ], + [ + "Question:What is the name of the proposed Vision-Language-Action model that decouples cognition and action capabilities?\nAnswer:", + " Octo-Base" + ] + ], + "resps": [ + [ + [ + -13.862266540527344, + false + ] + ], + [ + [ + -16.2746524810791, + false + ] + ], + [ + [ + -24.433616638183594, + false + ] + ], + [ + [ + -24.454200744628906, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.862266540527344, + false + ], + [ + -16.2746524810791, + false + ], + [ + -24.433616638183594, + false + ], + [ + -24.454200744628906, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a04fd2b4c0456439d4f9cc93484d3f91eb1bb16fd1778cf8d5620ec2f97456fe", + "prompt_hash": "72826baf9d6f3aa5c5d683c89926c74e74e9fd893440b479c679096eb8490c92", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 96, + "doc": { + "question": "Which diffusion-based architecture is employed as the backbone for the action module?", + "choices": [ + "Multi-Layer Perceptron", + "LSTM", + "Diffusion Transformer (DiT)", + "CNN" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which diffusion-based architecture is employed as the backbone for the action module?\nAnswer:", + " Multi-Layer Perceptron" + ], + [ + "Question:Which diffusion-based architecture is employed as the backbone for the action module?\nAnswer:", + " LSTM" + ], + [ + "Question:Which diffusion-based architecture is employed as the backbone for the action module?\nAnswer:", + " Diffusion Transformer (DiT)" + ], + [ + "Question:Which diffusion-based architecture is employed as the backbone for the action module?\nAnswer:", + " CNN" + ] + ], + "resps": [ + [ + [ + -13.12600040435791, + false + ] + ], + [ + [ + -11.416516304016113, + false + ] + ], + [ + [ + -15.304282188415527, + false + ] + ], + [ + [ + -8.963671684265137, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.12600040435791, + false + ], + [ + -11.416516304016113, + false + ], + [ + -15.304282188415527, + false + ], + [ + -8.963671684265137, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "84c99fd4a4fae935da1b3c680ee3eee1afb8bc8eee00708a6aab52cb645a2158", + "prompt_hash": "17c62f6ff563bc4bf63a76d509af13a212e7d4c94849538343b05a0b5cace1f8", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 97, + "doc": { + "question": "What dataset was primarily used to train the proposed Vision-Language-Action models?", + "choices": [ + "ImageNet", + "COCO", + "Open X-Embodiment", + "EPIC-Kitchens" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What dataset was primarily used to train the proposed Vision-Language-Action models?\nAnswer:", + " ImageNet" + ], + [ + "Question:What dataset was primarily used to train the proposed Vision-Language-Action models?\nAnswer:", + " COCO" + ], + [ + "Question:What dataset was primarily used to train the proposed Vision-Language-Action models?\nAnswer:", + " Open X-Embodiment" + ], + [ + "Question:What dataset was primarily used to train the proposed Vision-Language-Action models?\nAnswer:", + " EPIC-Kitchens" + ] + ], + "resps": [ + [ + [ + -5.665737628936768, + false + ] + ], + [ + [ + -8.678854942321777, + false + ] + ], + [ + [ + -20.266983032226562, + false + ] + ], + [ + [ + -11.777584075927734, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.665737628936768, + false + ], + [ + -8.678854942321777, + false + ], + [ + -20.266983032226562, + false + ], + [ + -11.777584075927734, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "01cc2688bea7779c120bdf57ac79db47d821c9d3edbde9575a3bf14448bc120b", + "prompt_hash": "4f256133214ffff03f62ca26ee941b0d8161ed7621307af23442a559654745f1", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 98, + "doc": { + "question": "Which language model serves as the backbone of the language module in the architecture?", + "choices": [ + "GPT-4", + "LLAMA-2", + "PaLI-X", + "Phi" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which language model serves as the backbone of the language module in the architecture?\nAnswer:", + " GPT-4" + ], + [ + "Question:Which language model serves as the backbone of the language module in the architecture?\nAnswer:", + " LLAMA-2" + ], + [ + "Question:Which language model serves as the backbone of the language module in the architecture?\nAnswer:", + " PaLI-X" + ], + [ + "Question:Which language model serves as the backbone of the language module in the architecture?\nAnswer:", + " Phi" + ] + ], + "resps": [ + [ + [ + -5.79275369644165, + false + ] + ], + [ + [ + -13.931233406066895, + false + ] + ], + [ + [ + -22.94920539855957, + false + ] + ], + [ + [ + -8.201404571533203, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.79275369644165, + false + ], + [ + -13.931233406066895, + false + ], + [ + -22.94920539855957, + false + ], + [ + -8.201404571533203, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ce57ae75fb5b4ea3ce6c73e36a821d5303a5d9ba755ea4042553aed15c4eea77", + "prompt_hash": "d4e11f8f25225894ee66ed8fa09cf6996ca299426f3a35a50dfaf0d3814221f3", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 99, + "doc": { + "question": "During training, what is the default number of future actions predicted by the action module?", + "choices": [ + "3", + "7", + "15", + "31" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:During training, what is the default number of future actions predicted by the action module?\nAnswer:", + " 3" + ], + [ + "Question:During training, what is the default number of future actions predicted by the action module?\nAnswer:", + " 7" + ], + [ + "Question:During training, what is the default number of future actions predicted by the action module?\nAnswer:", + " 15" + ], + [ + "Question:During training, what is the default number of future actions predicted by the action module?\nAnswer:", + " 31" + ] + ], + "resps": [ + [ + [ + -3.34032940864563, + false + ] + ], + [ + [ + -5.836397171020508, + false + ] + ], + [ + [ + -5.689146995544434, + false + ] + ], + [ + [ + -9.099987030029297, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.34032940864563, + false + ], + [ + -5.836397171020508, + false + ], + [ + -5.689146995544434, + false + ], + [ + -9.099987030029297, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b63f832f8590826c7cb5b85e52d3dd0482b5a1714952450815f2678b2d9bf5c6", + "prompt_hash": "2901e294ac7f5a4130458e7a3a5905c6480257231e5f73acda55a32a89ff24d9", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + } + ], + "arxiv_mc_2412": [ + { + "doc_id": 0, + "doc": { + "question": "What limitation does teacher forcing impose on current next-token prediction models?", + "choices": [ + "It prevents models from being trained on continuous data.", + "It restricts the model to generating only one token at a time.", + "It offers no mechanism to guide sequence sampling toward specific objectives.", + "It reduces the model’s ability to perform tree search efficiently." + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What limitation does teacher forcing impose on current next-token prediction models?\nAnswer:", + " It prevents models from being trained on continuous data." + ], + [ + "Question:What limitation does teacher forcing impose on current next-token prediction models?\nAnswer:", + " It restricts the model to generating only one token at a time." + ], + [ + "Question:What limitation does teacher forcing impose on current next-token prediction models?\nAnswer:", + " It offers no mechanism to guide sequence sampling toward specific objectives." + ], + [ + "Question:What limitation does teacher forcing impose on current next-token prediction models?\nAnswer:", + " It reduces the model’s ability to perform tree search efficiently." + ] + ], + "resps": [ + [ + [ + -22.44562530517578, + false + ] + ], + [ + [ + -18.013652801513672, + false + ] + ], + [ + [ + -50.749053955078125, + false + ] + ], + [ + [ + -35.38025665283203, + false + ] + ] + ], + "filtered_resps": [ + [ + -22.44562530517578, + false + ], + [ + -18.013652801513672, + false + ], + [ + -50.749053955078125, + false + ], + [ + -35.38025665283203, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4c3e7569e4468951b0a98135ccc738df8d31c703938930b8b26c77fbb26f0073", + "prompt_hash": "e341aaa05fa6db270029370f8f97e06ef1dfb36d003bd27ebfc634862f06ce5b", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 1, + "doc": { + "question": "What is the key limitation of full-sequence diffusion architectures mentioned in the text?", + "choices": [ + "They fail to handle discrete data like text or audio.", + "They rely on causal attention, making them inflexible for parallel training.", + "They require variable-length token prediction, which is inefficient.", + "They are universally parameterized via non-causal, unmasked architectures." + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the key limitation of full-sequence diffusion architectures mentioned in the text?\nAnswer:", + " They fail to handle discrete data like text or audio." + ], + [ + "Question:What is the key limitation of full-sequence diffusion architectures mentioned in the text?\nAnswer:", + " They rely on causal attention, making them inflexible for parallel training." + ], + [ + "Question:What is the key limitation of full-sequence diffusion architectures mentioned in the text?\nAnswer:", + " They require variable-length token prediction, which is inefficient." + ], + [ + "Question:What is the key limitation of full-sequence diffusion architectures mentioned in the text?\nAnswer:", + " They are universally parameterized via non-causal, unmasked architectures." + ] + ], + "resps": [ + [ + [ + -30.89609718322754, + false + ] + ], + [ + [ + -49.041259765625, + false + ] + ], + [ + [ + -38.779083251953125, + false + ] + ], + [ + [ + -59.433868408203125, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.89609718322754, + false + ], + [ + -49.041259765625, + false + ], + [ + -38.779083251953125, + false + ], + [ + -59.433868408203125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f06576e73899a009a57b91ade993db970065e63efbf66ddfe429c18a1799096e", + "prompt_hash": "20f709b25998ca1921a405b0aead0207a27404bfb30309345ba3233b4bc171c9", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 2, + "doc": { + "question": "What enables \\u03b1Seq to generate variable-length sequences stably even for continuous tokens?", + "choices": [ + "Use of teacher forcing with deterministic transitions.", + "Joint modeling of latent variables with fixed-length noise schedules.", + "Denoising tokens with independent noise levels via a causal architecture.", + "Training with non-causal architectures over full-token batches." + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What enables \\u03b1Seq to generate variable-length sequences stably even for continuous tokens?\nAnswer:", + " Use of teacher forcing with deterministic transitions." + ], + [ + "Question:What enables \\u03b1Seq to generate variable-length sequences stably even for continuous tokens?\nAnswer:", + " Joint modeling of latent variables with fixed-length noise schedules." + ], + [ + "Question:What enables \\u03b1Seq to generate variable-length sequences stably even for continuous tokens?\nAnswer:", + " Denoising tokens with independent noise levels via a causal architecture." + ], + [ + "Question:What enables \\u03b1Seq to generate variable-length sequences stably even for continuous tokens?\nAnswer:", + " Training with non-causal architectures over full-token batches." + ] + ], + "resps": [ + [ + [ + -35.210548400878906, + false + ] + ], + [ + [ + -46.058746337890625, + false + ] + ], + [ + [ + -57.82805252075195, + false + ] + ], + [ + [ + -56.03277587890625, + false + ] + ] + ], + "filtered_resps": [ + [ + -35.210548400878906, + false + ], + [ + -46.058746337890625, + false + ], + [ + -57.82805252075195, + false + ], + [ + -56.03277587890625, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8b7db974c971783be24d283fc0b30af1533a468dea2747d11116c8198ecdd878", + "prompt_hash": "d06deb2e244631f4c9a0cf6c9e7bb14db1b725a527eae640a488e97c6c05f6c8", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 3, + "doc": { + "question": "What does Monte Carlo Guidance (MCG) aim to improve in the \\u03b1Seq model?", + "choices": [ + "Training speed during multi-task learning.", + "Sampling of high-reward generations.", + "Data efficiency in few-shot settings.", + "Convergence of the forward diffusion process." + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What does Monte Carlo Guidance (MCG) aim to improve in the \\u03b1Seq model?\nAnswer:", + " Training speed during multi-task learning." + ], + [ + "Question:What does Monte Carlo Guidance (MCG) aim to improve in the \\u03b1Seq model?\nAnswer:", + " Sampling of high-reward generations." + ], + [ + "Question:What does Monte Carlo Guidance (MCG) aim to improve in the \\u03b1Seq model?\nAnswer:", + " Data efficiency in few-shot settings." + ], + [ + "Question:What does Monte Carlo Guidance (MCG) aim to improve in the \\u03b1Seq model?\nAnswer:", + " Convergence of the forward diffusion process." + ] + ], + "resps": [ + [ + [ + -32.38593673706055, + false + ] + ], + [ + [ + -31.022836685180664, + false + ] + ], + [ + [ + -31.55197525024414, + false + ] + ], + [ + [ + -23.415512084960938, + false + ] + ] + ], + "filtered_resps": [ + [ + -32.38593673706055, + false + ], + [ + -31.022836685180664, + false + ], + [ + -31.55197525024414, + false + ], + [ + -23.415512084960938, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f7cda5ca84495bc9eddcf9a40c309e189f0abd73a2f7aa79e2ededc3d98c0296", + "prompt_hash": "8c9ee044d4547fc50d79a0ac9df3d33c91b8e365151aad32160dfb35be79f6ca", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 4, + "doc": { + "question": "Which of the following architectures does the paper currently use for its causal implementation?", + "choices": [ + "Feedforward Transformer", + "Bidirectional LSTM", + "Recurrent Neural Network (RNN)", + "Masked Autoencoder" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which of the following architectures does the paper currently use for its causal implementation?\nAnswer:", + " Feedforward Transformer" + ], + [ + "Question:Which of the following architectures does the paper currently use for its causal implementation?\nAnswer:", + " Bidirectional LSTM" + ], + [ + "Question:Which of the following architectures does the paper currently use for its causal implementation?\nAnswer:", + " Recurrent Neural Network (RNN)" + ], + [ + "Question:Which of the following architectures does the paper currently use for its causal implementation?\nAnswer:", + " Masked Autoencoder" + ] + ], + "resps": [ + [ + [ + -20.343334197998047, + false + ] + ], + [ + [ + -13.779508590698242, + false + ] + ], + [ + [ + -11.171441078186035, + false + ] + ], + [ + [ + -12.286684036254883, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.343334197998047, + false + ], + [ + -13.779508590698242, + false + ], + [ + -11.171441078186035, + false + ], + [ + -12.286684036254883, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f618242e0a371a70f893f9d7b535d4b66c482e1fa31a0a617ec2998fb54d3a92", + "prompt_hash": "3a808e1733399c82d28c4218096a47230cb6816c0ced6a433152bf2b40d54df9", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 5, + "doc": { + "question": "How many editing samples does the dataset \\u201cUltraEdit\\u201d contain in total?", + "choices": [ + "750,000", + "1,600,000", + "4,108,262", + "313,010" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many editing samples does the dataset \\u201cUltraEdit\\u201d contain in total?\nAnswer:", + " 750,000" + ], + [ + "Question:How many editing samples does the dataset \\u201cUltraEdit\\u201d contain in total?\nAnswer:", + " 1,600,000" + ], + [ + "Question:How many editing samples does the dataset \\u201cUltraEdit\\u201d contain in total?\nAnswer:", + " 4,108,262" + ], + [ + "Question:How many editing samples does the dataset \\u201cUltraEdit\\u201d contain in total?\nAnswer:", + " 313,010" + ] + ], + "resps": [ + [ + [ + -12.545722007751465, + false + ] + ], + [ + [ + -12.534786224365234, + false + ] + ], + [ + [ + -22.415538787841797, + false + ] + ], + [ + [ + -19.220945358276367, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.545722007751465, + false + ], + [ + -12.534786224365234, + false + ], + [ + -22.415538787841797, + false + ], + [ + -19.220945358276367, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4fe6b16f5e73298a13d1372b8ea8db250730cd4e5aa0c877cc84dabb8477e46c", + "prompt_hash": "95b2a1fd0fc42d700a482f49c3f87e054be2a261fd58cfceff9c39f8af857f1b", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 6, + "doc": { + "question": "Which model is used as the diffusion backbone in the UltraEdit data generation pipeline?", + "choices": [ + "Stable Diffusion v1.5", + "GLIDE", + "DALL-E 3", + "SDXL-Turbo" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which model is used as the diffusion backbone in the UltraEdit data generation pipeline?\nAnswer:", + " Stable Diffusion v1.5" + ], + [ + "Question:Which model is used as the diffusion backbone in the UltraEdit data generation pipeline?\nAnswer:", + " GLIDE" + ], + [ + "Question:Which model is used as the diffusion backbone in the UltraEdit data generation pipeline?\nAnswer:", + " DALL-E 3" + ], + [ + "Question:Which model is used as the diffusion backbone in the UltraEdit data generation pipeline?\nAnswer:", + " SDXL-Turbo" + ] + ], + "resps": [ + [ + [ + -8.732109069824219, + false + ] + ], + [ + [ + -11.479008674621582, + false + ] + ], + [ + [ + -9.608929634094238, + false + ] + ], + [ + [ + -17.172548294067383, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.732109069824219, + false + ], + [ + -11.479008674621582, + false + ], + [ + -9.608929634094238, + false + ], + [ + -17.172548294067383, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d861624f0803bd0d0d0f30213b808194a8a804266415f902a306b7e6ad3ba97f", + "prompt_hash": "2213363c328ebf6a27f5f0227e8d976bf54f6e1ad54e50469dfc3b221f56b641", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 7, + "doc": { + "question": "Which benchmark reveals a scaling effect when using increasing volumes of the UltraEdit dataset?", + "choices": [ + "COCO", + "MagicBrush", + "ImageNet", + "EditBench" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which benchmark reveals a scaling effect when using increasing volumes of the UltraEdit dataset?\nAnswer:", + " COCO" + ], + [ + "Question:Which benchmark reveals a scaling effect when using increasing volumes of the UltraEdit dataset?\nAnswer:", + " MagicBrush" + ], + [ + "Question:Which benchmark reveals a scaling effect when using increasing volumes of the UltraEdit dataset?\nAnswer:", + " ImageNet" + ], + [ + "Question:Which benchmark reveals a scaling effect when using increasing volumes of the UltraEdit dataset?\nAnswer:", + " EditBench" + ] + ], + "resps": [ + [ + [ + -11.662349700927734, + false + ] + ], + [ + [ + -14.004988670349121, + false + ] + ], + [ + [ + -8.855036735534668, + false + ] + ], + [ + [ + -13.624947547912598, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.662349700927734, + false + ], + [ + -14.004988670349121, + false + ], + [ + -8.855036735534668, + false + ], + [ + -13.624947547912598, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f84d5958b5aead115ff6c3b25f650dc6b2342343d5f729c3015b9b4cae0f3f81", + "prompt_hash": "40b205e24e7829798397f2138536fb831ce0cded104c4feb1999e141ba7222c2", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 8, + "doc": { + "question": "Which technique is used to identify the object to be edited in region-based editing?", + "choices": [ + "ControlNet", + "GroundingDINO with SAM", + "VQGAN-CLIP", + "DreamBooth" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which technique is used to identify the object to be edited in region-based editing?\nAnswer:", + " ControlNet" + ], + [ + "Question:Which technique is used to identify the object to be edited in region-based editing?\nAnswer:", + " GroundingDINO with SAM" + ], + [ + "Question:Which technique is used to identify the object to be edited in region-based editing?\nAnswer:", + " VQGAN-CLIP" + ], + [ + "Question:Which technique is used to identify the object to be edited in region-based editing?\nAnswer:", + " DreamBooth" + ] + ], + "resps": [ + [ + [ + -12.154390335083008, + false + ] + ], + [ + [ + -23.500587463378906, + false + ] + ], + [ + [ + -23.389801025390625, + false + ] + ], + [ + [ + -13.390340805053711, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.154390335083008, + false + ], + [ + -23.500587463378906, + false + ], + [ + -23.389801025390625, + false + ], + [ + -13.390340805053711, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f117727818829cd8265a97648f0610caa2e19403f0de0ef7e1a28057ae87a256", + "prompt_hash": "8ce78380f484ba30f5a7489d5d0fcc31c80f3e78e95cd2c13345dbbe5f98e61c", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 9, + "doc": { + "question": "What is the primary metric used in the ablation study to evaluate the impact of region-based editing data?", + "choices": [ + "DINO similarity", + "CLIPout", + "L2 distance", + "SSIM" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the primary metric used in the ablation study to evaluate the impact of region-based editing data?\nAnswer:", + " DINO similarity" + ], + [ + "Question:What is the primary metric used in the ablation study to evaluate the impact of region-based editing data?\nAnswer:", + " CLIPout" + ], + [ + "Question:What is the primary metric used in the ablation study to evaluate the impact of region-based editing data?\nAnswer:", + " L2 distance" + ], + [ + "Question:What is the primary metric used in the ablation study to evaluate the impact of region-based editing data?\nAnswer:", + " SSIM" + ] + ], + "resps": [ + [ + [ + -16.91316795349121, + false + ] + ], + [ + [ + -20.059391021728516, + false + ] + ], + [ + [ + -11.492344856262207, + false + ] + ], + [ + [ + -10.103410720825195, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.91316795349121, + false + ], + [ + -20.059391021728516, + false + ], + [ + -11.492344856262207, + false + ], + [ + -10.103410720825195, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "115786c7d725aed69d3e54b16f9664cdfbedb4f49ff34d45f98a5333a18fba7e", + "prompt_hash": "0acb114e8b7c7e6ed8c4e39c8f039af3e557b3f5863020d89b83cc60db6f9837", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 10, + "doc": { + "question": "What is the increase in coverage observed when solving CodeContests problems using Gemma-2B with repeated sampling up to 10,000 samples?", + "choices": [ + "From 0.02% to 7.1%", + "From 0.1% to 25%", + "From 0.5% to 10%", + "From 1% to 50%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the increase in coverage observed when solving CodeContests problems using Gemma-2B with repeated sampling up to 10,000 samples?\nAnswer:", + " From 0.02% to 7.1%" + ], + [ + "Question:What is the increase in coverage observed when solving CodeContests problems using Gemma-2B with repeated sampling up to 10,000 samples?\nAnswer:", + " From 0.1% to 25%" + ], + [ + "Question:What is the increase in coverage observed when solving CodeContests problems using Gemma-2B with repeated sampling up to 10,000 samples?\nAnswer:", + " From 0.5% to 10%" + ], + [ + "Question:What is the increase in coverage observed when solving CodeContests problems using Gemma-2B with repeated sampling up to 10,000 samples?\nAnswer:", + " From 1% to 50%" + ] + ], + "resps": [ + [ + [ + -36.91026306152344, + false + ] + ], + [ + [ + -33.62900161743164, + false + ] + ], + [ + [ + -31.656620025634766, + false + ] + ], + [ + [ + -26.566728591918945, + false + ] + ] + ], + "filtered_resps": [ + [ + -36.91026306152344, + false + ], + [ + -33.62900161743164, + false + ], + [ + -31.656620025634766, + false + ], + [ + -26.566728591918945, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "44df8ac43cd9caa8e4599475afb3e79cd55e4101df9a6be10f6f9382602b688c", + "prompt_hash": "9ceebe5ac95b6ba8122bf265f958d7b9b6ea4934411a663b4bdc4d54f7e0862e", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 11, + "doc": { + "question": "Which datasets evaluated in the study include automatic verifiers for sample correctness?", + "choices": [ + "MiniF2F-MATH, CodeContests, SWE-bench Lite", + "GSM8K, MATH, SWE-bench Lite", + "MATH, MiniF2F-MATH, GSM8K", + "CodeContests, GSM8K, MATH" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which datasets evaluated in the study include automatic verifiers for sample correctness?\nAnswer:", + " MiniF2F-MATH, CodeContests, SWE-bench Lite" + ], + [ + "Question:Which datasets evaluated in the study include automatic verifiers for sample correctness?\nAnswer:", + " GSM8K, MATH, SWE-bench Lite" + ], + [ + "Question:Which datasets evaluated in the study include automatic verifiers for sample correctness?\nAnswer:", + " MATH, MiniF2F-MATH, GSM8K" + ], + [ + "Question:Which datasets evaluated in the study include automatic verifiers for sample correctness?\nAnswer:", + " CodeContests, GSM8K, MATH" + ] + ], + "resps": [ + [ + [ + -87.04169464111328, + false + ] + ], + [ + [ + -50.65325927734375, + false + ] + ], + [ + [ + -45.92802429199219, + false + ] + ], + [ + [ + -28.259180068969727, + false + ] + ] + ], + "filtered_resps": [ + [ + -87.04169464111328, + false + ], + [ + -50.65325927734375, + false + ], + [ + -45.92802429199219, + false + ], + [ + -28.259180068969727, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "16350209d0fd2de900dc667b7547b1c1acda347ffe7ad0e075a33bacbcd0b5ae", + "prompt_hash": "e7aca0fd8e0b4655d141766d237513ee15a41bedc0801a38614753bd866ea52b", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 12, + "doc": { + "question": "What type of function is used to model the relationship between coverage and number of samples?", + "choices": [ + "Exponentiated power law", + "Linear regression", + "Logarithmic function", + "Quadratic polynomial" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What type of function is used to model the relationship between coverage and number of samples?\nAnswer:", + " Exponentiated power law" + ], + [ + "Question:What type of function is used to model the relationship between coverage and number of samples?\nAnswer:", + " Linear regression" + ], + [ + "Question:What type of function is used to model the relationship between coverage and number of samples?\nAnswer:", + " Logarithmic function" + ], + [ + "Question:What type of function is used to model the relationship between coverage and number of samples?\nAnswer:", + " Quadratic polynomial" + ] + ], + "resps": [ + [ + [ + -24.710233688354492, + false + ] + ], + [ + [ + -7.671542167663574, + false + ] + ], + [ + [ + -11.807942390441895, + false + ] + ], + [ + [ + -13.855320930480957, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.710233688354492, + false + ], + [ + -7.671542167663574, + false + ], + [ + -11.807942390441895, + false + ], + [ + -13.855320930480957, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "871053292bba086d3b3b812b1c4b2bfff9fdf77ff9831ebdc84d75742938d72a", + "prompt_hash": "a33a8b9084d669044175e7ce9033ca380ab9718a2aff6a835db8f2709a7e5c73", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 13, + "doc": { + "question": "What was the maximum percentage of issues solved using DeepSeek-Coder-V2-Instruct with 5 attempts per issue on SWE-bench Lite?", + "choices": [ + "29.62%", + "43.00%", + "15.90%", + "56.00%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What was the maximum percentage of issues solved using DeepSeek-Coder-V2-Instruct with 5 attempts per issue on SWE-bench Lite?\nAnswer:", + " 29.62%" + ], + [ + "Question:What was the maximum percentage of issues solved using DeepSeek-Coder-V2-Instruct with 5 attempts per issue on SWE-bench Lite?\nAnswer:", + " 43.00%" + ], + [ + "Question:What was the maximum percentage of issues solved using DeepSeek-Coder-V2-Instruct with 5 attempts per issue on SWE-bench Lite?\nAnswer:", + " 15.90%" + ], + [ + "Question:What was the maximum percentage of issues solved using DeepSeek-Coder-V2-Instruct with 5 attempts per issue on SWE-bench Lite?\nAnswer:", + " 56.00%" + ] + ], + "resps": [ + [ + [ + -16.0406494140625, + false + ] + ], + [ + [ + -15.355518341064453, + false + ] + ], + [ + [ + -16.611568450927734, + false + ] + ], + [ + [ + -15.058615684509277, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.0406494140625, + false + ], + [ + -15.355518341064453, + false + ], + [ + -16.611568450927734, + false + ], + [ + -15.058615684509277, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "664e791dfb8e473eb440962fb2801a946420e6aacb189738b6b4cd3a6e590f88", + "prompt_hash": "5a1317a0c04f5290b0304168b6d6d77eb802b415403ee25b29fdaaf46cc42af3", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 14, + "doc": { + "question": "Which verification methods were tested for their scalability in identifying correct answers on GSM8K and MATH datasets?", + "choices": [ + "Majority vote, reward model best-of-N, reward model weighted vote", + "Gradient descent, beam search, ensemble voting", + "Prompt engineering, token sampling, output reranking", + "Syntax parsing, lexical similarity, token classification" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which verification methods were tested for their scalability in identifying correct answers on GSM8K and MATH datasets?\nAnswer:", + " Majority vote, reward model best-of-N, reward model weighted vote" + ], + [ + "Question:Which verification methods were tested for their scalability in identifying correct answers on GSM8K and MATH datasets?\nAnswer:", + " Gradient descent, beam search, ensemble voting" + ], + [ + "Question:Which verification methods were tested for their scalability in identifying correct answers on GSM8K and MATH datasets?\nAnswer:", + " Prompt engineering, token sampling, output reranking" + ], + [ + "Question:Which verification methods were tested for their scalability in identifying correct answers on GSM8K and MATH datasets?\nAnswer:", + " Syntax parsing, lexical similarity, token classification" + ] + ], + "resps": [ + [ + [ + -59.267059326171875, + false + ] + ], + [ + [ + -35.804481506347656, + false + ] + ], + [ + [ + -43.54387664794922, + false + ] + ], + [ + [ + -39.79151153564453, + false + ] + ] + ], + "filtered_resps": [ + [ + -59.267059326171875, + false + ], + [ + -35.804481506347656, + false + ], + [ + -43.54387664794922, + false + ], + [ + -39.79151153564453, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0b9f0bba11615e1363d62aca9634d306bd14429c924a8ad6885c2983578308c8", + "prompt_hash": "0d17baf56754596eaed49b4028801899ec6ba4fbb02532f3aa951b5948115bb2", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 15, + "doc": { + "question": "What Dice score does MedSAM-2 achieve on the BTCV 3D medical image segmentation dataset?", + "choices": [ + "86.2%", + "88.4%", + "89.0%", + "90.5%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What Dice score does MedSAM-2 achieve on the BTCV 3D medical image segmentation dataset?\nAnswer:", + " 86.2%" + ], + [ + "Question:What Dice score does MedSAM-2 achieve on the BTCV 3D medical image segmentation dataset?\nAnswer:", + " 88.4%" + ], + [ + "Question:What Dice score does MedSAM-2 achieve on the BTCV 3D medical image segmentation dataset?\nAnswer:", + " 89.0%" + ], + [ + "Question:What Dice score does MedSAM-2 achieve on the BTCV 3D medical image segmentation dataset?\nAnswer:", + " 90.5%" + ] + ], + "resps": [ + [ + [ + -14.722869873046875, + false + ] + ], + [ + [ + -15.132040977478027, + false + ] + ], + [ + [ + -15.362841606140137, + false + ] + ], + [ + [ + -15.039222717285156, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.722869873046875, + false + ], + [ + -15.132040977478027, + false + ], + [ + -15.362841606140137, + false + ], + [ + -15.039222717285156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "071c9543500216fe3fa54b274fda3405e1560d004da2672b30ff527777bbc92b", + "prompt_hash": "3a19aaa4ffdf21a14e504172df6f1b5b9f69e86067c43f481ddeb35b33473d2f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 16, + "doc": { + "question": "What is the primary function of the self-sorting memory bank in MedSAM-2?", + "choices": [ + "To store only the most recent embeddings", + "To sort image slices by acquisition order", + "To dynamically select embeddings based on confidence and dissimilarity", + "To increase the memory capacity of the model" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the primary function of the self-sorting memory bank in MedSAM-2?\nAnswer:", + " To store only the most recent embeddings" + ], + [ + "Question:What is the primary function of the self-sorting memory bank in MedSAM-2?\nAnswer:", + " To sort image slices by acquisition order" + ], + [ + "Question:What is the primary function of the self-sorting memory bank in MedSAM-2?\nAnswer:", + " To dynamically select embeddings based on confidence and dissimilarity" + ], + [ + "Question:What is the primary function of the self-sorting memory bank in MedSAM-2?\nAnswer:", + " To increase the memory capacity of the model" + ] + ], + "resps": [ + [ + [ + -27.875429153442383, + false + ] + ], + [ + [ + -35.84728240966797, + false + ] + ], + [ + [ + -43.798126220703125, + false + ] + ], + [ + [ + -17.682130813598633, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.875429153442383, + false + ], + [ + -35.84728240966797, + false + ], + [ + -43.798126220703125, + false + ], + [ + -17.682130813598633, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "85e9263b05f7c959e70b80cb683ee5ed4496855a5dcd697bf4d14086b8fefbd0", + "prompt_hash": "f49c6b7c63e3b7b2d841b5acf05e3014c385f72a0ee3273b1642084679e24684", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 17, + "doc": { + "question": "What prompt probability is used for 3D medical image frames during segmentation?", + "choices": [ + "0.1", + "0.25", + "0.3", + "0.5" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What prompt probability is used for 3D medical image frames during segmentation?\nAnswer:", + " 0.1" + ], + [ + "Question:What prompt probability is used for 3D medical image frames during segmentation?\nAnswer:", + " 0.25" + ], + [ + "Question:What prompt probability is used for 3D medical image frames during segmentation?\nAnswer:", + " 0.3" + ], + [ + "Question:What prompt probability is used for 3D medical image frames during segmentation?\nAnswer:", + " 0.5" + ] + ], + "resps": [ + [ + [ + -4.572741985321045, + false + ] + ], + [ + [ + -6.072905540466309, + false + ] + ], + [ + [ + -5.138263702392578, + false + ] + ], + [ + [ + -4.633764743804932, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.572741985321045, + false + ], + [ + -6.072905540466309, + false + ], + [ + -5.138263702392578, + false + ], + [ + -4.633764743804932, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f2344d55b6c745c4d2dc99b6f0644d1836baaef4bb9c0b3d1b2b74620fae1f41", + "prompt_hash": "564b43b57c335dd664c18821b1478abd181701f4a2f606c13138dcf6769ee592", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 18, + "doc": { + "question": "In the ablation study, what Dice score was achieved on the CadVidSet dataset when using all three components: IOU threshold, dissimilar templates, and resampling?", + "choices": [ + "53.9%", + "57.8%", + "64.5%", + "72.9%" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:In the ablation study, what Dice score was achieved on the CadVidSet dataset when using all three components: IOU threshold, dissimilar templates, and resampling?\nAnswer:", + " 53.9%" + ], + [ + "Question:In the ablation study, what Dice score was achieved on the CadVidSet dataset when using all three components: IOU threshold, dissimilar templates, and resampling?\nAnswer:", + " 57.8%" + ], + [ + "Question:In the ablation study, what Dice score was achieved on the CadVidSet dataset when using all three components: IOU threshold, dissimilar templates, and resampling?\nAnswer:", + " 64.5%" + ], + [ + "Question:In the ablation study, what Dice score was achieved on the CadVidSet dataset when using all three components: IOU threshold, dissimilar templates, and resampling?\nAnswer:", + " 72.9%" + ] + ], + "resps": [ + [ + [ + -16.8532657623291, + false + ] + ], + [ + [ + -16.828392028808594, + false + ] + ], + [ + [ + -16.361194610595703, + false + ] + ], + [ + [ + -15.42495059967041, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.8532657623291, + false + ], + [ + -16.828392028808594, + false + ], + [ + -16.361194610595703, + false + ], + [ + -15.42495059967041, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9a6a818e319d1a38943ed39f3744bf0ce7ff570586e188b790e8d99e5f098a92", + "prompt_hash": "feb4610e33e35a2ab7303ab190151070e7ca5d88a02104fb86048ba3508510f8", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 19, + "doc": { + "question": "How many GPUs were used for training and testing MedSAM-2?", + "choices": [ + "32", + "64", + "128", + "256" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many GPUs were used for training and testing MedSAM-2?\nAnswer:", + " 32" + ], + [ + "Question:How many GPUs were used for training and testing MedSAM-2?\nAnswer:", + " 64" + ], + [ + "Question:How many GPUs were used for training and testing MedSAM-2?\nAnswer:", + " 128" + ], + [ + "Question:How many GPUs were used for training and testing MedSAM-2?\nAnswer:", + " 256" + ] + ], + "resps": [ + [ + [ + -5.806703567504883, + false + ] + ], + [ + [ + -7.463939666748047, + false + ] + ], + [ + [ + -7.4690446853637695, + false + ] + ], + [ + [ + -8.828374862670898, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.806703567504883, + false + ], + [ + -7.463939666748047, + false + ], + [ + -7.4690446853637695, + false + ], + [ + -8.828374862670898, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a4494f34db05155b1bb4cb1416cd7e0a4c3c3f82427a97adb362ed5eadf60cd0", + "prompt_hash": "74c4c7819aed8c050224abc259669dfd661560212a0ea91485acf1c392669c06", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 20, + "doc": { + "question": "What is the maximum context length supported by the LongVILA training system?", + "choices": [ + "1 million tokens", + "1.5 million tokens", + "2 million tokens", + "2.5 million tokens" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the maximum context length supported by the LongVILA training system?\nAnswer:", + " 1 million tokens" + ], + [ + "Question:What is the maximum context length supported by the LongVILA training system?\nAnswer:", + " 1.5 million tokens" + ], + [ + "Question:What is the maximum context length supported by the LongVILA training system?\nAnswer:", + " 2 million tokens" + ], + [ + "Question:What is the maximum context length supported by the LongVILA training system?\nAnswer:", + " 2.5 million tokens" + ] + ], + "resps": [ + [ + [ + -10.128260612487793, + false + ] + ], + [ + [ + -12.194124221801758, + false + ] + ], + [ + [ + -11.196681022644043, + false + ] + ], + [ + [ + -13.080391883850098, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.128260612487793, + false + ], + [ + -12.194124221801758, + false + ], + [ + -11.196681022644043, + false + ], + [ + -13.080391883850098, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a141a5184d5a62903bf04e38d70a7794df07219d83153e0c1b68e2a515c1804e", + "prompt_hash": "56db71b047e286f882824cf68ce6953f0db72790be1553e20bb18eedb012cfac", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 21, + "doc": { + "question": "Which video benchmark did LongVILA-7B achieve 65.1% accuracy on with subtitles?", + "choices": [ + "ActivityNet-QA", + "MVBench", + "VideoMME", + "NExT-QA" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which video benchmark did LongVILA-7B achieve 65.1% accuracy on with subtitles?\nAnswer:", + " ActivityNet-QA" + ], + [ + "Question:Which video benchmark did LongVILA-7B achieve 65.1% accuracy on with subtitles?\nAnswer:", + " MVBench" + ], + [ + "Question:Which video benchmark did LongVILA-7B achieve 65.1% accuracy on with subtitles?\nAnswer:", + " VideoMME" + ], + [ + "Question:Which video benchmark did LongVILA-7B achieve 65.1% accuracy on with subtitles?\nAnswer:", + " NExT-QA" + ] + ], + "resps": [ + [ + [ + -16.321186065673828, + false + ] + ], + [ + [ + -9.568792343139648, + false + ] + ], + [ + [ + -14.908329963684082, + false + ] + ], + [ + [ + -15.955806732177734, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.321186065673828, + false + ], + [ + -9.568792343139648, + false + ], + [ + -14.908329963684082, + false + ], + [ + -15.955806732177734, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a5cd9da82e58ee1724813d414580a6214f097549786cc1672569bcd8345ed30c", + "prompt_hash": "1b2dd67bbfee3227cc9ed70d506261123a60e49132a90b5d4f7604645c7fc245", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 22, + "doc": { + "question": "During Stage 1 of training, which components are trainable in the LongVILA pipeline?", + "choices": [ + "Only the LLM decoder", + "Only the vision encoder", + "Only the multi-modal projector", + "All components" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:During Stage 1 of training, which components are trainable in the LongVILA pipeline?\nAnswer:", + " Only the LLM decoder" + ], + [ + "Question:During Stage 1 of training, which components are trainable in the LongVILA pipeline?\nAnswer:", + " Only the vision encoder" + ], + [ + "Question:During Stage 1 of training, which components are trainable in the LongVILA pipeline?\nAnswer:", + " Only the multi-modal projector" + ], + [ + "Question:During Stage 1 of training, which components are trainable in the LongVILA pipeline?\nAnswer:", + " All components" + ] + ], + "resps": [ + [ + [ + -14.713794708251953, + false + ] + ], + [ + [ + -11.596830368041992, + false + ] + ], + [ + [ + -20.188129425048828, + false + ] + ], + [ + [ + -5.85800313949585, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.713794708251953, + false + ], + [ + -11.596830368041992, + false + ], + [ + -20.188129425048828, + false + ], + [ + -5.85800313949585, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b47ec35348a110052f19639f2b9499bc0c46cf8429488ca42df5e378e1202a22", + "prompt_hash": "a2b2d866314cdfb5427c81ef7684c3e5c20dfb869a71d5187e6a42b9dd54261d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 23, + "doc": { + "question": "Which dataset is used in Stage 4 to extend the context length of the LLM?", + "choices": [ + "COYO-25M", + "SlimPajama", + "YouCook2", + "Shot2Story" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset is used in Stage 4 to extend the context length of the LLM?\nAnswer:", + " COYO-25M" + ], + [ + "Question:Which dataset is used in Stage 4 to extend the context length of the LLM?\nAnswer:", + " SlimPajama" + ], + [ + "Question:Which dataset is used in Stage 4 to extend the context length of the LLM?\nAnswer:", + " YouCook2" + ], + [ + "Question:Which dataset is used in Stage 4 to extend the context length of the LLM?\nAnswer:", + " Shot2Story" + ] + ], + "resps": [ + [ + [ + -27.8762264251709, + false + ] + ], + [ + [ + -17.669025421142578, + false + ] + ], + [ + [ + -21.235336303710938, + false + ] + ], + [ + [ + -26.448081970214844, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.8762264251709, + false + ], + [ + -17.669025421142578, + false + ], + [ + -21.235336303710938, + false + ], + [ + -26.448081970214844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a4faf5c9264ce653b3bedfdacea3d798fea9a79474a73ef34ecd0ee6cfdb95ed", + "prompt_hash": "32cb48456ade5784ba620662840f9a7e452b30ccc3f89208bd37508e8fc34185", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 24, + "doc": { + "question": "How many videos are included in the LongVILA long video supervised fine-tuning dataset?", + "choices": [ + "12,345", + "15,292", + "18,764", + "20,001" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many videos are included in the LongVILA long video supervised fine-tuning dataset?\nAnswer:", + " 12,345" + ], + [ + "Question:How many videos are included in the LongVILA long video supervised fine-tuning dataset?\nAnswer:", + " 15,292" + ], + [ + "Question:How many videos are included in the LongVILA long video supervised fine-tuning dataset?\nAnswer:", + " 18,764" + ], + [ + "Question:How many videos are included in the LongVILA long video supervised fine-tuning dataset?\nAnswer:", + " 20,001" + ] + ], + "resps": [ + [ + [ + -13.590742111206055, + false + ] + ], + [ + [ + -14.44241714477539, + false + ] + ], + [ + [ + -14.642477035522461, + false + ] + ], + [ + [ + -13.420373916625977, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.590742111206055, + false + ], + [ + -14.44241714477539, + false + ], + [ + -14.642477035522461, + false + ], + [ + -13.420373916625977, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d46fb59bfb0fb6f34a89faf012d5d165c19301c9ff502c47f685e65636163fab", + "prompt_hash": "5db7d08dc8a19b42b72123d0cf2b236d0a4105c971f8cffbcdc8c9f1ca6ea2e2", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 25, + "doc": { + "question": "What is the main function of the Visual Summarization Token (VST) in Video-XL?", + "choices": [ + "To reduce model size by pruning layers", + "To summarize visual information within intervals into compressed KVs", + "To generate subtitles for video content", + "To align visual tokens with language tokens" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main function of the Visual Summarization Token (VST) in Video-XL?\nAnswer:", + " To reduce model size by pruning layers" + ], + [ + "Question:What is the main function of the Visual Summarization Token (VST) in Video-XL?\nAnswer:", + " To summarize visual information within intervals into compressed KVs" + ], + [ + "Question:What is the main function of the Visual Summarization Token (VST) in Video-XL?\nAnswer:", + " To generate subtitles for video content" + ], + [ + "Question:What is the main function of the Visual Summarization Token (VST) in Video-XL?\nAnswer:", + " To align visual tokens with language tokens" + ] + ], + "resps": [ + [ + [ + -30.24161720275879, + false + ] + ], + [ + [ + -52.90513229370117, + false + ] + ], + [ + [ + -17.981550216674805, + false + ] + ], + [ + [ + -19.51709747314453, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.24161720275879, + false + ], + [ + -52.90513229370117, + false + ], + [ + -17.981550216674805, + false + ], + [ + -19.51709747314453, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ab6b3e9da155f4e091c7adca38ca3f2181393cff5bcd17547010c14389815ad3", + "prompt_hash": "4f3a04b3e1ee5bb04e70300064c2ec7ae91e27ae9093f854d97d886610c6e9c8", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 26, + "doc": { + "question": "Which visual encoder does Video-XL use to process input images?", + "choices": [ + "ViViT", + "TimeSformer", + "CLIP-ViT-L", + "Swin Transformer" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which visual encoder does Video-XL use to process input images?\nAnswer:", + " ViViT" + ], + [ + "Question:Which visual encoder does Video-XL use to process input images?\nAnswer:", + " TimeSformer" + ], + [ + "Question:Which visual encoder does Video-XL use to process input images?\nAnswer:", + " CLIP-ViT-L" + ], + [ + "Question:Which visual encoder does Video-XL use to process input images?\nAnswer:", + " Swin Transformer" + ] + ], + "resps": [ + [ + [ + -14.41057300567627, + false + ] + ], + [ + [ + -13.634620666503906, + false + ] + ], + [ + [ + -14.097136497497559, + false + ] + ], + [ + [ + -12.50584888458252, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.41057300567627, + false + ], + [ + -13.634620666503906, + false + ], + [ + -14.097136497497559, + false + ], + [ + -12.50584888458252, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "19029acc1632e0ce04138bde642186adfdb088364c6e22820cbced7dbd8b47a7", + "prompt_hash": "48a33f308bc8305cb506bf394dd05b82a1d9948d6b21d4320d2a0e16afcbb07d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 27, + "doc": { + "question": "What technique does Video-XL use to enable training with increasing compression ratios?", + "choices": [ + "Dropout regularization", + "Curriculum learning", + "Contrastive learning", + "Token shuffling" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What technique does Video-XL use to enable training with increasing compression ratios?\nAnswer:", + " Dropout regularization" + ], + [ + "Question:What technique does Video-XL use to enable training with increasing compression ratios?\nAnswer:", + " Curriculum learning" + ], + [ + "Question:What technique does Video-XL use to enable training with increasing compression ratios?\nAnswer:", + " Contrastive learning" + ], + [ + "Question:What technique does Video-XL use to enable training with increasing compression ratios?\nAnswer:", + " Token shuffling" + ] + ], + "resps": [ + [ + [ + -15.249528884887695, + false + ] + ], + [ + [ + -12.305272102355957, + false + ] + ], + [ + [ + -10.114755630493164, + false + ] + ], + [ + [ + -15.036741256713867, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.249528884887695, + false + ], + [ + -12.305272102355957, + false + ], + [ + -10.114755630493164, + false + ], + [ + -15.036741256713867, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "05ed3ed749c5082a395841f692f4e756ff6c490ecc6fd11c77bd2c9d78f01e2f", + "prompt_hash": "9326395753e20d8f6a210f80607a0defaf3caee8d5cf3017c7650ac04added78", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 28, + "doc": { + "question": "How many frames can Video-XL process using a single A100-80GB GPU?", + "choices": [ + "1024", + "4096", + "2048", + "512" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many frames can Video-XL process using a single A100-80GB GPU?\nAnswer:", + " 1024" + ], + [ + "Question:How many frames can Video-XL process using a single A100-80GB GPU?\nAnswer:", + " 4096" + ], + [ + "Question:How many frames can Video-XL process using a single A100-80GB GPU?\nAnswer:", + " 2048" + ], + [ + "Question:How many frames can Video-XL process using a single A100-80GB GPU?\nAnswer:", + " 512" + ] + ], + "resps": [ + [ + [ + -6.188669204711914, + false + ] + ], + [ + [ + -7.195087432861328, + false + ] + ], + [ + [ + -6.599781513214111, + false + ] + ], + [ + [ + -5.575624465942383, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.188669204711914, + false + ], + [ + -7.195087432861328, + false + ], + [ + -6.599781513214111, + false + ], + [ + -5.575624465942383, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7b7c34cd606d716032e76d0177ab07a608e8af518d855eb8985638d584495d78", + "prompt_hash": "976819c662fd052e46e58e47c454e8d74ee47a3c9d5efeff61089cb2f90e7296", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 29, + "doc": { + "question": "What is the name of the synthetic dataset introduced to improve multi-detail retrieval capabilities in Video-XL?", + "choices": [ + "VideoQA++", + "V-Language", + "VICO", + "VisMulti" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the name of the synthetic dataset introduced to improve multi-detail retrieval capabilities in Video-XL?\nAnswer:", + " VideoQA++" + ], + [ + "Question:What is the name of the synthetic dataset introduced to improve multi-detail retrieval capabilities in Video-XL?\nAnswer:", + " V-Language" + ], + [ + "Question:What is the name of the synthetic dataset introduced to improve multi-detail retrieval capabilities in Video-XL?\nAnswer:", + " VICO" + ], + [ + "Question:What is the name of the synthetic dataset introduced to improve multi-detail retrieval capabilities in Video-XL?\nAnswer:", + " VisMulti" + ] + ], + "resps": [ + [ + [ + -13.095142364501953, + false + ] + ], + [ + [ + -17.66976547241211, + false + ] + ], + [ + [ + -13.14517879486084, + false + ] + ], + [ + [ + -18.147605895996094, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.095142364501953, + false + ], + [ + -17.66976547241211, + false + ], + [ + -13.14517879486084, + false + ], + [ + -18.147605895996094, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ee12ffee2d13e230f203929df631c9db7b4bb411b9d369b3cd82831de0779b9a", + "prompt_hash": "46c9ab5339134e7daf84834412de21e0f378044a50bd1769753d972666508580", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 30, + "doc": { + "question": "How many problems in the Omni-MATH dataset are sourced from the HMMT_2 contest?", + "choices": [ + "896", + "1385", + "106", + "75" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many problems in the Omni-MATH dataset are sourced from the HMMT_2 contest?\nAnswer:", + " 896" + ], + [ + "Question:How many problems in the Omni-MATH dataset are sourced from the HMMT_2 contest?\nAnswer:", + " 1385" + ], + [ + "Question:How many problems in the Omni-MATH dataset are sourced from the HMMT_2 contest?\nAnswer:", + " 106" + ], + [ + "Question:How many problems in the Omni-MATH dataset are sourced from the HMMT_2 contest?\nAnswer:", + " 75" + ] + ], + "resps": [ + [ + [ + -10.793913841247559, + false + ] + ], + [ + [ + -12.004586219787598, + false + ] + ], + [ + [ + -7.941033363342285, + false + ] + ], + [ + [ + -7.036806106567383, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.793913841247559, + false + ], + [ + -12.004586219787598, + false + ], + [ + -7.941033363342285, + false + ], + [ + -7.036806106567383, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0a5e5235b6e1fe059e6bf460d653870620d294c41e45b803d11e32dcdf5a318f", + "prompt_hash": "ace468d42a70969b8a40545c6810c92b0928178610da37347fa2ff704a75febe", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 31, + "doc": { + "question": "What is the maximum value of the trend intensity metric \\\\( \\\\mathcal{A} \\\\) defined in the difficulty consistency analysis?", + "choices": [ + "75", + "200", + "-200", + "100" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the maximum value of the trend intensity metric \\\\( \\\\mathcal{A} \\\\) defined in the difficulty consistency analysis?\nAnswer:", + " 75" + ], + [ + "Question:What is the maximum value of the trend intensity metric \\\\( \\\\mathcal{A} \\\\) defined in the difficulty consistency analysis?\nAnswer:", + " 200" + ], + [ + "Question:What is the maximum value of the trend intensity metric \\\\( \\\\mathcal{A} \\\\) defined in the difficulty consistency analysis?\nAnswer:", + " -200" + ], + [ + "Question:What is the maximum value of the trend intensity metric \\\\( \\\\mathcal{A} \\\\) defined in the difficulty consistency analysis?\nAnswer:", + " 100" + ] + ], + "resps": [ + [ + [ + -7.222622871398926, + false + ] + ], + [ + [ + -7.002682685852051, + false + ] + ], + [ + [ + -11.823040962219238, + false + ] + ], + [ + [ + -5.583542346954346, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.222622871398926, + false + ], + [ + -7.002682685852051, + false + ], + [ + -11.823040962219238, + false + ], + [ + -5.583542346954346, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "95c164209da12cf48779f8fd02937d4b49ef90e767068725ecef87b8de14fe2f", + "prompt_hash": "ebaff398c564036b069093bb1d4499c667f7c8f74c86f8941bcd4e0cfb64d5d6", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 32, + "doc": { + "question": "Which error type is identified as the most frequent during process-level assessment of model-generated mathematical solutions?", + "choices": [ + "Calculation", + "Accumulation", + "Unrelated", + "Logic" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which error type is identified as the most frequent during process-level assessment of model-generated mathematical solutions?\nAnswer:", + " Calculation" + ], + [ + "Question:Which error type is identified as the most frequent during process-level assessment of model-generated mathematical solutions?\nAnswer:", + " Accumulation" + ], + [ + "Question:Which error type is identified as the most frequent during process-level assessment of model-generated mathematical solutions?\nAnswer:", + " Unrelated" + ], + [ + "Question:Which error type is identified as the most frequent during process-level assessment of model-generated mathematical solutions?\nAnswer:", + " Logic" + ] + ], + "resps": [ + [ + [ + -8.842470169067383, + false + ] + ], + [ + [ + -12.819282531738281, + false + ] + ], + [ + [ + -12.745844841003418, + false + ] + ], + [ + [ + -11.778804779052734, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.842470169067383, + false + ], + [ + -12.819282531738281, + false + ], + [ + -12.745844841003418, + false + ], + [ + -11.778804779052734, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a965a9a067cdd28c14092a1c2c8424370bd560fe3dbe71408eb6149645fc2916", + "prompt_hash": "055439271e3814e4eee77b7f8ca47bd8364a1de10e29066297dd459b83408328", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 33, + "doc": { + "question": "Which model achieved the highest consistency rate with GPT-4o in the Omni-Judge evaluation?", + "choices": [ + "LLaMA-2-7b-Chat", + "LLaMA-3-8b-Instruct", + "LLaMA-3.1-8b-Instruct", + "MetaLlama-3.1-70B-instruct" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model achieved the highest consistency rate with GPT-4o in the Omni-Judge evaluation?\nAnswer:", + " LLaMA-2-7b-Chat" + ], + [ + "Question:Which model achieved the highest consistency rate with GPT-4o in the Omni-Judge evaluation?\nAnswer:", + " LLaMA-3-8b-Instruct" + ], + [ + "Question:Which model achieved the highest consistency rate with GPT-4o in the Omni-Judge evaluation?\nAnswer:", + " LLaMA-3.1-8b-Instruct" + ], + [ + "Question:Which model achieved the highest consistency rate with GPT-4o in the Omni-Judge evaluation?\nAnswer:", + " MetaLlama-3.1-70B-instruct" + ] + ], + "resps": [ + [ + [ + -16.803985595703125, + false + ] + ], + [ + [ + -15.235649108886719, + false + ] + ], + [ + [ + -29.402576446533203, + false + ] + ], + [ + [ + -33.316680908203125, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.803985595703125, + false + ], + [ + -15.235649108886719, + false + ], + [ + -29.402576446533203, + false + ], + [ + -33.316680908203125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6be0fdba9c7e5a2761a177a9bd233dce391802d7ff89b3039b893e560c5aefef", + "prompt_hash": "7679b8758e83a28a1b38c63d37f3adaa9c9102f41967a6f095a737d4fc63f07f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 34, + "doc": { + "question": "What proportion of answers in the 200 randomly sampled Omni-MATH problems were classified as 'Text' type?", + "choices": [ + "34", + "51", + "95", + "8" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What proportion of answers in the 200 randomly sampled Omni-MATH problems were classified as 'Text' type?\nAnswer:", + " 34" + ], + [ + "Question:What proportion of answers in the 200 randomly sampled Omni-MATH problems were classified as 'Text' type?\nAnswer:", + " 51" + ], + [ + "Question:What proportion of answers in the 200 randomly sampled Omni-MATH problems were classified as 'Text' type?\nAnswer:", + " 95" + ], + [ + "Question:What proportion of answers in the 200 randomly sampled Omni-MATH problems were classified as 'Text' type?\nAnswer:", + " 8" + ] + ], + "resps": [ + [ + [ + -6.218066692352295, + false + ] + ], + [ + [ + -6.431605815887451, + false + ] + ], + [ + [ + -7.052123546600342, + false + ] + ], + [ + [ + -4.304784774780273, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.218066692352295, + false + ], + [ + -6.431605815887451, + false + ], + [ + -7.052123546600342, + false + ], + [ + -4.304784774780273, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3a86737612a9683d7d370b0913cb5f25f494e298179962b646ce55733319d905", + "prompt_hash": "b8e869651696406e58e09116ea37bc3dc7a38b4745aa4d2aed0fd7d7d97e688d", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 35, + "doc": { + "question": "What is the success rate of leading language models on the FrontierMath benchmark?", + "choices": [ + "A. Over 50%", + "B. Between 10% and 20%", + "C. Between 2% and 5%", + "D. Less than 2%" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the success rate of leading language models on the FrontierMath benchmark?\nAnswer:", + " A. Over 50%" + ], + [ + "Question:What is the success rate of leading language models on the FrontierMath benchmark?\nAnswer:", + " B. Between 10% and 20%" + ], + [ + "Question:What is the success rate of leading language models on the FrontierMath benchmark?\nAnswer:", + " C. Between 2% and 5%" + ], + [ + "Question:What is the success rate of leading language models on the FrontierMath benchmark?\nAnswer:", + " D. Less than 2%" + ] + ], + "resps": [ + [ + [ + -22.314029693603516, + false + ] + ], + [ + [ + -32.98554611206055, + false + ] + ], + [ + [ + -36.0869255065918, + false + ] + ], + [ + [ + -28.037778854370117, + false + ] + ] + ], + "filtered_resps": [ + [ + -22.314029693603516, + false + ], + [ + -32.98554611206055, + false + ], + [ + -36.0869255065918, + false + ], + [ + -28.037778854370117, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fb724d7d2863885f3cf6a1163886e3738f7ac37c4f79ddca68cd0f40d50cb26a", + "prompt_hash": "e9dafc91799916a2de278f5ce5de05e1083033fa4e79b19efd23065a0ddc8416", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 36, + "doc": { + "question": "Which two MSC classifications account for approximately 34% of the FrontierMath dataset?", + "choices": [ + "A. Number theory and group theory", + "B. Combinatorics and algebraic geometry", + "C. Number theory and combinatorics", + "D. Linear algebra and category theory" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which two MSC classifications account for approximately 34% of the FrontierMath dataset?\nAnswer:", + " A. Number theory and group theory" + ], + [ + "Question:Which two MSC classifications account for approximately 34% of the FrontierMath dataset?\nAnswer:", + " B. Combinatorics and algebraic geometry" + ], + [ + "Question:Which two MSC classifications account for approximately 34% of the FrontierMath dataset?\nAnswer:", + " C. Number theory and combinatorics" + ], + [ + "Question:Which two MSC classifications account for approximately 34% of the FrontierMath dataset?\nAnswer:", + " D. Linear algebra and category theory" + ] + ], + "resps": [ + [ + [ + -19.048017501831055, + false + ] + ], + [ + [ + -22.854753494262695, + false + ] + ], + [ + [ + -23.063589096069336, + false + ] + ], + [ + [ + -28.897056579589844, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.048017501831055, + false + ], + [ + -22.854753494262695, + false + ], + [ + -23.063589096069336, + false + ], + [ + -28.897056579589844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "604fd78a586878ff4f413ed6f22f71b442200ec02108500692fc0245ad46b421", + "prompt_hash": "467b4440ea470c77b5f52afebf24eb0321adac3836f625e66bb243d957d956ea", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 37, + "doc": { + "question": "What is the estimated critical error rate of the FrontierMath benchmark based on blind review?", + "choices": [ + "A. 3%", + "B. 6%", + "C. 10%", + "D. 15%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the estimated critical error rate of the FrontierMath benchmark based on blind review?\nAnswer:", + " A. 3%" + ], + [ + "Question:What is the estimated critical error rate of the FrontierMath benchmark based on blind review?\nAnswer:", + " B. 6%" + ], + [ + "Question:What is the estimated critical error rate of the FrontierMath benchmark based on blind review?\nAnswer:", + " C. 10%" + ], + [ + "Question:What is the estimated critical error rate of the FrontierMath benchmark based on blind review?\nAnswer:", + " D. 15%" + ] + ], + "resps": [ + [ + [ + -19.265060424804688, + false + ] + ], + [ + [ + -24.73288345336914, + false + ] + ], + [ + [ + -22.54848289489746, + false + ] + ], + [ + [ + -24.39875602722168, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.265060424804688, + false + ], + [ + -24.73288345336914, + false + ], + [ + -22.54848289489746, + false + ], + [ + -24.39875602722168, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "73232ae25a7ec8e87e70a6a8c15642854c2045fe57f97631f7a1c18f334cccb1", + "prompt_hash": "ffd96735ddf9f5aea6c88664038e563748ffb9d2f1703a9e58da228705766970", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 38, + "doc": { + "question": "How many Fields Medalists were interviewed to assess the difficulty and relevance of FrontierMath?", + "choices": [ + "A. One", + "B. Two", + "C. Three", + "D. Four" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many Fields Medalists were interviewed to assess the difficulty and relevance of FrontierMath?\nAnswer:", + " A. One" + ], + [ + "Question:How many Fields Medalists were interviewed to assess the difficulty and relevance of FrontierMath?\nAnswer:", + " B. Two" + ], + [ + "Question:How many Fields Medalists were interviewed to assess the difficulty and relevance of FrontierMath?\nAnswer:", + " C. Three" + ], + [ + "Question:How many Fields Medalists were interviewed to assess the difficulty and relevance of FrontierMath?\nAnswer:", + " D. Four" + ] + ], + "resps": [ + [ + [ + -12.897353172302246, + false + ] + ], + [ + [ + -15.479686737060547, + false + ] + ], + [ + [ + -15.306407928466797, + false + ] + ], + [ + [ + -16.060991287231445, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.897353172302246, + false + ], + [ + -15.479686737060547, + false + ], + [ + -15.306407928466797, + false + ], + [ + -16.060991287231445, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "92536fa76b86f3416d618123ba7e1ac49f79cf14a5ef14270b202fbb22f4a8f0", + "prompt_hash": "c021c1ebb446b3bda19c667d9c143e5fb5afc75d3c09f41c5d7e24568f777fe6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 39, + "doc": { + "question": "What is the primary reason existing benchmarks like MATH and GSM8K are considered insufficient for evaluating advanced mathematical reasoning in AI?", + "choices": [ + "A. They contain too many open-ended problems.", + "B. They are based on outdated mathematical theory.", + "C. Models have achieved near-perfect scores on them.", + "D. They are limited to symbolic computation only." + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the primary reason existing benchmarks like MATH and GSM8K are considered insufficient for evaluating advanced mathematical reasoning in AI?\nAnswer:", + " A. They contain too many open-ended problems." + ], + [ + "Question:What is the primary reason existing benchmarks like MATH and GSM8K are considered insufficient for evaluating advanced mathematical reasoning in AI?\nAnswer:", + " B. They are based on outdated mathematical theory." + ], + [ + "Question:What is the primary reason existing benchmarks like MATH and GSM8K are considered insufficient for evaluating advanced mathematical reasoning in AI?\nAnswer:", + " C. Models have achieved near-perfect scores on them." + ], + [ + "Question:What is the primary reason existing benchmarks like MATH and GSM8K are considered insufficient for evaluating advanced mathematical reasoning in AI?\nAnswer:", + " D. They are limited to symbolic computation only." + ] + ], + "resps": [ + [ + [ + -28.125661849975586, + false + ] + ], + [ + [ + -32.29228973388672, + false + ] + ], + [ + [ + -39.08148193359375, + false + ] + ], + [ + [ + -35.807430267333984, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.125661849975586, + false + ], + [ + -32.29228973388672, + false + ], + [ + -39.08148193359375, + false + ], + [ + -35.807430267333984, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "47eb164109115e79f1c451f50e7f4fa925728fd1217329b9466740453a1f5756", + "prompt_hash": "b2ecb72db476336ef9065e8ba6707ea0b2787e97183e1e00aeca7136b40e0d26", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 40, + "doc": { + "question": "What does the parameter \\\\( \\\\gamma_h \\\\) represent in the Shannon capacity formula?", + "choices": [ + "Available bandwidth", + "Effective channel gain", + "Signal power", + "Noise power" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What does the parameter \\\\( \\\\gamma_h \\\\) represent in the Shannon capacity formula?\nAnswer:", + " Available bandwidth" + ], + [ + "Question:What does the parameter \\\\( \\\\gamma_h \\\\) represent in the Shannon capacity formula?\nAnswer:", + " Effective channel gain" + ], + [ + "Question:What does the parameter \\\\( \\\\gamma_h \\\\) represent in the Shannon capacity formula?\nAnswer:", + " Signal power" + ], + [ + "Question:What does the parameter \\\\( \\\\gamma_h \\\\) represent in the Shannon capacity formula?\nAnswer:", + " Noise power" + ] + ], + "resps": [ + [ + [ + -14.229288101196289, + false + ] + ], + [ + [ + -15.019903182983398, + false + ] + ], + [ + [ + -11.57690715789795, + false + ] + ], + [ + [ + -11.713316917419434, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.229288101196289, + false + ], + [ + -15.019903182983398, + false + ], + [ + -11.57690715789795, + false + ], + [ + -11.713316917419434, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "caa8bbea2cb52d55ccc287406d790e3015ce631c0834d8f17178d18642335aef", + "prompt_hash": "1265b2989bfbc8a4cfb448e69880f5bd0e58c00570492ec1ff561b9a461ad8f0", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 41, + "doc": { + "question": "Which communication technology allows \\\\( P_n \\\\) to be treated as a configurable system parameter?", + "choices": [ + "NOMA", + "MIMO", + "Noise modulation", + "RIS" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which communication technology allows \\\\( P_n \\\\) to be treated as a configurable system parameter?\nAnswer:", + " NOMA" + ], + [ + "Question:Which communication technology allows \\\\( P_n \\\\) to be treated as a configurable system parameter?\nAnswer:", + " MIMO" + ], + [ + "Question:Which communication technology allows \\\\( P_n \\\\) to be treated as a configurable system parameter?\nAnswer:", + " Noise modulation" + ], + [ + "Question:Which communication technology allows \\\\( P_n \\\\) to be treated as a configurable system parameter?\nAnswer:", + " RIS" + ] + ], + "resps": [ + [ + [ + -16.54871940612793, + false + ] + ], + [ + [ + -12.308393478393555, + false + ] + ], + [ + [ + -18.554058074951172, + false + ] + ], + [ + [ + -12.541747093200684, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.54871940612793, + false + ], + [ + -12.308393478393555, + false + ], + [ + -18.554058074951172, + false + ], + [ + -12.541747093200684, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8794ede48847d991771909da45d875c91fdb156214c8e27d545be83fca7ce78d", + "prompt_hash": "72ca7c5f2ac66b9929b418ce19a700167b6aa456de73d3c075407a8bbb0c7213", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 42, + "doc": { + "question": "What is one unique feature of pinching antennas demonstrated by DOCOMO in 2022?", + "choices": [ + "High-cost antenna reconfiguration", + "Inability to support LoS communication", + "Fixed installation at transceivers", + "Flexible deployment to support strong LoS links" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is one unique feature of pinching antennas demonstrated by DOCOMO in 2022?\nAnswer:", + " High-cost antenna reconfiguration" + ], + [ + "Question:What is one unique feature of pinching antennas demonstrated by DOCOMO in 2022?\nAnswer:", + " Inability to support LoS communication" + ], + [ + "Question:What is one unique feature of pinching antennas demonstrated by DOCOMO in 2022?\nAnswer:", + " Fixed installation at transceivers" + ], + [ + "Question:What is one unique feature of pinching antennas demonstrated by DOCOMO in 2022?\nAnswer:", + " Flexible deployment to support strong LoS links" + ] + ], + "resps": [ + [ + [ + -30.219932556152344, + false + ] + ], + [ + [ + -29.853424072265625, + false + ] + ], + [ + [ + -34.562137603759766, + false + ] + ], + [ + [ + -44.639793395996094, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.219932556152344, + false + ], + [ + -29.853424072265625, + false + ], + [ + -34.562137603759766, + false + ], + [ + -44.639793395996094, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "df2bf0e8ec5836ac8c9b56aa2cdd36f047860e4145da2b160983e3ae8e237236", + "prompt_hash": "688dfac913d0eba440ad3d2edcf0e107af2f03802368c86f07a686245728bb32", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 43, + "doc": { + "question": "In the high SNR regime, what is the approximate ergodic sum rate gain of pinching antennas over conventional antennas dependent on?", + "choices": [ + "The number of waveguides", + "The deployment cost of antennas", + "The ratio \\( \\frac{D}{d} \\)", + "The modulation scheme" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In the high SNR regime, what is the approximate ergodic sum rate gain of pinching antennas over conventional antennas dependent on?\nAnswer:", + " The number of waveguides" + ], + [ + "Question:In the high SNR regime, what is the approximate ergodic sum rate gain of pinching antennas over conventional antennas dependent on?\nAnswer:", + " The deployment cost of antennas" + ], + [ + "Question:In the high SNR regime, what is the approximate ergodic sum rate gain of pinching antennas over conventional antennas dependent on?\nAnswer:", + " The ratio \\( \\frac{D}{d} \\)" + ], + [ + "Question:In the high SNR regime, what is the approximate ergodic sum rate gain of pinching antennas over conventional antennas dependent on?\nAnswer:", + " The modulation scheme" + ] + ], + "resps": [ + [ + [ + -17.938186645507812, + false + ] + ], + [ + [ + -19.054763793945312, + false + ] + ], + [ + [ + -31.16602325439453, + false + ] + ], + [ + [ + -16.148653030395508, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.938186645507812, + false + ], + [ + -19.054763793945312, + false + ], + [ + -31.16602325439453, + false + ], + [ + -16.148653030395508, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "913ffe5ab2c4e9354f9da825d2adc0d08a777f3210d7d06478e42072760a2a7c", + "prompt_hash": "dacddc5de4ac44a8476a44d2839dc73456a28baf1523b2e9f0544fd579e7e568", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 44, + "doc": { + "question": "What condition is required for multiple pinching antennas on a single waveguide to achieve constructive interference?", + "choices": [ + "Using independent RF chains for each antenna", + "Ensuring equal distance from all users", + "Aligning signal phase shifts across antennas", + "Increasing the carrier frequency" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What condition is required for multiple pinching antennas on a single waveguide to achieve constructive interference?\nAnswer:", + " Using independent RF chains for each antenna" + ], + [ + "Question:What condition is required for multiple pinching antennas on a single waveguide to achieve constructive interference?\nAnswer:", + " Ensuring equal distance from all users" + ], + [ + "Question:What condition is required for multiple pinching antennas on a single waveguide to achieve constructive interference?\nAnswer:", + " Aligning signal phase shifts across antennas" + ], + [ + "Question:What condition is required for multiple pinching antennas on a single waveguide to achieve constructive interference?\nAnswer:", + " Increasing the carrier frequency" + ] + ], + "resps": [ + [ + [ + -25.47564697265625, + false + ] + ], + [ + [ + -34.380126953125, + false + ] + ], + [ + [ + -38.81388854980469, + false + ] + ], + [ + [ + -12.850831031799316, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.47564697265625, + false + ], + [ + -34.380126953125, + false + ], + [ + -38.81388854980469, + false + ], + [ + -12.850831031799316, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e1bcf5e1bdbd66506843b97b3911efdb450ec267b9757b4852dd24d46d79dc12", + "prompt_hash": "376a8a5df82e3d345c9f436b865e152193185c84d22742a7bb2709adff2478a4", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 45, + "doc": { + "question": "What frame rate is used by the speech tokenizer in GLM-4-Voice?", + "choices": [ + "A. 50Hz", + "B. 25Hz", + "C. 12.5Hz", + "D. 6.25Hz" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What frame rate is used by the speech tokenizer in GLM-4-Voice?\nAnswer:", + " A. 50Hz" + ], + [ + "Question:What frame rate is used by the speech tokenizer in GLM-4-Voice?\nAnswer:", + " B. 25Hz" + ], + [ + "Question:What frame rate is used by the speech tokenizer in GLM-4-Voice?\nAnswer:", + " C. 12.5Hz" + ], + [ + "Question:What frame rate is used by the speech tokenizer in GLM-4-Voice?\nAnswer:", + " D. 6.25Hz" + ] + ], + "resps": [ + [ + [ + -14.651763916015625, + false + ] + ], + [ + [ + -19.52559471130371, + false + ] + ], + [ + [ + -21.01659393310547, + false + ] + ], + [ + [ + -27.753231048583984, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.651763916015625, + false + ], + [ + -19.52559471130371, + false + ], + [ + -21.01659393310547, + false + ], + [ + -27.753231048583984, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3f47d465aee898fe5f2858c9697a72df4ec05fa08dcd205a7b7dba6b3385f64a", + "prompt_hash": "bbd6919b9e251cf2028e21fa63346428252f3a62bf9c9dd8658015c1f9739415", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 46, + "doc": { + "question": "Which model is used as the initialization base for GLM-4-Voice?", + "choices": [ + "A. Qwen2-7B", + "B. Whisper-Large-V3", + "C. GLM-4-9B-Base", + "D. ChatGLM3-6B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model is used as the initialization base for GLM-4-Voice?\nAnswer:", + " A. Qwen2-7B" + ], + [ + "Question:Which model is used as the initialization base for GLM-4-Voice?\nAnswer:", + " B. Whisper-Large-V3" + ], + [ + "Question:Which model is used as the initialization base for GLM-4-Voice?\nAnswer:", + " C. GLM-4-9B-Base" + ], + [ + "Question:Which model is used as the initialization base for GLM-4-Voice?\nAnswer:", + " D. ChatGLM3-6B" + ] + ], + "resps": [ + [ + [ + -20.858482360839844, + false + ] + ], + [ + [ + -18.8957462310791, + false + ] + ], + [ + [ + -30.00905418395996, + false + ] + ], + [ + [ + -27.548831939697266, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.858482360839844, + false + ], + [ + -18.8957462310791, + false + ], + [ + -30.00905418395996, + false + ], + [ + -27.548831939697266, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a293f6d8dadaf6bf2f7043fc920412bce4bc44eea4364f1f16750dd30ed27e60", + "prompt_hash": "736edfb168814c76d08998796804177f146b07aacb0759253d8bfd2e1aeb3486", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 47, + "doc": { + "question": "Which speech decoder architecture is used in GLM-4-Voice?", + "choices": [ + "A. FastSpeech 2", + "B. CosyVoice", + "C. Tacotron 2", + "D. VALL-E" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which speech decoder architecture is used in GLM-4-Voice?\nAnswer:", + " A. FastSpeech 2" + ], + [ + "Question:Which speech decoder architecture is used in GLM-4-Voice?\nAnswer:", + " B. CosyVoice" + ], + [ + "Question:Which speech decoder architecture is used in GLM-4-Voice?\nAnswer:", + " C. Tacotron 2" + ], + [ + "Question:Which speech decoder architecture is used in GLM-4-Voice?\nAnswer:", + " D. VALL-E" + ] + ], + "resps": [ + [ + [ + -15.32555866241455, + false + ] + ], + [ + [ + -22.239765167236328, + false + ] + ], + [ + [ + -18.153594970703125, + false + ] + ], + [ + [ + -21.598918914794922, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.32555866241455, + false + ], + [ + -22.239765167236328, + false + ], + [ + -18.153594970703125, + false + ], + [ + -21.598918914794922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "850102a872cc87cacfad06d4e0d5cb4794eba6950191dc24f32f5ba7d381e3ab", + "prompt_hash": "a9fc375bc89c3d93c91a932a96b0528b1f5b93895c18af2d3aa297b0eb2ff84f", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 48, + "doc": { + "question": "What is the total number of tokens used during pre-training of GLM-4-Voice?", + "choices": [ + "A. 700 million", + "B. 7 billion", + "C. 1 trillion", + "D. 12 billion" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the total number of tokens used during pre-training of GLM-4-Voice?\nAnswer:", + " A. 700 million" + ], + [ + "Question:What is the total number of tokens used during pre-training of GLM-4-Voice?\nAnswer:", + " B. 7 billion" + ], + [ + "Question:What is the total number of tokens used during pre-training of GLM-4-Voice?\nAnswer:", + " C. 1 trillion" + ], + [ + "Question:What is the total number of tokens used during pre-training of GLM-4-Voice?\nAnswer:", + " D. 12 billion" + ] + ], + "resps": [ + [ + [ + -16.722270965576172, + false + ] + ], + [ + [ + -20.472637176513672, + false + ] + ], + [ + [ + -22.23086929321289, + false + ] + ], + [ + [ + -22.3320369720459, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.722270965576172, + false + ], + [ + -20.472637176513672, + false + ], + [ + -22.23086929321289, + false + ], + [ + -22.3320369720459, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "74e876c3453e4c4b42efbcfb18fb317a6029dbe0e340688a0d15ebb475e1c0c0", + "prompt_hash": "a7766b046084beb4895caed4bb212405c27f90baeae0cb0f1c8d744b3a4902e5", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 49, + "doc": { + "question": "Which model provides the text predictions for English recognition in the TTS task?", + "choices": [ + "A. Whisper-Large-V3", + "B. Paraformer-Large", + "C. ChatGPT", + "D. CosyVoice" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which model provides the text predictions for English recognition in the TTS task?\nAnswer:", + " A. Whisper-Large-V3" + ], + [ + "Question:Which model provides the text predictions for English recognition in the TTS task?\nAnswer:", + " B. Paraformer-Large" + ], + [ + "Question:Which model provides the text predictions for English recognition in the TTS task?\nAnswer:", + " C. ChatGPT" + ], + [ + "Question:Which model provides the text predictions for English recognition in the TTS task?\nAnswer:", + " D. CosyVoice" + ] + ], + "resps": [ + [ + [ + -15.751648902893066, + false + ] + ], + [ + [ + -26.65165138244629, + false + ] + ], + [ + [ + -15.32774543762207, + false + ] + ], + [ + [ + -22.42506980895996, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.751648902893066, + false + ], + [ + -26.65165138244629, + false + ], + [ + -15.32774543762207, + false + ], + [ + -22.42506980895996, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "70b1b8c729dc4de68d708dae2634d94cd7992df78b1a011c91085f51b37f64b9", + "prompt_hash": "559397a2d29372587b9ffb95e5efbe13ca23fa3ed596ae5a96b78a1d3780be3d", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 50, + "doc": { + "question": "What is the name of the newly introduced dataset that uses randomly generated DAGs for logical reasoning tasks?", + "choices": [ + "TreeQA", + "GraphProofQA", + "ProsQA", + "LogicDAG" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the name of the newly introduced dataset that uses randomly generated DAGs for logical reasoning tasks?\nAnswer:", + " TreeQA" + ], + [ + "Question:What is the name of the newly introduced dataset that uses randomly generated DAGs for logical reasoning tasks?\nAnswer:", + " GraphProofQA" + ], + [ + "Question:What is the name of the newly introduced dataset that uses randomly generated DAGs for logical reasoning tasks?\nAnswer:", + " ProsQA" + ], + [ + "Question:What is the name of the newly introduced dataset that uses randomly generated DAGs for logical reasoning tasks?\nAnswer:", + " LogicDAG" + ] + ], + "resps": [ + [ + [ + -11.236031532287598, + false + ] + ], + [ + [ + -19.69611358642578, + false + ] + ], + [ + [ + -20.450382232666016, + false + ] + ], + [ + [ + -15.820966720581055, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.236031532287598, + false + ], + [ + -19.69611358642578, + false + ], + [ + -20.450382232666016, + false + ], + [ + -15.820966720581055, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "217edf1170ad8ce8a10962ef9838cb1c147f69453ff3b88e325419607a5c912f", + "prompt_hash": "f1f4f13fd8429cb29c05b70cdf386e00ca3b54085197be705ad3102830e66f08", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 51, + "doc": { + "question": "Which pre-trained model is used as the base model for all experiments?", + "choices": [ + "GPT-2", + "LLaMA-2", + "GPT-Neo", + "T5" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which pre-trained model is used as the base model for all experiments?\nAnswer:", + " GPT-2" + ], + [ + "Question:Which pre-trained model is used as the base model for all experiments?\nAnswer:", + " LLaMA-2" + ], + [ + "Question:Which pre-trained model is used as the base model for all experiments?\nAnswer:", + " GPT-Neo" + ], + [ + "Question:Which pre-trained model is used as the base model for all experiments?\nAnswer:", + " T5" + ] + ], + "resps": [ + [ + [ + -7.498136520385742, + false + ] + ], + [ + [ + -9.902589797973633, + false + ] + ], + [ + [ + -9.747665405273438, + false + ] + ], + [ + [ + -6.549315452575684, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.498136520385742, + false + ], + [ + -9.902589797973633, + false + ], + [ + -9.747665405273438, + false + ], + [ + -6.549315452575684, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "717cadaa616a2bdcb3b4bf53a886de4b82ea0600a2c355ba95d9d1c16c134cd3", + "prompt_hash": "4b95d935fa7bcd67d739d9d8436016f42a8f32e5e23471f1ae1a0285f58d13fa", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 52, + "doc": { + "question": "Which token is used to mark the end of the latent thought mode in the proposed method?", + "choices": [ + "", + "", + "", + "" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which token is used to mark the end of the latent thought mode in the proposed method?\nAnswer:", + " " + ], + [ + "Question:Which token is used to mark the end of the latent thought mode in the proposed method?\nAnswer:", + " " + ], + [ + "Question:Which token is used to mark the end of the latent thought mode in the proposed method?\nAnswer:", + " " + ], + [ + "Question:Which token is used to mark the end of the latent thought mode in the proposed method?\nAnswer:", + " " + ] + ], + "resps": [ + [ + [ + -13.798894882202148, + false + ] + ], + [ + [ + -16.525150299072266, + false + ] + ], + [ + [ + -18.213878631591797, + false + ] + ], + [ + [ + -21.657358169555664, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.798894882202148, + false + ], + [ + -16.525150299072266, + false + ], + [ + -18.213878631591797, + false + ], + [ + -21.657358169555664, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "cf0f9f5525037bbc013d7bb008d93bb912e491d8bfc8a8a3066055cd421ac454", + "prompt_hash": "48b90eea372a285d08bd038c7dfcb3a0d2529d59ea81e7e38d313a5a14106a8e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 53, + "doc": { + "question": "What is the default value of the hyperparameter 'c' used for math reasoning tasks in this study?", + "choices": [ + "1", + "2", + "3", + "4" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the default value of the hyperparameter 'c' used for math reasoning tasks in this study?\nAnswer:", + " 1" + ], + [ + "Question:What is the default value of the hyperparameter 'c' used for math reasoning tasks in this study?\nAnswer:", + " 2" + ], + [ + "Question:What is the default value of the hyperparameter 'c' used for math reasoning tasks in this study?\nAnswer:", + " 3" + ], + [ + "Question:What is the default value of the hyperparameter 'c' used for math reasoning tasks in this study?\nAnswer:", + " 4" + ] + ], + "resps": [ + [ + [ + -3.3783435821533203, + false + ] + ], + [ + [ + -3.9628448486328125, + false + ] + ], + [ + [ + -4.146780014038086, + false + ] + ], + [ + [ + -4.759138107299805, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.3783435821533203, + false + ], + [ + -3.9628448486328125, + false + ], + [ + -4.146780014038086, + false + ], + [ + -4.759138107299805, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ca09e725c397a1c7347e96ad136378e5dd2d335a63df0a84e35597299b2a3ca1", + "prompt_hash": "174c8e3c250a846c3b26420ac08a5f026b3b9b41c8763c39a85758bd0a1c3f99", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 54, + "doc": { + "question": "Which baseline method removes tokens at the beginning of the reasoning chain during training and predicts the answer directly during inference?", + "choices": [ + "No-CoT", + "Pause Token", + "iCoT", + "w/o curriculum" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which baseline method removes tokens at the beginning of the reasoning chain during training and predicts the answer directly during inference?\nAnswer:", + " No-CoT" + ], + [ + "Question:Which baseline method removes tokens at the beginning of the reasoning chain during training and predicts the answer directly during inference?\nAnswer:", + " Pause Token" + ], + [ + "Question:Which baseline method removes tokens at the beginning of the reasoning chain during training and predicts the answer directly during inference?\nAnswer:", + " iCoT" + ], + [ + "Question:Which baseline method removes tokens at the beginning of the reasoning chain during training and predicts the answer directly during inference?\nAnswer:", + " w/o curriculum" + ] + ], + "resps": [ + [ + [ + -16.552953720092773, + false + ] + ], + [ + [ + -15.976731300354004, + false + ] + ], + [ + [ + -17.974000930786133, + false + ] + ], + [ + [ + -22.248485565185547, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.552953720092773, + false + ], + [ + -15.976731300354004, + false + ], + [ + -17.974000930786133, + false + ], + [ + -22.248485565185547, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "19c4a79c779e1d991b74627e99dab9307bd5fb398d814d3660d24947b9d54611", + "prompt_hash": "5dee7cee91640ff580164a7cdeffb2da7d4cdd23e33467b9ca6a99a056c7090c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 55, + "doc": { + "question": "Which base model is used for STILL-2?", + "choices": [ + "LLaMA-3-70B", + "Qwen2.5-32B-Instruct", + "DeepSeek-R1", + "Claude 3.5 Sonnet" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which base model is used for STILL-2?\nAnswer:", + " LLaMA-3-70B" + ], + [ + "Question:Which base model is used for STILL-2?\nAnswer:", + " Qwen2.5-32B-Instruct" + ], + [ + "Question:Which base model is used for STILL-2?\nAnswer:", + " DeepSeek-R1" + ], + [ + "Question:Which base model is used for STILL-2?\nAnswer:", + " Claude 3.5 Sonnet" + ] + ], + "resps": [ + [ + [ + -11.793059349060059, + false + ] + ], + [ + [ + -26.26325798034668, + false + ] + ], + [ + [ + -20.53174591064453, + false + ] + ], + [ + [ + -9.624313354492188, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.793059349060059, + false + ], + [ + -26.26325798034668, + false + ], + [ + -20.53174591064453, + false + ], + [ + -9.624313354492188, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c2b849793784a9505396f5301f1eeb76cc79db06232dd82424a84b5740afb1b1", + "prompt_hash": "302486dd58a15c3fe94f662eaa51f2c5361e954848f56c235b63a56b30c3e011", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 56, + "doc": { + "question": "Which dataset consists of 198 multiple-choice questions from biology, physics, and chemistry?", + "choices": [ + "MATH-OAI", + "AIME2024", + "GPQA", + "LeetCode-Hard" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which dataset consists of 198 multiple-choice questions from biology, physics, and chemistry?\nAnswer:", + " MATH-OAI" + ], + [ + "Question:Which dataset consists of 198 multiple-choice questions from biology, physics, and chemistry?\nAnswer:", + " AIME2024" + ], + [ + "Question:Which dataset consists of 198 multiple-choice questions from biology, physics, and chemistry?\nAnswer:", + " GPQA" + ], + [ + "Question:Which dataset consists of 198 multiple-choice questions from biology, physics, and chemistry?\nAnswer:", + " LeetCode-Hard" + ] + ], + "resps": [ + [ + [ + -28.460281372070312, + false + ] + ], + [ + [ + -24.10089683532715, + false + ] + ], + [ + [ + -13.763631820678711, + false + ] + ], + [ + [ + -31.144329071044922, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.460281372070312, + false + ], + [ + -24.10089683532715, + false + ], + [ + -13.763631820678711, + false + ], + [ + -31.144329071044922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "271cf3dbb0d16a2e5d6d5fa731d7bb108df4fb9b052fc9423f3a85759d29312b", + "prompt_hash": "83f2afe83c85e40c1582a64dd22ef90a1058c1da9529598f225e3d203b4bd707", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 57, + "doc": { + "question": "Which training approach in STILL-2 includes aligning only the thought process and incorporating SFT for stability?", + "choices": [ + "Chain-of-thought prompting", + "Supervised fine-tuning only", + "Direct Preference Optimization", + "Tree search with reward model" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which training approach in STILL-2 includes aligning only the thought process and incorporating SFT for stability?\nAnswer:", + " Chain-of-thought prompting" + ], + [ + "Question:Which training approach in STILL-2 includes aligning only the thought process and incorporating SFT for stability?\nAnswer:", + " Supervised fine-tuning only" + ], + [ + "Question:Which training approach in STILL-2 includes aligning only the thought process and incorporating SFT for stability?\nAnswer:", + " Direct Preference Optimization" + ], + [ + "Question:Which training approach in STILL-2 includes aligning only the thought process and incorporating SFT for stability?\nAnswer:", + " Tree search with reward model" + ] + ], + "resps": [ + [ + [ + -13.133655548095703, + false + ] + ], + [ + [ + -15.158451080322266, + false + ] + ], + [ + [ + -8.369895935058594, + false + ] + ], + [ + [ + -27.84103775024414, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.133655548095703, + false + ], + [ + -15.158451080322266, + false + ], + [ + -8.369895935058594, + false + ], + [ + -27.84103775024414, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f94afbbe7b52724b8214b29114f6b669a4a89d5bbae545496a00f2acdde9df6d", + "prompt_hash": "1d4c18ff649f9f9360faee6f55c223cbb9e600f28419bcf3be5485a504f4a566", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 58, + "doc": { + "question": "Which problem domain is prioritized in the data mixture due to its prevalence of challenging reasoning problems?", + "choices": [ + "Coding", + "Puzzle", + "Mathematics", + "Science" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which problem domain is prioritized in the data mixture due to its prevalence of challenging reasoning problems?\nAnswer:", + " Coding" + ], + [ + "Question:Which problem domain is prioritized in the data mixture due to its prevalence of challenging reasoning problems?\nAnswer:", + " Puzzle" + ], + [ + "Question:Which problem domain is prioritized in the data mixture due to its prevalence of challenging reasoning problems?\nAnswer:", + " Mathematics" + ], + [ + "Question:Which problem domain is prioritized in the data mixture due to its prevalence of challenging reasoning problems?\nAnswer:", + " Science" + ] + ], + "resps": [ + [ + [ + -6.636571884155273, + false + ] + ], + [ + [ + -10.519530296325684, + false + ] + ], + [ + [ + -6.341315269470215, + false + ] + ], + [ + [ + -8.144645690917969, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.636571884155273, + false + ], + [ + -10.519530296325684, + false + ], + [ + -6.341315269470215, + false + ], + [ + -8.144645690917969, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1804ef6e48eb4018a1b8cd33d9dc8a777fddc068fabcf07a1fd4c5a7abb51113", + "prompt_hash": "c673cd913e516594b6879b3185f72c468d1657f0a76609fafed1044316aa1a3a", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 59, + "doc": { + "question": "What type of tasks does AIME2024 aim to challenge students with?", + "choices": [ + "Basic arithmetic and algebra", + "Simple geometry questions", + "Complex problem solving tasks", + "Physics-based numerical estimation" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What type of tasks does AIME2024 aim to challenge students with?\nAnswer:", + " Basic arithmetic and algebra" + ], + [ + "Question:What type of tasks does AIME2024 aim to challenge students with?\nAnswer:", + " Simple geometry questions" + ], + [ + "Question:What type of tasks does AIME2024 aim to challenge students with?\nAnswer:", + " Complex problem solving tasks" + ], + [ + "Question:What type of tasks does AIME2024 aim to challenge students with?\nAnswer:", + " Physics-based numerical estimation" + ] + ], + "resps": [ + [ + [ + -16.77252960205078, + false + ] + ], + [ + [ + -22.267976760864258, + false + ] + ], + [ + [ + -15.13084888458252, + false + ] + ], + [ + [ + -25.581417083740234, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.77252960205078, + false + ], + [ + -22.267976760864258, + false + ], + [ + -15.13084888458252, + false + ], + [ + -25.581417083740234, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "422af2770208747c4876585172d73a22ab5ff58500e254276602c33f94a7b6d5", + "prompt_hash": "05bc871e2ef2fc7fafe36650b6c5087172d2af1beedb7446f8da9686ecf8cdf2", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 60, + "doc": { + "question": "What type of quantization module does CosyVoice 2 insert into the encoder of the SenseVoice-Large ASR model?", + "choices": [ + "Vector Quantization (VQ)", + "Scalar Product Quantization (SPQ)", + "Finite Scalar Quantization (FSQ)", + "Dynamic Quantization (DQ)" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What type of quantization module does CosyVoice 2 insert into the encoder of the SenseVoice-Large ASR model?\nAnswer:", + " Vector Quantization (VQ)" + ], + [ + "Question:What type of quantization module does CosyVoice 2 insert into the encoder of the SenseVoice-Large ASR model?\nAnswer:", + " Scalar Product Quantization (SPQ)" + ], + [ + "Question:What type of quantization module does CosyVoice 2 insert into the encoder of the SenseVoice-Large ASR model?\nAnswer:", + " Finite Scalar Quantization (FSQ)" + ], + [ + "Question:What type of quantization module does CosyVoice 2 insert into the encoder of the SenseVoice-Large ASR model?\nAnswer:", + " Dynamic Quantization (DQ)" + ] + ], + "resps": [ + [ + [ + -11.393033981323242, + false + ] + ], + [ + [ + -20.69158172607422, + false + ] + ], + [ + [ + -21.93514633178711, + false + ] + ], + [ + [ + -11.593354225158691, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.393033981323242, + false + ], + [ + -20.69158172607422, + false + ], + [ + -21.93514633178711, + false + ], + [ + -11.593354225158691, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "32388d6a9407a2fcfc1e9858f4d73cc33a35b7225d109348bbc29bbee35de03d", + "prompt_hash": "32b3cf4085d62ea87f7bdc218fb3adf337a472f25c450abaaab4d2cbacbe42c6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 61, + "doc": { + "question": "What is the token rate of CosyVoice 2's speech tokenizer?", + "choices": [ + "50 Hz", + "25 Hz", + "16 Hz", + "100 Hz" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the token rate of CosyVoice 2's speech tokenizer?\nAnswer:", + " 50 Hz" + ], + [ + "Question:What is the token rate of CosyVoice 2's speech tokenizer?\nAnswer:", + " 25 Hz" + ], + [ + "Question:What is the token rate of CosyVoice 2's speech tokenizer?\nAnswer:", + " 16 Hz" + ], + [ + "Question:What is the token rate of CosyVoice 2's speech tokenizer?\nAnswer:", + " 100 Hz" + ] + ], + "resps": [ + [ + [ + -9.393841743469238, + false + ] + ], + [ + [ + -9.112127304077148, + false + ] + ], + [ + [ + -11.231280326843262, + false + ] + ], + [ + [ + -9.16458511352539, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.393841743469238, + false + ], + [ + -9.112127304077148, + false + ], + [ + -11.231280326843262, + false + ], + [ + -9.16458511352539, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6bd69fa3a8b3b05be09aef518a19ec78c4b03b5fabb4e150d4336d1b9adbeaf4", + "prompt_hash": "7280d11880ab1693eb3cf76da610ec6c43fdfbc75219ef6f7aab4661a959250d", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 62, + "doc": { + "question": "Which pre-trained language model is used in CosyVoice 2 as the text-speech language model?", + "choices": [ + "ChatGLM3", + "Qwen1.5-1.8B", + "Qwen2.5-0.5B", + "GPT-NeoX-20B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which pre-trained language model is used in CosyVoice 2 as the text-speech language model?\nAnswer:", + " ChatGLM3" + ], + [ + "Question:Which pre-trained language model is used in CosyVoice 2 as the text-speech language model?\nAnswer:", + " Qwen1.5-1.8B" + ], + [ + "Question:Which pre-trained language model is used in CosyVoice 2 as the text-speech language model?\nAnswer:", + " Qwen2.5-0.5B" + ], + [ + "Question:Which pre-trained language model is used in CosyVoice 2 as the text-speech language model?\nAnswer:", + " GPT-NeoX-20B" + ] + ], + "resps": [ + [ + [ + -7.910797595977783, + false + ] + ], + [ + [ + -13.443157196044922, + false + ] + ], + [ + [ + -28.324539184570312, + false + ] + ], + [ + [ + -19.48931121826172, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.910797595977783, + false + ], + [ + -13.443157196044922, + false + ], + [ + -28.324539184570312, + false + ], + [ + -19.48931121826172, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "59de127e082d56bb6a752e5a2a7289961596af15dbf9353afd2237d9808dcc5d", + "prompt_hash": "41dd334a9a9ce470c060aff357efbcd2b2ba5e2c6095c1a9c4c3eed65a8f3ed2", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 63, + "doc": { + "question": "What chunk sizes (N:M) are used in CosyVoice 2’s streaming mode for mixing text and speech tokens?", + "choices": [ + "3:9", + "4:10", + "5:15", + "6:18" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What chunk sizes (N:M) are used in CosyVoice 2’s streaming mode for mixing text and speech tokens?\nAnswer:", + " 3:9" + ], + [ + "Question:What chunk sizes (N:M) are used in CosyVoice 2’s streaming mode for mixing text and speech tokens?\nAnswer:", + " 4:10" + ], + [ + "Question:What chunk sizes (N:M) are used in CosyVoice 2’s streaming mode for mixing text and speech tokens?\nAnswer:", + " 5:15" + ], + [ + "Question:What chunk sizes (N:M) are used in CosyVoice 2’s streaming mode for mixing text and speech tokens?\nAnswer:", + " 6:18" + ] + ], + "resps": [ + [ + [ + -13.878074645996094, + false + ] + ], + [ + [ + -9.731094360351562, + false + ] + ], + [ + [ + -12.27220344543457, + false + ] + ], + [ + [ + -14.032604217529297, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.878074645996094, + false + ], + [ + -9.731094360351562, + false + ], + [ + -12.27220344543457, + false + ], + [ + -14.032604217529297, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e0e5ce7d1693f556c0ab935f6a2e22c31119e91529cd638537632814431bf87a", + "prompt_hash": "cd61ef147018a2a8b6e7ba44d85766b1813c0bbc1d26a6e5566dbb64ec287bbe", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 64, + "doc": { + "question": "What mask type in CosyVoice 2 is designed for extremely low latency by attending only past frames?", + "choices": [ + "Chunk-2M Mask", + "Non-causal Mask", + "Full-causal Mask", + "Streaming Mask" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What mask type in CosyVoice 2 is designed for extremely low latency by attending only past frames?\nAnswer:", + " Chunk-2M Mask" + ], + [ + "Question:What mask type in CosyVoice 2 is designed for extremely low latency by attending only past frames?\nAnswer:", + " Non-causal Mask" + ], + [ + "Question:What mask type in CosyVoice 2 is designed for extremely low latency by attending only past frames?\nAnswer:", + " Full-causal Mask" + ], + [ + "Question:What mask type in CosyVoice 2 is designed for extremely low latency by attending only past frames?\nAnswer:", + " Streaming Mask" + ] + ], + "resps": [ + [ + [ + -28.615375518798828, + false + ] + ], + [ + [ + -15.082904815673828, + false + ] + ], + [ + [ + -25.027320861816406, + false + ] + ], + [ + [ + -12.463473320007324, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.615375518798828, + false + ], + [ + -15.082904815673828, + false + ], + [ + -25.027320861816406, + false + ], + [ + -12.463473320007324, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "18b2cd20433a68dc5171a40403d0a99ab28e22454bd72874e3985c717b9d38a8", + "prompt_hash": "4749cce3e3a50d1979d5653490c00dbc8f21b3f348996d5f08a68f1e9d2b5f07", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 65, + "doc": { + "question": "What is the base resolution used by the SigLIP-SO400M-384 vision encoder in DeepSeek-VL2?", + "choices": [ + "256 × 256", + "384 × 384", + "512 × 512", + "1024 × 1024" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the base resolution used by the SigLIP-SO400M-384 vision encoder in DeepSeek-VL2?\nAnswer:", + " 256 × 256" + ], + [ + "Question:What is the base resolution used by the SigLIP-SO400M-384 vision encoder in DeepSeek-VL2?\nAnswer:", + " 384 × 384" + ], + [ + "Question:What is the base resolution used by the SigLIP-SO400M-384 vision encoder in DeepSeek-VL2?\nAnswer:", + " 512 × 512" + ], + [ + "Question:What is the base resolution used by the SigLIP-SO400M-384 vision encoder in DeepSeek-VL2?\nAnswer:", + " 1024 × 1024" + ] + ], + "resps": [ + [ + [ + -8.130940437316895, + false + ] + ], + [ + [ + -7.248654365539551, + false + ] + ], + [ + [ + -8.044690132141113, + false + ] + ], + [ + [ + -8.730853080749512, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.130940437316895, + false + ], + [ + -7.248654365539551, + false + ], + [ + -8.044690132141113, + false + ], + [ + -8.730853080749512, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a92a91bc1c33a56adb94738fafdca24088730e185939f28ca1970ae3277e1cf3", + "prompt_hash": "eb055f1388f8ac10285b290c69173e7c821461cfee313aacde2b270f3a7a20dd", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 66, + "doc": { + "question": "Which mechanism in DeepSeek-VL2 reduces inference computational cost by compressing the Key-Value cache?", + "choices": [ + "Multi-query Attention", + "Low-rank Decomposition", + "Multi-head Latent Attention", + "Sparse Transformer Blocks" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which mechanism in DeepSeek-VL2 reduces inference computational cost by compressing the Key-Value cache?\nAnswer:", + " Multi-query Attention" + ], + [ + "Question:Which mechanism in DeepSeek-VL2 reduces inference computational cost by compressing the Key-Value cache?\nAnswer:", + " Low-rank Decomposition" + ], + [ + "Question:Which mechanism in DeepSeek-VL2 reduces inference computational cost by compressing the Key-Value cache?\nAnswer:", + " Multi-head Latent Attention" + ], + [ + "Question:Which mechanism in DeepSeek-VL2 reduces inference computational cost by compressing the Key-Value cache?\nAnswer:", + " Sparse Transformer Blocks" + ] + ], + "resps": [ + [ + [ + -16.46713638305664, + false + ] + ], + [ + [ + -14.968793869018555, + false + ] + ], + [ + [ + -24.092559814453125, + false + ] + ], + [ + [ + -18.959882736206055, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.46713638305664, + false + ], + [ + -14.968793869018555, + false + ], + [ + -24.092559814453125, + false + ], + [ + -18.959882736206055, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "20c9fc230fca349220bc87a8955d8e099f027dbd565e0122628db50fdbfae1c8", + "prompt_hash": "565efc2f8caf679698d0c9ea9eedaa81a8f5ac5ff6aeade1e2212c25f6cdcabd", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 67, + "doc": { + "question": "Which resolution constraint limited the original DeepSeek-VL in handling high-resolution images?", + "choices": [ + "512 × 512", + "256 × 256", + "1024 × 1024", + "2048 × 2048" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which resolution constraint limited the original DeepSeek-VL in handling high-resolution images?\nAnswer:", + " 512 × 512" + ], + [ + "Question:Which resolution constraint limited the original DeepSeek-VL in handling high-resolution images?\nAnswer:", + " 256 × 256" + ], + [ + "Question:Which resolution constraint limited the original DeepSeek-VL in handling high-resolution images?\nAnswer:", + " 1024 × 1024" + ], + [ + "Question:Which resolution constraint limited the original DeepSeek-VL in handling high-resolution images?\nAnswer:", + " 2048 × 2048" + ] + ], + "resps": [ + [ + [ + -10.976244926452637, + false + ] + ], + [ + [ + -10.028334617614746, + false + ] + ], + [ + [ + -11.025457382202148, + false + ] + ], + [ + [ + -12.071310997009277, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.976244926452637, + false + ], + [ + -10.028334617614746, + false + ], + [ + -11.025457382202148, + false + ], + [ + -12.071310997009277, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "339d792bbc3b019970a774aa67ab7b5ca6ce8bd4b5eb6fad1b35f879e699edff", + "prompt_hash": "8d2317f726c17f3e3b8939e1a9ec7485286c12b4f26288b2b1be18212ba0aa70", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 68, + "doc": { + "question": "Which dataset was used in the vision-language alignment stage of DeepSeek-VL2 training?", + "choices": [ + "WIT", + "OBELICS", + "ShareGPT4V", + "WikiHow" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which dataset was used in the vision-language alignment stage of DeepSeek-VL2 training?\nAnswer:", + " WIT" + ], + [ + "Question:Which dataset was used in the vision-language alignment stage of DeepSeek-VL2 training?\nAnswer:", + " OBELICS" + ], + [ + "Question:Which dataset was used in the vision-language alignment stage of DeepSeek-VL2 training?\nAnswer:", + " ShareGPT4V" + ], + [ + "Question:Which dataset was used in the vision-language alignment stage of DeepSeek-VL2 training?\nAnswer:", + " WikiHow" + ] + ], + "resps": [ + [ + [ + -11.62100601196289, + false + ] + ], + [ + [ + -19.688602447509766, + false + ] + ], + [ + [ + -13.465234756469727, + false + ] + ], + [ + [ + -13.078892707824707, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.62100601196289, + false + ], + [ + -19.688602447509766, + false + ], + [ + -13.465234756469727, + false + ], + [ + -13.078892707824707, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "99d67a96b52fca73877197e256f8e112981eda8d491873d62f896bd8e54fb9fc", + "prompt_hash": "c65656c4aea83a2ce79f67dc2e6cd18335eed3f311b95e24a22a6e3d99dff98d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 69, + "doc": { + "question": "What vocabulary size is used in both DeepSeek-VL2-Tiny and DeepSeek-VL2 models?", + "choices": [ + "102,400", + "129,280", + "96,000", + "88,512" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What vocabulary size is used in both DeepSeek-VL2-Tiny and DeepSeek-VL2 models?\nAnswer:", + " 102,400" + ], + [ + "Question:What vocabulary size is used in both DeepSeek-VL2-Tiny and DeepSeek-VL2 models?\nAnswer:", + " 129,280" + ], + [ + "Question:What vocabulary size is used in both DeepSeek-VL2-Tiny and DeepSeek-VL2 models?\nAnswer:", + " 96,000" + ], + [ + "Question:What vocabulary size is used in both DeepSeek-VL2-Tiny and DeepSeek-VL2 models?\nAnswer:", + " 88,512" + ] + ], + "resps": [ + [ + [ + -12.950544357299805, + false + ] + ], + [ + [ + -20.23044204711914, + false + ] + ], + [ + [ + -12.914684295654297, + false + ] + ], + [ + [ + -19.146526336669922, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.950544357299805, + false + ], + [ + -20.23044204711914, + false + ], + [ + -12.914684295654297, + false + ], + [ + -19.146526336669922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0f05867b00d406bf5e9952c229cbd0f2bf1767270f14ad1556db51142a068e5e", + "prompt_hash": "8d602975dad9718428ec256f9279453382b67fb55fc0159e498c69f4298b04bf", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 70, + "doc": { + "question": "What is the parameter count of the Phi-4 model?", + "choices": [ + "7 billion", + "14 billion", + "70 billion", + "32 billion" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the parameter count of the Phi-4 model?\nAnswer:", + " 7 billion" + ], + [ + "Question:What is the parameter count of the Phi-4 model?\nAnswer:", + " 14 billion" + ], + [ + "Question:What is the parameter count of the Phi-4 model?\nAnswer:", + " 70 billion" + ], + [ + "Question:What is the parameter count of the Phi-4 model?\nAnswer:", + " 32 billion" + ] + ], + "resps": [ + [ + [ + -8.51523208618164, + false + ] + ], + [ + [ + -10.523784637451172, + false + ] + ], + [ + [ + -10.680200576782227, + false + ] + ], + [ + [ + -10.754088401794434, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.51523208618164, + false + ], + [ + -10.523784637451172, + false + ], + [ + -10.680200576782227, + false + ], + [ + -10.754088401794434, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "116f467649a52acc054bc1f84838646ea02114520a9156a41b9605a49738f1d0", + "prompt_hash": "be084df62358405412baac8c81d863a2a0b37a8d5ce5019beddd342c0af16f65", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 71, + "doc": { + "question": "Which benchmark does Phi-4 outperform GPT-4o on?", + "choices": [ + "HumanEval+", + "DROP", + "GPQA", + "MMLU" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which benchmark does Phi-4 outperform GPT-4o on?\nAnswer:", + " HumanEval+" + ], + [ + "Question:Which benchmark does Phi-4 outperform GPT-4o on?\nAnswer:", + " DROP" + ], + [ + "Question:Which benchmark does Phi-4 outperform GPT-4o on?\nAnswer:", + " GPQA" + ], + [ + "Question:Which benchmark does Phi-4 outperform GPT-4o on?\nAnswer:", + " MMLU" + ] + ], + "resps": [ + [ + [ + -13.359102249145508, + false + ] + ], + [ + [ + -13.945856094360352, + false + ] + ], + [ + [ + -14.53800106048584, + false + ] + ], + [ + [ + -8.430791854858398, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.359102249145508, + false + ], + [ + -13.945856094360352, + false + ], + [ + -14.53800106048584, + false + ], + [ + -8.430791854858398, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f2865fffe2519020fe97fce33fc3d4425d49db303f5659dfa1d85c4b590e04a8", + "prompt_hash": "8e29f0852f692ccde8ae0c99c89c51add71e311dcad1bace9a4821757f310040", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 72, + "doc": { + "question": "What percentage of Phi-4’s pretraining data is composed of synthetic data?", + "choices": [ + "20%", + "30%", + "40%", + "50%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What percentage of Phi-4’s pretraining data is composed of synthetic data?\nAnswer:", + " 20%" + ], + [ + "Question:What percentage of Phi-4’s pretraining data is composed of synthetic data?\nAnswer:", + " 30%" + ], + [ + "Question:What percentage of Phi-4’s pretraining data is composed of synthetic data?\nAnswer:", + " 40%" + ], + [ + "Question:What percentage of Phi-4’s pretraining data is composed of synthetic data?\nAnswer:", + " 50%" + ] + ], + "resps": [ + [ + [ + -9.83897590637207, + false + ] + ], + [ + [ + -10.036050796508789, + false + ] + ], + [ + [ + -9.986572265625, + false + ] + ], + [ + [ + -9.598316192626953, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.83897590637207, + false + ], + [ + -10.036050796508789, + false + ], + [ + -9.986572265625, + false + ], + [ + -9.598316192626953, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "74f028a642a4ea81b40a1d46b7967c490245ef183efc0afee04acb9a4fae3253", + "prompt_hash": "5653b6673de59f32813aaf6883f72dc84b2d3055b444083e8650a07442f1c14d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 73, + "doc": { + "question": "Which technique is introduced to generate DPO pairs in the first stage of post-training?", + "choices": [ + "Multi-agent prompting", + "Instruction reversal", + "Pivotal Token Search", + "Majority Voting" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which technique is introduced to generate DPO pairs in the first stage of post-training?\nAnswer:", + " Multi-agent prompting" + ], + [ + "Question:Which technique is introduced to generate DPO pairs in the first stage of post-training?\nAnswer:", + " Instruction reversal" + ], + [ + "Question:Which technique is introduced to generate DPO pairs in the first stage of post-training?\nAnswer:", + " Pivotal Token Search" + ], + [ + "Question:Which technique is introduced to generate DPO pairs in the first stage of post-training?\nAnswer:", + " Majority Voting" + ] + ], + "resps": [ + [ + [ + -20.125656127929688, + false + ] + ], + [ + [ + -21.587875366210938, + false + ] + ], + [ + [ + -29.10210418701172, + false + ] + ], + [ + [ + -12.24392318725586, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.125656127929688, + false + ], + [ + -21.587875366210938, + false + ], + [ + -29.10210418701172, + false + ], + [ + -12.24392318725586, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a97af93f7f739a6c4de4934931a620cc1923def75f835215ff4af99c48c3f76c", + "prompt_hash": "203b12dfaa50c24d8451eb1877ce2f71f2cd03a96619328418be7cc1d1a9f44f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 74, + "doc": { + "question": "Which benchmark category showed the largest performance gap between synthetic and web-trained models in Table 8?", + "choices": [ + "Human-Eval", + "TQA", + "GSM8k", + "MATH" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which benchmark category showed the largest performance gap between synthetic and web-trained models in Table 8?\nAnswer:", + " Human-Eval" + ], + [ + "Question:Which benchmark category showed the largest performance gap between synthetic and web-trained models in Table 8?\nAnswer:", + " TQA" + ], + [ + "Question:Which benchmark category showed the largest performance gap between synthetic and web-trained models in Table 8?\nAnswer:", + " GSM8k" + ], + [ + "Question:Which benchmark category showed the largest performance gap between synthetic and web-trained models in Table 8?\nAnswer:", + " MATH" + ] + ], + "resps": [ + [ + [ + -14.4087495803833, + false + ] + ], + [ + [ + -13.021769523620605, + false + ] + ], + [ + [ + -11.766997337341309, + false + ] + ], + [ + [ + -11.660591125488281, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.4087495803833, + false + ], + [ + -13.021769523620605, + false + ], + [ + -11.766997337341309, + false + ], + [ + -11.660591125488281, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c39391ac6f10f661ca68d0a620f88fd776aa9c7bbc14c437618e02fdf3231050", + "prompt_hash": "1082716728d8aa876731dc059daf270a0d97c1ea049c003518d0de258d9dff00", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 75, + "doc": { + "question": "What percentage of Black and African American adults encounter online racial harassment?", + "choices": [ + "30%", + "46%", + "60%", + "75%" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What percentage of Black and African American adults encounter online racial harassment?\nAnswer:", + " 30%" + ], + [ + "Question:What percentage of Black and African American adults encounter online racial harassment?\nAnswer:", + " 46%" + ], + [ + "Question:What percentage of Black and African American adults encounter online racial harassment?\nAnswer:", + " 60%" + ], + [ + "Question:What percentage of Black and African American adults encounter online racial harassment?\nAnswer:", + " 75%" + ] + ], + "resps": [ + [ + [ + -9.151792526245117, + false + ] + ], + [ + [ + -11.020578384399414, + false + ] + ], + [ + [ + -9.901688575744629, + false + ] + ], + [ + [ + -10.72553825378418, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.151792526245117, + false + ], + [ + -11.020578384399414, + false + ], + [ + -9.901688575744629, + false + ], + [ + -10.72553825378418, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "defc532debfa109bd5786680e55093a0ac8f615953535eba76bfe5f1fe603eb3", + "prompt_hash": "5c5d70de7eece4a6a6270db730f523fd46d4b061e8b3c5c2c438cb55b5afded8", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 76, + "doc": { + "question": "What is the purpose of the Digital Services Act (DSA) enforced in 2024?", + "choices": [ + "To ban all forms of social media advertisements", + "To promote blockchain-based content verification", + "To enforce clear obligations for content moderation and explanation on online platforms", + "To encourage anonymous content posting" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the purpose of the Digital Services Act (DSA) enforced in 2024?\nAnswer:", + " To ban all forms of social media advertisements" + ], + [ + "Question:What is the purpose of the Digital Services Act (DSA) enforced in 2024?\nAnswer:", + " To promote blockchain-based content verification" + ], + [ + "Question:What is the purpose of the Digital Services Act (DSA) enforced in 2024?\nAnswer:", + " To enforce clear obligations for content moderation and explanation on online platforms" + ], + [ + "Question:What is the purpose of the Digital Services Act (DSA) enforced in 2024?\nAnswer:", + " To encourage anonymous content posting" + ] + ], + "resps": [ + [ + [ + -32.01473617553711, + false + ] + ], + [ + [ + -30.65254020690918, + false + ] + ], + [ + [ + -49.995948791503906, + false + ] + ], + [ + [ + -29.79315948486328, + false + ] + ] + ], + "filtered_resps": [ + [ + -32.01473617553711, + false + ], + [ + -30.65254020690918, + false + ], + [ + -49.995948791503906, + false + ], + [ + -29.79315948486328, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2b67327383aea1b2216df1851800ead1eda5cf24431edc6d2a49f3db192ea7ef", + "prompt_hash": "bf0c0fa613fc1e2f1ce5b58a456d14b2d1ca407b1e52a9726f3918adcb59213f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 77, + "doc": { + "question": "Which large language model is used as the teacher in the knowledge distillation approach?", + "choices": [ + "Llama-2-13B-Instruct", + "GPT-4", + "Llama-3-70B-Instruct", + "BERT-Large" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which large language model is used as the teacher in the knowledge distillation approach?\nAnswer:", + " Llama-2-13B-Instruct" + ], + [ + "Question:Which large language model is used as the teacher in the knowledge distillation approach?\nAnswer:", + " GPT-4" + ], + [ + "Question:Which large language model is used as the teacher in the knowledge distillation approach?\nAnswer:", + " Llama-3-70B-Instruct" + ], + [ + "Question:Which large language model is used as the teacher in the knowledge distillation approach?\nAnswer:", + " BERT-Large" + ] + ], + "resps": [ + [ + [ + -14.030332565307617, + false + ] + ], + [ + [ + -4.68262243270874, + false + ] + ], + [ + [ + -11.608107566833496, + false + ] + ], + [ + [ + -10.403993606567383, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.030332565307617, + false + ], + [ + -4.68262243270874, + false + ], + [ + -11.608107566833496, + false + ], + [ + -10.403993606567383, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "afbccc7ee3433afdb0a3b523c58b2d0138396c971149c757784844a5bc74d212", + "prompt_hash": "f89a8b088d90665636d8081e6fb13e0824bf92d489fe6a13c648631f685be539", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 78, + "doc": { + "question": "How many samples were correctly classified and used for distillation in MiniMetaHate Distil?", + "choices": [ + "2993", + "2296", + "2001", + "1500" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many samples were correctly classified and used for distillation in MiniMetaHate Distil?\nAnswer:", + " 2993" + ], + [ + "Question:How many samples were correctly classified and used for distillation in MiniMetaHate Distil?\nAnswer:", + " 2296" + ], + [ + "Question:How many samples were correctly classified and used for distillation in MiniMetaHate Distil?\nAnswer:", + " 2001" + ], + [ + "Question:How many samples were correctly classified and used for distillation in MiniMetaHate Distil?\nAnswer:", + " 1500" + ] + ], + "resps": [ + [ + [ + -11.522136688232422, + false + ] + ], + [ + [ + -11.24272346496582, + false + ] + ], + [ + [ + -10.519241333007812, + false + ] + ], + [ + [ + -7.333746910095215, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.522136688232422, + false + ], + [ + -11.24272346496582, + false + ], + [ + -10.519241333007812, + false + ], + [ + -7.333746910095215, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "54212365c2cf87b74ae0a6054ba14a270ab2ba0fb809d6595173a0de4afccd2e", + "prompt_hash": "4e9f63f3a6233cfb52db93ae4a1f485352f3a5cdad59fcc7aeec3b2230696e02", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 79, + "doc": { + "question": "What F1-score did the Llama-3-8B-Distil-MetaHate model achieve in the classification task?", + "choices": [ + "0.7796", + "0.8499", + "0.9037", + "0.7467" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What F1-score did the Llama-3-8B-Distil-MetaHate model achieve in the classification task?\nAnswer:", + " 0.7796" + ], + [ + "Question:What F1-score did the Llama-3-8B-Distil-MetaHate model achieve in the classification task?\nAnswer:", + " 0.8499" + ], + [ + "Question:What F1-score did the Llama-3-8B-Distil-MetaHate model achieve in the classification task?\nAnswer:", + " 0.9037" + ], + [ + "Question:What F1-score did the Llama-3-8B-Distil-MetaHate model achieve in the classification task?\nAnswer:", + " 0.7467" + ] + ], + "resps": [ + [ + [ + -11.481164932250977, + false + ] + ], + [ + [ + -11.787622451782227, + false + ] + ], + [ + [ + -11.448420524597168, + false + ] + ], + [ + [ + -11.004653930664062, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.481164932250977, + false + ], + [ + -11.787622451782227, + false + ], + [ + -11.448420524597168, + false + ], + [ + -11.004653930664062, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "682153f7962b13a3fca29e0ba040de1d172c39f2dab56ff965e7e9fd9342127e", + "prompt_hash": "0d9e96020a25c9fb03893ba494994b2140c438c15dcdc7e9106794b7f9f93ab8", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 80, + "doc": { + "question": "What is the native sequence length of ModernBERT?", + "choices": [ + "512", + "1024", + "4096", + "8192" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the native sequence length of ModernBERT?\nAnswer:", + " 512" + ], + [ + "Question:What is the native sequence length of ModernBERT?\nAnswer:", + " 1024" + ], + [ + "Question:What is the native sequence length of ModernBERT?\nAnswer:", + " 4096" + ], + [ + "Question:What is the native sequence length of ModernBERT?\nAnswer:", + " 8192" + ] + ], + "resps": [ + [ + [ + -2.9802372455596924, + false + ] + ], + [ + [ + -3.8660225868225098, + false + ] + ], + [ + [ + -5.204238414764404, + false + ] + ], + [ + [ + -5.671204566955566, + false + ] + ] + ], + "filtered_resps": [ + [ + -2.9802372455596924, + false + ], + [ + -3.8660225868225098, + false + ], + [ + -5.204238414764404, + false + ], + [ + -5.671204566955566, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4b66277959befb396d08c8d6472ff3dddff308651cc9aa98dc5093a5a147cb85", + "prompt_hash": "ab090baa0e014fb2cd8c2f57d2d56f86c9a8d4d9abdff9b048ee4bbda6745d16", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 81, + "doc": { + "question": "Which optimizer does ModernBERT use during pretraining?", + "choices": [ + "AdamW", + "StableAdamW", + "Adafactor", + "SGD" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which optimizer does ModernBERT use during pretraining?\nAnswer:", + " AdamW" + ], + [ + "Question:Which optimizer does ModernBERT use during pretraining?\nAnswer:", + " StableAdamW" + ], + [ + "Question:Which optimizer does ModernBERT use during pretraining?\nAnswer:", + " Adafactor" + ], + [ + "Question:Which optimizer does ModernBERT use during pretraining?\nAnswer:", + " SGD" + ] + ], + "resps": [ + [ + [ + -4.6689043045043945, + false + ] + ], + [ + [ + -20.804319381713867, + false + ] + ], + [ + [ + -9.3514986038208, + false + ] + ], + [ + [ + -7.600836277008057, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.6689043045043945, + false + ], + [ + -20.804319381713867, + false + ], + [ + -9.3514986038208, + false + ], + [ + -7.600836277008057, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "92bdbed4335dbf1e07e12e87e5059eda5d1bbb35503877e35e8c58a33b52b7f8", + "prompt_hash": "9a4aace23c9f432f3dee6b16fc1443472f954fe5fabf84c7504029042ad05ee5", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 82, + "doc": { + "question": "What type of activation function does ModernBERT adopt?", + "choices": [ + "ReLU", + "GeLU", + "Swish", + "GeGLU" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What type of activation function does ModernBERT adopt?\nAnswer:", + " ReLU" + ], + [ + "Question:What type of activation function does ModernBERT adopt?\nAnswer:", + " GeLU" + ], + [ + "Question:What type of activation function does ModernBERT adopt?\nAnswer:", + " Swish" + ], + [ + "Question:What type of activation function does ModernBERT adopt?\nAnswer:", + " GeGLU" + ] + ], + "resps": [ + [ + [ + -3.3449418544769287, + false + ] + ], + [ + [ + -9.327370643615723, + false + ] + ], + [ + [ + -7.710070610046387, + false + ] + ], + [ + [ + -20.871734619140625, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.3449418544769287, + false + ], + [ + -9.327370643615723, + false + ], + [ + -7.710070610046387, + false + ], + [ + -20.871734619140625, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "997911e4a17162c72338d72279b22fa19fa06027bb1a1084b09c859116effcbc", + "prompt_hash": "c9f1b89aadf77bba01b0161e076af8b57f527721531ac665734d68c788100eee", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 83, + "doc": { + "question": "What tokenizer does ModernBERT use?", + "choices": [ + "Original BERT tokenizer", + "SentencePiece", + "OLMo tokenizer (modified)", + "GPT2 tokenizer" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What tokenizer does ModernBERT use?\nAnswer:", + " Original BERT tokenizer" + ], + [ + "Question:What tokenizer does ModernBERT use?\nAnswer:", + " SentencePiece" + ], + [ + "Question:What tokenizer does ModernBERT use?\nAnswer:", + " OLMo tokenizer (modified)" + ], + [ + "Question:What tokenizer does ModernBERT use?\nAnswer:", + " GPT2 tokenizer" + ] + ], + "resps": [ + [ + [ + -10.338162422180176, + false + ] + ], + [ + [ + -8.317133903503418, + false + ] + ], + [ + [ + -37.740684509277344, + false + ] + ], + [ + [ + -12.453644752502441, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.338162422180176, + false + ], + [ + -8.317133903503418, + false + ], + [ + -37.740684509277344, + false + ], + [ + -12.453644752502441, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c4ded2d8c96422c7f89f3d5ddfb7fd19a7e6a387b40d9228443164a9ca00d5bc", + "prompt_hash": "d7848dcdb9379b33a0cf5ba68f10be54f71c1a2d2d1707173f52ac4845dddd3a", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 84, + "doc": { + "question": "What is the batch size warmup schedule used for ModernBERT-base?", + "choices": [ + "From 448 to 4608 over 50B tokens", + "From 768 to 4608 over 50B tokens", + "From 448 to 4928 over 10B tokens", + "From 768 to 4928 over 10B tokens" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the batch size warmup schedule used for ModernBERT-base?\nAnswer:", + " From 448 to 4608 over 50B tokens" + ], + [ + "Question:What is the batch size warmup schedule used for ModernBERT-base?\nAnswer:", + " From 768 to 4608 over 50B tokens" + ], + [ + "Question:What is the batch size warmup schedule used for ModernBERT-base?\nAnswer:", + " From 448 to 4928 over 10B tokens" + ], + [ + "Question:What is the batch size warmup schedule used for ModernBERT-base?\nAnswer:", + " From 768 to 4928 over 10B tokens" + ] + ], + "resps": [ + [ + [ + -54.454219818115234, + false + ] + ], + [ + [ + -48.003021240234375, + false + ] + ], + [ + [ + -53.04581832885742, + false + ] + ], + [ + [ + -49.49294662475586, + false + ] + ] + ], + "filtered_resps": [ + [ + -54.454219818115234, + false + ], + [ + -48.003021240234375, + false + ], + [ + -53.04581832885742, + false + ], + [ + -49.49294662475586, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "794189ecdd29ef84800d4d01b8ee091f8374a1baded0b020ce27070595a49962", + "prompt_hash": "b489d3303e83291a02e9332aac2328a83ff57fc1b2e696f1c0d58e4a0cd10347", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 85, + "doc": { + "question": "What is the maximum number of iterations allowed for generating a correct reasoning trajectory in Stage One?", + "choices": [ + "2", + "3", + "5", + "10" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the maximum number of iterations allowed for generating a correct reasoning trajectory in Stage One?\nAnswer:", + " 2" + ], + [ + "Question:What is the maximum number of iterations allowed for generating a correct reasoning trajectory in Stage One?\nAnswer:", + " 3" + ], + [ + "Question:What is the maximum number of iterations allowed for generating a correct reasoning trajectory in Stage One?\nAnswer:", + " 5" + ], + [ + "Question:What is the maximum number of iterations allowed for generating a correct reasoning trajectory in Stage One?\nAnswer:", + " 10" + ] + ], + "resps": [ + [ + [ + -3.444279432296753, + false + ] + ], + [ + [ + -3.4499194622039795, + false + ] + ], + [ + [ + -3.39045786857605, + false + ] + ], + [ + [ + -3.5267815589904785, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.444279432296753, + false + ], + [ + -3.4499194622039795, + false + ], + [ + -3.39045786857605, + false + ], + [ + -3.5267815589904785, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9523e0cd07b614fe1c199437bf4ffccc04af52eb1ecfc73735f856cdfec656c9", + "prompt_hash": "38df8aadb3c572104715f5ba3d1ca0b9a1204eae23019f3688978c3f1575359e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 86, + "doc": { + "question": "Which LLM is used as the verifier to check answers in the proposed method?", + "choices": [ + "GPT-4o", + "LLaMA-3.1", + "Qwen2.5", + "Gemma2" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which LLM is used as the verifier to check answers in the proposed method?\nAnswer:", + " GPT-4o" + ], + [ + "Question:Which LLM is used as the verifier to check answers in the proposed method?\nAnswer:", + " LLaMA-3.1" + ], + [ + "Question:Which LLM is used as the verifier to check answers in the proposed method?\nAnswer:", + " Qwen2.5" + ], + [ + "Question:Which LLM is used as the verifier to check answers in the proposed method?\nAnswer:", + " Gemma2" + ] + ], + "resps": [ + [ + [ + -4.99644136428833, + false + ] + ], + [ + [ + -19.128623962402344, + false + ] + ], + [ + [ + -15.504560470581055, + false + ] + ], + [ + [ + -11.37159538269043, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.99644136428833, + false + ], + [ + -19.128623962402344, + false + ], + [ + -15.504560470581055, + false + ], + [ + -11.37159538269043, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5eb7fa165860610eaab01627a1850068a08f3eb2da634b92c0023d437a7d7dfe", + "prompt_hash": "427983833d46e8c41fd395b7ee6c9234517a1a2aca2a96f5cac95570de9717af", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 87, + "doc": { + "question": "Which search strategy involves the LLM critiquing and correcting the current reasoning?", + "choices": [ + "Exploring New Paths", + "Backtracking", + "Verification", + "Corrections" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which search strategy involves the LLM critiquing and correcting the current reasoning?\nAnswer:", + " Exploring New Paths" + ], + [ + "Question:Which search strategy involves the LLM critiquing and correcting the current reasoning?\nAnswer:", + " Backtracking" + ], + [ + "Question:Which search strategy involves the LLM critiquing and correcting the current reasoning?\nAnswer:", + " Verification" + ], + [ + "Question:Which search strategy involves the LLM critiquing and correcting the current reasoning?\nAnswer:", + " Corrections" + ] + ], + "resps": [ + [ + [ + -21.948320388793945, + false + ] + ], + [ + [ + -8.776028633117676, + false + ] + ], + [ + [ + -7.903186321258545, + false + ] + ], + [ + [ + -12.345247268676758, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.948320388793945, + false + ], + [ + -8.776028633117676, + false + ], + [ + -7.903186321258545, + false + ], + [ + -12.345247268676758, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4e3eeed368a493138f5f47e056a32d6fe7d22c1104564f124cb321c66e5905f0", + "prompt_hash": "61066c3003ef4b501f9a45537746350905d46ccd3a17aad075de8ecae6d3cd9b", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 88, + "doc": { + "question": "What was the total number of medical exam questions collected from MedQA-USMLE and MedMcQA?", + "choices": [ + "40K", + "192K", + "60K", + "20K" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What was the total number of medical exam questions collected from MedQA-USMLE and MedMcQA?\nAnswer:", + " 40K" + ], + [ + "Question:What was the total number of medical exam questions collected from MedQA-USMLE and MedMcQA?\nAnswer:", + " 192K" + ], + [ + "Question:What was the total number of medical exam questions collected from MedQA-USMLE and MedMcQA?\nAnswer:", + " 60K" + ], + [ + "Question:What was the total number of medical exam questions collected from MedQA-USMLE and MedMcQA?\nAnswer:", + " 20K" + ] + ], + "resps": [ + [ + [ + -12.511892318725586, + false + ] + ], + [ + [ + -15.900592803955078, + false + ] + ], + [ + [ + -13.070356369018555, + false + ] + ], + [ + [ + -11.873079299926758, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.511892318725586, + false + ], + [ + -15.900592803955078, + false + ], + [ + -13.070356369018555, + false + ], + [ + -11.873079299926758, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e53c3edd00284a53963bdd1efe103a9bfe9388c5c2e84ac2badd3b9054ba2e05", + "prompt_hash": "f0dbc688373e4cddcbae3e6123c97a37e92a4b7bec4f53459bdf0d44a20e93a6", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 89, + "doc": { + "question": "What is the batch size used during reinforcement learning with PPO in Stage 2?", + "choices": [ + "32", + "64", + "128", + "256" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the batch size used during reinforcement learning with PPO in Stage 2?\nAnswer:", + " 32" + ], + [ + "Question:What is the batch size used during reinforcement learning with PPO in Stage 2?\nAnswer:", + " 64" + ], + [ + "Question:What is the batch size used during reinforcement learning with PPO in Stage 2?\nAnswer:", + " 128" + ], + [ + "Question:What is the batch size used during reinforcement learning with PPO in Stage 2?\nAnswer:", + " 256" + ] + ], + "resps": [ + [ + [ + -4.313339710235596, + false + ] + ], + [ + [ + -4.934946060180664, + false + ] + ], + [ + [ + -5.580106735229492, + false + ] + ], + [ + [ + -5.842267036437988, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.313339710235596, + false + ], + [ + -4.934946060180664, + false + ], + [ + -5.580106735229492, + false + ], + [ + -5.842267036437988, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "14a8b0bdb1427efa8f3f46a44fe01b61bf71b7e9be24826304669f9d38aec6be", + "prompt_hash": "494d5b020999ea125ad81c2789a6df16a06f4d8ce120053db53ed5f24704b07f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 90, + "doc": { + "question": "What is the total duration of the 30 million video clips used in the dataset?", + "choices": [ + "50k hours", + "60k hours", + "70k hours", + "80k hours" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the total duration of the 30 million video clips used in the dataset?\nAnswer:", + " 50k hours" + ], + [ + "Question:What is the total duration of the 30 million video clips used in the dataset?\nAnswer:", + " 60k hours" + ], + [ + "Question:What is the total duration of the 30 million video clips used in the dataset?\nAnswer:", + " 70k hours" + ], + [ + "Question:What is the total duration of the 30 million video clips used in the dataset?\nAnswer:", + " 80k hours" + ] + ], + "resps": [ + [ + [ + -13.65107250213623, + false + ] + ], + [ + [ + -14.15365219116211, + false + ] + ], + [ + [ + -14.824312210083008, + false + ] + ], + [ + [ + -14.518945693969727, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.65107250213623, + false + ], + [ + -14.15365219116211, + false + ], + [ + -14.824312210083008, + false + ], + [ + -14.518945693969727, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b50a124a9aa6ab1d5bd0f4632abb993e6849b6e9cb9696575ab03d0a9095094f", + "prompt_hash": "810bc5d920fb28c94ecd9f2fc18f5a8fa44857c36784468bb6965d9551f53cb9", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 91, + "doc": { + "question": "Which model is used for Optical Flow Score computation in the data preprocessing pipeline?", + "choices": [ + "RAFT", + "FlowNet", + "UniMatch", + "PWC-Net" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model is used for Optical Flow Score computation in the data preprocessing pipeline?\nAnswer:", + " RAFT" + ], + [ + "Question:Which model is used for Optical Flow Score computation in the data preprocessing pipeline?\nAnswer:", + " FlowNet" + ], + [ + "Question:Which model is used for Optical Flow Score computation in the data preprocessing pipeline?\nAnswer:", + " UniMatch" + ], + [ + "Question:Which model is used for Optical Flow Score computation in the data preprocessing pipeline?\nAnswer:", + " PWC-Net" + ] + ], + "resps": [ + [ + [ + -6.261064529418945, + false + ] + ], + [ + [ + -6.434628009796143, + false + ] + ], + [ + [ + -11.806937217712402, + false + ] + ], + [ + [ + -14.45037841796875, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.261064529418945, + false + ], + [ + -6.434628009796143, + false + ], + [ + -11.806937217712402, + false + ], + [ + -14.45037841796875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "88cec1a4202acd1a92b3ada2606234ac3e46712884ab27ab2a207ce3ecf652b9", + "prompt_hash": "102011f9bdda56af7503599cf0188d34031ccdf8079b632707254ad248936e1d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 92, + "doc": { + "question": "How many total parameters does the video compression network contain?", + "choices": [ + "300M", + "384M", + "400M", + "420M" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many total parameters does the video compression network contain?\nAnswer:", + " 300M" + ], + [ + "Question:How many total parameters does the video compression network contain?\nAnswer:", + " 384M" + ], + [ + "Question:How many total parameters does the video compression network contain?\nAnswer:", + " 400M" + ], + [ + "Question:How many total parameters does the video compression network contain?\nAnswer:", + " 420M" + ] + ], + "resps": [ + [ + [ + -9.516914367675781, + false + ] + ], + [ + [ + -12.171443939208984, + false + ] + ], + [ + [ + -9.804231643676758, + false + ] + ], + [ + [ + -11.168252944946289, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.516914367675781, + false + ], + [ + -12.171443939208984, + false + ], + [ + -9.804231643676758, + false + ], + [ + -11.168252944946289, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7b265aa594361a305450a7af84e6f6ac5abb1d6aa1c4d9a9f888b90df577f0b7", + "prompt_hash": "4d644297715562776f3430e89931dfe7c0191a6a028d9143ffcd10bf90ea95e1", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 93, + "doc": { + "question": "Which model was used to initialize the 2D VAE in Open-Sora 1.2?", + "choices": [ + "Latte", + "PixArt", + "SDXL", + "Magvit-v2" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model was used to initialize the 2D VAE in Open-Sora 1.2?\nAnswer:", + " Latte" + ], + [ + "Question:Which model was used to initialize the 2D VAE in Open-Sora 1.2?\nAnswer:", + " PixArt" + ], + [ + "Question:Which model was used to initialize the 2D VAE in Open-Sora 1.2?\nAnswer:", + " SDXL" + ], + [ + "Question:Which model was used to initialize the 2D VAE in Open-Sora 1.2?\nAnswer:", + " Magvit-v2" + ] + ], + "resps": [ + [ + [ + -15.276412963867188, + false + ] + ], + [ + [ + -10.8336820602417, + false + ] + ], + [ + [ + -12.574335098266602, + false + ] + ], + [ + [ + -30.271799087524414, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.276412963867188, + false + ], + [ + -10.8336820602417, + false + ], + [ + -12.574335098266602, + false + ], + [ + -30.271799087524414, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "faa427f70ad5b6417f6829fa64227f2aebbbca4760908a4dfe6693da596a2743", + "prompt_hash": "67248fc329805136e7e6d18124cb719c6c66be23e0c2c66618faf8c2f43c333d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 94, + "doc": { + "question": "What was the VBench total score achieved by Open-Sora 1.2?", + "choices": [ + "75.91%", + "77.23%", + "78.93%", + "79.76%" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What was the VBench total score achieved by Open-Sora 1.2?\nAnswer:", + " 75.91%" + ], + [ + "Question:What was the VBench total score achieved by Open-Sora 1.2?\nAnswer:", + " 77.23%" + ], + [ + "Question:What was the VBench total score achieved by Open-Sora 1.2?\nAnswer:", + " 78.93%" + ], + [ + "Question:What was the VBench total score achieved by Open-Sora 1.2?\nAnswer:", + " 79.76%" + ] + ], + "resps": [ + [ + [ + -18.253808975219727, + false + ] + ], + [ + [ + -17.968246459960938, + false + ] + ], + [ + [ + -17.808147430419922, + false + ] + ], + [ + [ + -17.937244415283203, + false + ] + ] + ], + "filtered_resps": [ + [ + -18.253808975219727, + false + ], + [ + -17.968246459960938, + false + ], + [ + -17.808147430419922, + false + ], + [ + -17.937244415283203, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "996839ff1783e652a55de6089818380d8cb1ac4077a073134dd9e0eb43823415", + "prompt_hash": "a6d9d72a424e7f01bcdbf2bc42ea87b275074d665b20737780add009981a4a02", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 95, + "doc": { + "question": "What is the total spatio-temporal compression ratio achieved by the Video-VAE in LTX-Video?", + "choices": [ + "1:48", + "1:96", + "1:192", + "1:512" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the total spatio-temporal compression ratio achieved by the Video-VAE in LTX-Video?\nAnswer:", + " 1:48" + ], + [ + "Question:What is the total spatio-temporal compression ratio achieved by the Video-VAE in LTX-Video?\nAnswer:", + " 1:96" + ], + [ + "Question:What is the total spatio-temporal compression ratio achieved by the Video-VAE in LTX-Video?\nAnswer:", + " 1:192" + ], + [ + "Question:What is the total spatio-temporal compression ratio achieved by the Video-VAE in LTX-Video?\nAnswer:", + " 1:512" + ] + ], + "resps": [ + [ + [ + -14.28443717956543, + false + ] + ], + [ + [ + -14.824919700622559, + false + ] + ], + [ + [ + -17.184690475463867, + false + ] + ], + [ + [ + -17.0477294921875, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.28443717956543, + false + ], + [ + -14.824919700622559, + false + ], + [ + -17.184690475463867, + false + ], + [ + -17.0477294921875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "601e102d0995f5a8d207082151dede875cf07ba21217427be49a4b7f25e9d2f8", + "prompt_hash": "4f2d44c5a4304265e085751f62a4638452aca0db5263a9948634e4742ee519c7", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 96, + "doc": { + "question": "Which positional embedding method is used in LTX-Video instead of absolute positional embeddings?", + "choices": [ + "Sinusoidal Embeddings", + "Rotary Positional Embeddings (RoPE)", + "Learned Positional Vectors", + "Relative Position Bias" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which positional embedding method is used in LTX-Video instead of absolute positional embeddings?\nAnswer:", + " Sinusoidal Embeddings" + ], + [ + "Question:Which positional embedding method is used in LTX-Video instead of absolute positional embeddings?\nAnswer:", + " Rotary Positional Embeddings (RoPE)" + ], + [ + "Question:Which positional embedding method is used in LTX-Video instead of absolute positional embeddings?\nAnswer:", + " Learned Positional Vectors" + ], + [ + "Question:Which positional embedding method is used in LTX-Video instead of absolute positional embeddings?\nAnswer:", + " Relative Position Bias" + ] + ], + "resps": [ + [ + [ + -14.810482025146484, + false + ] + ], + [ + [ + -15.664023399353027, + false + ] + ], + [ + [ + -22.923189163208008, + false + ] + ], + [ + [ + -15.321342468261719, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.810482025146484, + false + ], + [ + -15.664023399353027, + false + ], + [ + -22.923189163208008, + false + ], + [ + -15.321342468261719, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5af380cb6067d02b663a4861dfaac101b9b090a294539c623fdb1abefe4150cf", + "prompt_hash": "9c151ca1e15f6348d2eb1c9188f2ee28970395133d7cad4793fd4a80bdc60f30", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 97, + "doc": { + "question": "On which GPU did LTX-Video achieve faster-than-real-time generation of 121 frames in 2 seconds?", + "choices": [ + "Nvidia A100", + "Nvidia V100", + "Nvidia H100", + "Nvidia RTX 4090" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:On which GPU did LTX-Video achieve faster-than-real-time generation of 121 frames in 2 seconds?\nAnswer:", + " Nvidia A100" + ], + [ + "Question:On which GPU did LTX-Video achieve faster-than-real-time generation of 121 frames in 2 seconds?\nAnswer:", + " Nvidia V100" + ], + [ + "Question:On which GPU did LTX-Video achieve faster-than-real-time generation of 121 frames in 2 seconds?\nAnswer:", + " Nvidia H100" + ], + [ + "Question:On which GPU did LTX-Video achieve faster-than-real-time generation of 121 frames in 2 seconds?\nAnswer:", + " Nvidia RTX 4090" + ] + ], + "resps": [ + [ + [ + -7.209535598754883, + false + ] + ], + [ + [ + -9.387601852416992, + false + ] + ], + [ + [ + -9.177935600280762, + false + ] + ], + [ + [ + -8.791450500488281, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.209535598754883, + false + ], + [ + -9.387601852416992, + false + ], + [ + -9.177935600280762, + false + ], + [ + -8.791450500488281, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "44eedd9be0c5f64eaa9704b437069887928e26c434da8c23defedc0628d95e41", + "prompt_hash": "02c313177014283fc2d2ceb3fb889083c088a5ecb4651826427b61531023f9c4", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 98, + "doc": { + "question": "Which GAN variant is proposed in LTX-Video to improve stability and reconstruction quality?", + "choices": [ + "PatchGAN", + "CycleGAN", + "Reconstruction GAN", + "StyleGAN" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which GAN variant is proposed in LTX-Video to improve stability and reconstruction quality?\nAnswer:", + " PatchGAN" + ], + [ + "Question:Which GAN variant is proposed in LTX-Video to improve stability and reconstruction quality?\nAnswer:", + " CycleGAN" + ], + [ + "Question:Which GAN variant is proposed in LTX-Video to improve stability and reconstruction quality?\nAnswer:", + " Reconstruction GAN" + ], + [ + "Question:Which GAN variant is proposed in LTX-Video to improve stability and reconstruction quality?\nAnswer:", + " StyleGAN" + ] + ], + "resps": [ + [ + [ + -8.652844429016113, + false + ] + ], + [ + [ + -5.183656692504883, + false + ] + ], + [ + [ + -11.628631591796875, + false + ] + ], + [ + [ + -5.3855133056640625, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.652844429016113, + false + ], + [ + -5.183656692504883, + false + ], + [ + -11.628631591796875, + false + ], + [ + -5.3855133056640625, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4d11afddf855da867964775b6d488b39809add0de3182b038f66d8e7bc791255", + "prompt_hash": "3a76cad05d271f8f47d95186dc29e621e8ecd538ca06bc1cd34fa32a22af9d62", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 99, + "doc": { + "question": "What pre-trained text encoder is used for generating initial text embeddings in LTX-Video?", + "choices": [ + "BERT", + "CLIP", + "T5", + "GPT-2" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What pre-trained text encoder is used for generating initial text embeddings in LTX-Video?\nAnswer:", + " BERT" + ], + [ + "Question:What pre-trained text encoder is used for generating initial text embeddings in LTX-Video?\nAnswer:", + " CLIP" + ], + [ + "Question:What pre-trained text encoder is used for generating initial text embeddings in LTX-Video?\nAnswer:", + " T5" + ], + [ + "Question:What pre-trained text encoder is used for generating initial text embeddings in LTX-Video?\nAnswer:", + " GPT-2" + ] + ], + "resps": [ + [ + [ + -3.287250518798828, + false + ] + ], + [ + [ + -3.3353793621063232, + false + ] + ], + [ + [ + -5.748715400695801, + false + ] + ], + [ + [ + -7.809186935424805, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.287250518798828, + false + ], + [ + -3.3353793621063232, + false + ], + [ + -5.748715400695801, + false + ], + [ + -7.809186935424805, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "72e9c47fb45eea2ab469fb1ddf4f6bf24453e7843260a0b9eadfb41eb325081d", + "prompt_hash": "296ebf6f03fd43f3f28379d8d020e501ea08a43ed5d73900fd472c11cfbc5053", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + } + ], + "arxiv_mc_2501": [ + { + "doc_id": 0, + "doc": { + "question": "What does an approximate unitary k-design aim to replicate?", + "choices": [ + "The behavior of a classical circuit with k gates", + "The full entanglement structure of any quantum state", + "The behavior of Haar-random unitaries within any quantum experiment that queries the unitary k times", + "The output of a fully depolarizing quantum channel" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does an approximate unitary k-design aim to replicate?\nAnswer:", + " The behavior of a classical circuit with k gates" + ], + [ + "Question:What does an approximate unitary k-design aim to replicate?\nAnswer:", + " The full entanglement structure of any quantum state" + ], + [ + "Question:What does an approximate unitary k-design aim to replicate?\nAnswer:", + " The behavior of Haar-random unitaries within any quantum experiment that queries the unitary k times" + ], + [ + "Question:What does an approximate unitary k-design aim to replicate?\nAnswer:", + " The output of a fully depolarizing quantum channel" + ] + ], + "resps": [ + [ + [ + -29.659969329833984, + false + ] + ], + [ + [ + -27.531269073486328, + false + ] + ], + [ + [ + -79.02543640136719, + false + ] + ], + [ + [ + -26.968955993652344, + false + ] + ] + ], + "filtered_resps": [ + [ + -29.659969329833984, + false + ], + [ + -27.531269073486328, + false + ], + [ + -79.02543640136719, + false + ], + [ + -26.968955993652344, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7e116631a34f93e263eba5412718a33e303dfe605621305a2d9162d5f93b5e30", + "prompt_hash": "878dab46c9267372fec18d8fa960c96a12614f77781a17b6117cc17825a75e71", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 1, + "doc": { + "question": "What is the required local patch size ξ to ensure the ensemble forms an ε-approximate unitary k-design on n qubits?", + "choices": [ + "ξ ≥ √(nk/ε)", + "ξ ≥ log₂(nk²/ε)", + "ξ ≥ ε / nk", + "ξ ≥ (n + k + ε) / 2" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the required local patch size ξ to ensure the ensemble forms an ε-approximate unitary k-design on n qubits?\nAnswer:", + " ξ ≥ √(nk/ε)" + ], + [ + "Question:What is the required local patch size ξ to ensure the ensemble forms an ε-approximate unitary k-design on n qubits?\nAnswer:", + " ξ ≥ log₂(nk²/ε)" + ], + [ + "Question:What is the required local patch size ξ to ensure the ensemble forms an ε-approximate unitary k-design on n qubits?\nAnswer:", + " ξ ≥ ε / nk" + ], + [ + "Question:What is the required local patch size ξ to ensure the ensemble forms an ε-approximate unitary k-design on n qubits?\nAnswer:", + " ξ ≥ (n + k + ε) / 2" + ] + ], + "resps": [ + [ + [ + -49.53298568725586, + false + ] + ], + [ + [ + -54.27104187011719, + false + ] + ], + [ + [ + -37.01532745361328, + false + ] + ], + [ + [ + -42.08514404296875, + false + ] + ] + ], + "filtered_resps": [ + [ + -49.53298568725586, + false + ], + [ + -54.27104187011719, + false + ], + [ + -37.01532745361328, + false + ], + [ + -42.08514404296875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "102db8f0a456c5c4a4df5e6a3b89c1d0a0c152ab74bc9b0d55821233e4201790", + "prompt_hash": "ec0c42f5b8cc384eacbbfe02d3206f129a055df415b34dfaf626feddae9d4ec9", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 2, + "doc": { + "question": "What circuit depth is sufficient to form pseudorandom unitaries on all-to-all-connected circuits?", + "choices": [ + "poly(n)", + "poly(log log n)", + "poly(n log n)", + "poly(sqrt(n))" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What circuit depth is sufficient to form pseudorandom unitaries on all-to-all-connected circuits?\nAnswer:", + " poly(n)" + ], + [ + "Question:What circuit depth is sufficient to form pseudorandom unitaries on all-to-all-connected circuits?\nAnswer:", + " poly(log log n)" + ], + [ + "Question:What circuit depth is sufficient to form pseudorandom unitaries on all-to-all-connected circuits?\nAnswer:", + " poly(n log n)" + ], + [ + "Question:What circuit depth is sufficient to form pseudorandom unitaries on all-to-all-connected circuits?\nAnswer:", + " poly(sqrt(n))" + ] + ], + "resps": [ + [ + [ + -16.263832092285156, + false + ] + ], + [ + [ + -24.4702091217041, + false + ] + ], + [ + [ + -24.825349807739258, + false + ] + ], + [ + [ + -23.393047332763672, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.263832092285156, + false + ], + [ + -24.4702091217041, + false + ], + [ + -24.825349807739258, + false + ], + [ + -23.393047332763672, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "dbf5f6346c8e1d783a57e99b79cfa23e88d4e708818bfc4a3a556358d731c5fe", + "prompt_hash": "fd0f10c0a89cb75f0efbdb7c152212d7e9c6fb21be2bf75043c1a09204a8b9b3", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 3, + "doc": { + "question": "Why can random quantum circuits appear Haar-random in exponentially lower depth compared to classical circuits?", + "choices": [ + "Quantum circuits use fewer gates", + "Quantum circuits operate only on classical bits", + "Quantum circuits can hide information in non-commuting observables", + "Quantum circuits avoid measurement altogether" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Why can random quantum circuits appear Haar-random in exponentially lower depth compared to classical circuits?\nAnswer:", + " Quantum circuits use fewer gates" + ], + [ + "Question:Why can random quantum circuits appear Haar-random in exponentially lower depth compared to classical circuits?\nAnswer:", + " Quantum circuits operate only on classical bits" + ], + [ + "Question:Why can random quantum circuits appear Haar-random in exponentially lower depth compared to classical circuits?\nAnswer:", + " Quantum circuits can hide information in non-commuting observables" + ], + [ + "Question:Why can random quantum circuits appear Haar-random in exponentially lower depth compared to classical circuits?\nAnswer:", + " Quantum circuits avoid measurement altogether" + ] + ], + "resps": [ + [ + [ + -17.355615615844727, + false + ] + ], + [ + [ + -24.588275909423828, + false + ] + ], + [ + [ + -35.55263900756836, + false + ] + ], + [ + [ + -26.179462432861328, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.355615615844727, + false + ], + [ + -24.588275909423828, + false + ], + [ + -35.55263900756836, + false + ], + [ + -26.179462432861328, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8942b343b5900cf80761e00cc194c309d61c7b88a17de5c8c41435096b338eb9", + "prompt_hash": "5f3c72693fae80533b25db692fda49f4cf0382dd96efb58bd5a1b3a719962ca7", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 4, + "doc": { + "question": "What is the minimum circuit depth required for orthogonal quantum circuits to match the second moment of a Haar-random orthogonal matrix in 1D?", + "choices": [ + "Ω(log log n)", + "Ω(n)", + "Ω(log n)", + "Ω(1)" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the minimum circuit depth required for orthogonal quantum circuits to match the second moment of a Haar-random orthogonal matrix in 1D?\nAnswer:", + " Ω(log log n)" + ], + [ + "Question:What is the minimum circuit depth required for orthogonal quantum circuits to match the second moment of a Haar-random orthogonal matrix in 1D?\nAnswer:", + " Ω(n)" + ], + [ + "Question:What is the minimum circuit depth required for orthogonal quantum circuits to match the second moment of a Haar-random orthogonal matrix in 1D?\nAnswer:", + " Ω(log n)" + ], + [ + "Question:What is the minimum circuit depth required for orthogonal quantum circuits to match the second moment of a Haar-random orthogonal matrix in 1D?\nAnswer:", + " Ω(1)" + ] + ], + "resps": [ + [ + [ + -32.244606018066406, + false + ] + ], + [ + [ + -26.079303741455078, + false + ] + ], + [ + [ + -28.070377349853516, + false + ] + ], + [ + [ + -25.273380279541016, + false + ] + ] + ], + "filtered_resps": [ + [ + -32.244606018066406, + false + ], + [ + -26.079303741455078, + false + ], + [ + -28.070377349853516, + false + ], + [ + -25.273380279541016, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c14fd056edad82ec32d1bed34a49d500307e4ba5d553ffaa2100964e31b87fd7", + "prompt_hash": "82814ab236197202c0eaa4e72f1fb33f558dcd6525b1e3e6b2f1db9e735ec159", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 5, + "doc": { + "question": "What type of product is denoted by the symbol \\\\( \\\\odot \\\\) in the notation section?", + "choices": [ + "Kronecker product", + "Outer product", + "Hadamard product", + "Khatri-Rao product" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What type of product is denoted by the symbol \\\\( \\\\odot \\\\) in the notation section?\nAnswer:", + " Kronecker product" + ], + [ + "Question:What type of product is denoted by the symbol \\\\( \\\\odot \\\\) in the notation section?\nAnswer:", + " Outer product" + ], + [ + "Question:What type of product is denoted by the symbol \\\\( \\\\odot \\\\) in the notation section?\nAnswer:", + " Hadamard product" + ], + [ + "Question:What type of product is denoted by the symbol \\\\( \\\\odot \\\\) in the notation section?\nAnswer:", + " Khatri-Rao product" + ] + ], + "resps": [ + [ + [ + -13.476451873779297, + false + ] + ], + [ + [ + -15.377878189086914, + false + ] + ], + [ + [ + -12.29111099243164, + false + ] + ], + [ + [ + -15.445086479187012, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.476451873779297, + false + ], + [ + -15.377878189086914, + false + ], + [ + -12.29111099243164, + false + ], + [ + -15.445086479187012, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7059aba6a93a63f69559df1560312849767e51568200900697eb0d21ba820d66", + "prompt_hash": "63775c730f165c2cdf883a1db71000f8749c0563ab2ce07d1e2bb009870f4d05", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 6, + "doc": { + "question": "What is the role of the matrix \\\\( \\\\bm{\\\\Sigma} \\\\) in the MIMO channel model?", + "choices": [ + "It represents the antenna gain matrix.", + "It denotes the noise correlation matrix.", + "It denotes the path response matrix (PRM) between the Tx and Rx regions.", + "It is the pilot signal matrix." + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the role of the matrix \\\\( \\\\bm{\\\\Sigma} \\\\) in the MIMO channel model?\nAnswer:", + " It represents the antenna gain matrix." + ], + [ + "Question:What is the role of the matrix \\\\( \\\\bm{\\\\Sigma} \\\\) in the MIMO channel model?\nAnswer:", + " It denotes the noise correlation matrix." + ], + [ + "Question:What is the role of the matrix \\\\( \\\\bm{\\\\Sigma} \\\\) in the MIMO channel model?\nAnswer:", + " It denotes the path response matrix (PRM) between the Tx and Rx regions." + ], + [ + "Question:What is the role of the matrix \\\\( \\\\bm{\\\\Sigma} \\\\) in the MIMO channel model?\nAnswer:", + " It is the pilot signal matrix." + ] + ], + "resps": [ + [ + [ + -21.347972869873047, + false + ] + ], + [ + [ + -20.47518539428711, + false + ] + ], + [ + [ + -52.65840148925781, + false + ] + ], + [ + [ + -24.194866180419922, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.347972869873047, + false + ], + [ + -20.47518539428711, + false + ], + [ + -52.65840148925781, + false + ], + [ + -24.194866180419922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f3ea440e3b590ca8bf5e1a9ff43163c026a3988f19f3265b244603813e64d127", + "prompt_hash": "cf332b053f0076252bfb678cd15aae63ca2ba6dd97670499a3fb759be57a4062", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 7, + "doc": { + "question": "What is the purpose of introducing a two-stage Tx-Rx successive antenna movement pattern?", + "choices": [ + "To minimize the number of antenna ports.", + "To simplify the hardware implementation.", + "To allow the received pilot signals to be expressed as a third-order tensor.", + "To estimate the pilot noise covariance matrix." + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the purpose of introducing a two-stage Tx-Rx successive antenna movement pattern?\nAnswer:", + " To minimize the number of antenna ports." + ], + [ + "Question:What is the purpose of introducing a two-stage Tx-Rx successive antenna movement pattern?\nAnswer:", + " To simplify the hardware implementation." + ], + [ + "Question:What is the purpose of introducing a two-stage Tx-Rx successive antenna movement pattern?\nAnswer:", + " To allow the received pilot signals to be expressed as a third-order tensor." + ], + [ + "Question:What is the purpose of introducing a two-stage Tx-Rx successive antenna movement pattern?\nAnswer:", + " To estimate the pilot noise covariance matrix." + ] + ], + "resps": [ + [ + [ + -21.519088745117188, + false + ] + ], + [ + [ + -15.229227066040039, + false + ] + ], + [ + [ + -46.79536819458008, + false + ] + ], + [ + [ + -31.928115844726562, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.519088745117188, + false + ], + [ + -15.229227066040039, + false + ], + [ + -46.79536819458008, + false + ], + [ + -31.928115844726562, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "cd078c68e5271f1789f1ad38aff800702fe05b176f29fd943f37fefa4ef38bd7", + "prompt_hash": "b7c805f3a0e82a3b9b69051224bf844789cea13c43f241f8e8a5735274003aac", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 8, + "doc": { + "question": "Which algorithm is used to estimate the factor matrices from the tensor in the proposed method?", + "choices": [ + "Gradient Descent", + "Stochastic Gradient Descent", + "Alternating Least Squares (ALS)", + "Expectation Maximization (EM)" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which algorithm is used to estimate the factor matrices from the tensor in the proposed method?\nAnswer:", + " Gradient Descent" + ], + [ + "Question:Which algorithm is used to estimate the factor matrices from the tensor in the proposed method?\nAnswer:", + " Stochastic Gradient Descent" + ], + [ + "Question:Which algorithm is used to estimate the factor matrices from the tensor in the proposed method?\nAnswer:", + " Alternating Least Squares (ALS)" + ], + [ + "Question:Which algorithm is used to estimate the factor matrices from the tensor in the proposed method?\nAnswer:", + " Expectation Maximization (EM)" + ] + ], + "resps": [ + [ + [ + -10.288975715637207, + false + ] + ], + [ + [ + -10.533903121948242, + false + ] + ], + [ + [ + -10.85214614868164, + false + ] + ], + [ + [ + -9.711283683776855, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.288975715637207, + false + ], + [ + -10.533903121948242, + false + ], + [ + -10.85214614868164, + false + ], + [ + -9.711283683776855, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "448aca9f20641c880937f0058af91cb1edc19f3cfd05c3b986c49cfacbe4131b", + "prompt_hash": "34a9086f0fa49009032bd6411419757a73cbd4732e082dd3f55b163c47f3d75e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 9, + "doc": { + "question": "According to Lemma 1, what is the inequality condition for the uniqueness of the CP decomposition of a third-order tensor?", + "choices": [ + "\\( \\min(I_1, L) + \\min(I_2, L) + \\min(I_3, L) \\geq 2L + 1 \\)", + "\\( \\min(I_1, L) + \\min(I_2, L) + \\min(I_3, L) \\leq 2L + 2 \\)", + "\\( \\min(I_1, L) + \\min(I_2, L) + \\min(I_3, L) \\geq 2L + 2 \\)", + "\\( I_1 + I_2 + I_3 \\geq 3L \\)" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:According to Lemma 1, what is the inequality condition for the uniqueness of the CP decomposition of a third-order tensor?\nAnswer:", + " \\( \\min(I_1, L) + \\min(I_2, L) + \\min(I_3, L) \\geq 2L + 1 \\)" + ], + [ + "Question:According to Lemma 1, what is the inequality condition for the uniqueness of the CP decomposition of a third-order tensor?\nAnswer:", + " \\( \\min(I_1, L) + \\min(I_2, L) + \\min(I_3, L) \\leq 2L + 2 \\)" + ], + [ + "Question:According to Lemma 1, what is the inequality condition for the uniqueness of the CP decomposition of a third-order tensor?\nAnswer:", + " \\( \\min(I_1, L) + \\min(I_2, L) + \\min(I_3, L) \\geq 2L + 2 \\)" + ], + [ + "Question:According to Lemma 1, what is the inequality condition for the uniqueness of the CP decomposition of a third-order tensor?\nAnswer:", + " \\( I_1 + I_2 + I_3 \\geq 3L \\)" + ] + ], + "resps": [ + [ + [ + -57.100379943847656, + false + ] + ], + [ + [ + -57.3333740234375, + false + ] + ], + [ + [ + -57.717796325683594, + false + ] + ], + [ + [ + -37.311668395996094, + false + ] + ] + ], + "filtered_resps": [ + [ + -57.100379943847656, + false + ], + [ + -57.3333740234375, + false + ], + [ + -57.717796325683594, + false + ], + [ + -37.311668395996094, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "73566678c7792ba0c12d20d4df5544540424788b02bbdfc5520581d7005ffa11", + "prompt_hash": "042d5972389ff5a870b7a6352b566ad9b4781c8c36433ffe586761307d59ad12", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 10, + "doc": { + "question": "What is the total number of parameters activated per text token in the described MoE model?", + "choices": [ + "3.5B", + "24.9B", + "66M", + "438M" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the total number of parameters activated per text token in the described MoE model?\nAnswer:", + " 3.5B" + ], + [ + "Question:What is the total number of parameters activated per text token in the described MoE model?\nAnswer:", + " 24.9B" + ], + [ + "Question:What is the total number of parameters activated per text token in the described MoE model?\nAnswer:", + " 66M" + ], + [ + "Question:What is the total number of parameters activated per text token in the described MoE model?\nAnswer:", + " 438M" + ] + ], + "resps": [ + [ + [ + -13.006896018981934, + false + ] + ], + [ + [ + -17.268253326416016, + false + ] + ], + [ + [ + -10.897493362426758, + false + ] + ], + [ + [ + -13.124433517456055, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.006896018981934, + false + ], + [ + -17.268253326416016, + false + ], + [ + -10.897493362426758, + false + ], + [ + -13.124433517456055, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c6afc5680e8d3dceb6ef438860605a3e0542ca1cce44927569aac288145823d6", + "prompt_hash": "c50de2733f2cf5cd1bba88a2366bb4ae37ade03fd2675d4d2e9050a4239b6220", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 11, + "doc": { + "question": "Which model is used to initialize the weights of the Vision Transformer (ViT) in the visual encoder?", + "choices": [ + "BLIP2", + "SigLIP-SO400M", + "CLIP", + "Qwen-VL" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which model is used to initialize the weights of the Vision Transformer (ViT) in the visual encoder?\nAnswer:", + " BLIP2" + ], + [ + "Question:Which model is used to initialize the weights of the Vision Transformer (ViT) in the visual encoder?\nAnswer:", + " SigLIP-SO400M" + ], + [ + "Question:Which model is used to initialize the weights of the Vision Transformer (ViT) in the visual encoder?\nAnswer:", + " CLIP" + ], + [ + "Question:Which model is used to initialize the weights of the Vision Transformer (ViT) in the visual encoder?\nAnswer:", + " Qwen-VL" + ] + ], + "resps": [ + [ + [ + -12.5610933303833, + false + ] + ], + [ + [ + -35.3951301574707, + false + ] + ], + [ + [ + -6.808763027191162, + false + ] + ], + [ + [ + -10.143580436706543, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.5610933303833, + false + ], + [ + -35.3951301574707, + false + ], + [ + -6.808763027191162, + false + ], + [ + -10.143580436706543, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "439f9fb56c44368ec0d4edbf4d6e0c4b213a2838c243fa648166b17164147572", + "prompt_hash": "12e07657b42f63eb9020cf762b1b06f77d3ea8123fa5b01679b6e03bfa24502f", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 12, + "doc": { + "question": "How many experts are there in each MoE layer of the model?", + "choices": [ + "12", + "32", + "66", + "128" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many experts are there in each MoE layer of the model?\nAnswer:", + " 12" + ], + [ + "Question:How many experts are there in each MoE layer of the model?\nAnswer:", + " 32" + ], + [ + "Question:How many experts are there in each MoE layer of the model?\nAnswer:", + " 66" + ], + [ + "Question:How many experts are there in each MoE layer of the model?\nAnswer:", + " 128" + ] + ], + "resps": [ + [ + [ + -5.806491374969482, + false + ] + ], + [ + [ + -6.29629373550415, + false + ] + ], + [ + [ + -11.010457992553711, + false + ] + ], + [ + [ + -7.832982540130615, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.806491374969482, + false + ], + [ + -6.29629373550415, + false + ], + [ + -11.010457992553711, + false + ], + [ + -7.832982540130615, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "64fcd7f21e4f146b99315a2ab8ffecfccbe38f757e4e2e4e741d44cc318cedff", + "prompt_hash": "79143893a7c5b0703ecf4631e7bc9aec135cd93570015c66ad796a6643c95c2a", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 13, + "doc": { + "question": "What type of loss is used to prevent routing collapse during training?", + "choices": [ + "Cross-entropy loss", + "Z-loss", + "Load balancing loss", + "Triplet loss" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What type of loss is used to prevent routing collapse during training?\nAnswer:", + " Cross-entropy loss" + ], + [ + "Question:What type of loss is used to prevent routing collapse during training?\nAnswer:", + " Z-loss" + ], + [ + "Question:What type of loss is used to prevent routing collapse during training?\nAnswer:", + " Load balancing loss" + ], + [ + "Question:What type of loss is used to prevent routing collapse during training?\nAnswer:", + " Triplet loss" + ] + ], + "resps": [ + [ + [ + -7.436159610748291, + false + ] + ], + [ + [ + -12.510422706604004, + false + ] + ], + [ + [ + -13.109281539916992, + false + ] + ], + [ + [ + -12.308365821838379, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.436159610748291, + false + ], + [ + -12.510422706604004, + false + ], + [ + -13.109281539916992, + false + ], + [ + -12.308365821838379, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d4376c2623b39f4a9dba9b0a13d6e872a8b5737b8bd05087639cc083c35eceb8", + "prompt_hash": "4f6cd51e4cda9517a2579b60770da74323879616ff0c76869abad7f2a69feaa2", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 14, + "doc": { + "question": "How many tokens are used in the final multimodal post-training stage?", + "choices": [ + "12B", + "20B", + "33B", + "400B" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many tokens are used in the final multimodal post-training stage?\nAnswer:", + " 12B" + ], + [ + "Question:How many tokens are used in the final multimodal post-training stage?\nAnswer:", + " 20B" + ], + [ + "Question:How many tokens are used in the final multimodal post-training stage?\nAnswer:", + " 33B" + ], + [ + "Question:How many tokens are used in the final multimodal post-training stage?\nAnswer:", + " 400B" + ] + ], + "resps": [ + [ + [ + -11.14369010925293, + false + ] + ], + [ + [ + -11.791297912597656, + false + ] + ], + [ + [ + -12.262496948242188, + false + ] + ], + [ + [ + -13.315629959106445, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.14369010925293, + false + ], + [ + -11.791297912597656, + false + ], + [ + -12.262496948242188, + false + ], + [ + -13.315629959106445, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0825e1d5fb8ee7c6bd644f052a2488cb7be215fb467757bfa2555f98a47c09d0", + "prompt_hash": "440baadf0422a6fba1ef6e8c9f6d53343c33a81f59427b90fdbeee2a2a507e87", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 15, + "doc": { + "question": "What are the four ability modules defined in the evaluation interface for embodied agents?", + "choices": [ + "Goal Interpretation, Subgoal Decomposition, Action Sequencing, Transition Modeling", + "Goal Interpretation, Motion Control, Planning, Feedback Analysis", + "Goal Setting, Path Planning, Execution, Memory Encoding", + "Object Recognition, Scene Understanding, Action Prediction, Feedback Control" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What are the four ability modules defined in the evaluation interface for embodied agents?\nAnswer:", + " Goal Interpretation, Subgoal Decomposition, Action Sequencing, Transition Modeling" + ], + [ + "Question:What are the four ability modules defined in the evaluation interface for embodied agents?\nAnswer:", + " Goal Interpretation, Motion Control, Planning, Feedback Analysis" + ], + [ + "Question:What are the four ability modules defined in the evaluation interface for embodied agents?\nAnswer:", + " Goal Setting, Path Planning, Execution, Memory Encoding" + ], + [ + "Question:What are the four ability modules defined in the evaluation interface for embodied agents?\nAnswer:", + " Object Recognition, Scene Understanding, Action Prediction, Feedback Control" + ] + ], + "resps": [ + [ + [ + -52.57338333129883, + false + ] + ], + [ + [ + -45.967445373535156, + false + ] + ], + [ + [ + -43.07550811767578, + false + ] + ], + [ + [ + -40.704917907714844, + false + ] + ] + ], + "filtered_resps": [ + [ + -52.57338333129883, + false + ], + [ + -45.967445373535156, + false + ], + [ + -43.07550811767578, + false + ], + [ + -40.704917907714844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "446907946e6be1ea3eddc9b9023aa680aab7631bc23164a68c1f2ceaae9ba2ab", + "prompt_hash": "c6e08ed15bc7ade20774b38a2bd6126aa9628961781f8bf1134f6e6a9efa8b2c", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 16, + "doc": { + "question": "Which logic formalism is used to unify goal representation and interface communication across modules?", + "choices": [ + "First-order logic", + "Linear Temporal Logic (LTL)", + "Propositional Logic", + "Modal Logic" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which logic formalism is used to unify goal representation and interface communication across modules?\nAnswer:", + " First-order logic" + ], + [ + "Question:Which logic formalism is used to unify goal representation and interface communication across modules?\nAnswer:", + " Linear Temporal Logic (LTL)" + ], + [ + "Question:Which logic formalism is used to unify goal representation and interface communication across modules?\nAnswer:", + " Propositional Logic" + ], + [ + "Question:Which logic formalism is used to unify goal representation and interface communication across modules?\nAnswer:", + " Modal Logic" + ] + ], + "resps": [ + [ + [ + -9.375009536743164, + false + ] + ], + [ + [ + -10.933296203613281, + false + ] + ], + [ + [ + -8.579730987548828, + false + ] + ], + [ + [ + -11.018582344055176, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.375009536743164, + false + ], + [ + -10.933296203613281, + false + ], + [ + -8.579730987548828, + false + ], + [ + -11.018582344055176, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a34d1de6c62668e56480d62ad4a4d294667162537293a265b2a23149e6cd5b3c", + "prompt_hash": "fbc065b0157fa4517c361a13aad431e8683795a12f276ad4162de1750f148b19", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 17, + "doc": { + "question": "Which operator in the LTL extension replaces traditional 'Next' and 'Eventually' operators?", + "choices": [ + "until", + "followed_by", + "then", + "after" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which operator in the LTL extension replaces traditional 'Next' and 'Eventually' operators?\nAnswer:", + " until" + ], + [ + "Question:Which operator in the LTL extension replaces traditional 'Next' and 'Eventually' operators?\nAnswer:", + " followed_by" + ], + [ + "Question:Which operator in the LTL extension replaces traditional 'Next' and 'Eventually' operators?\nAnswer:", + " then" + ], + [ + "Question:Which operator in the LTL extension replaces traditional 'Next' and 'Eventually' operators?\nAnswer:", + " after" + ] + ], + "resps": [ + [ + [ + -8.634527206420898, + false + ] + ], + [ + [ + -16.37803840637207, + false + ] + ], + [ + [ + -8.34276008605957, + false + ] + ], + [ + [ + -8.659734725952148, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.634527206420898, + false + ], + [ + -16.37803840637207, + false + ], + [ + -8.34276008605957, + false + ], + [ + -8.659734725952148, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b853127ae9c576f6be3da574057481ba1e3eac84d279279889e5df6bf8bf2088", + "prompt_hash": "8ee4cf7b73048e86482c23135a7db52e17051368b55de95740771faa55e50b7c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 18, + "doc": { + "question": "What are the three main categories used to evaluate Goal Interpretation?", + "choices": [ + "Grammar Accuracy, Object Mapping, Action Simulation", + "State Goals, Relation Goals, Action Goals", + "Visual Matching, Semantic Parsing, Execution Accuracy", + "Input Coherence, Goal Satisfaction, Subgoal Length" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What are the three main categories used to evaluate Goal Interpretation?\nAnswer:", + " Grammar Accuracy, Object Mapping, Action Simulation" + ], + [ + "Question:What are the three main categories used to evaluate Goal Interpretation?\nAnswer:", + " State Goals, Relation Goals, Action Goals" + ], + [ + "Question:What are the three main categories used to evaluate Goal Interpretation?\nAnswer:", + " Visual Matching, Semantic Parsing, Execution Accuracy" + ], + [ + "Question:What are the three main categories used to evaluate Goal Interpretation?\nAnswer:", + " Input Coherence, Goal Satisfaction, Subgoal Length" + ] + ], + "resps": [ + [ + [ + -46.37894821166992, + false + ] + ], + [ + [ + -30.310394287109375, + false + ] + ], + [ + [ + -46.54631423950195, + false + ] + ], + [ + [ + -45.22802734375, + false + ] + ] + ], + "filtered_resps": [ + [ + -46.37894821166992, + false + ], + [ + -30.310394287109375, + false + ], + [ + -46.54631423950195, + false + ], + [ + -45.22802734375, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e39038c99ac3dba99213e032531639c25f9198784fa744f6528f32358acaf39e", + "prompt_hash": "65380e72e4caa365a491095aaae576924e67da98e0d80326c8e24bce0b8c3f8c", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 19, + "doc": { + "question": "Which type of error occurs when an action is correct in format but involves an object or action not present in the environment?", + "choices": [ + "Parsing Error", + "Affordance Error", + "Hallucination Error", + "Missing Step" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which type of error occurs when an action is correct in format but involves an object or action not present in the environment?\nAnswer:", + " Parsing Error" + ], + [ + "Question:Which type of error occurs when an action is correct in format but involves an object or action not present in the environment?\nAnswer:", + " Affordance Error" + ], + [ + "Question:Which type of error occurs when an action is correct in format but involves an object or action not present in the environment?\nAnswer:", + " Hallucination Error" + ], + [ + "Question:Which type of error occurs when an action is correct in format but involves an object or action not present in the environment?\nAnswer:", + " Missing Step" + ] + ], + "resps": [ + [ + [ + -11.865713119506836, + false + ] + ], + [ + [ + -12.490418434143066, + false + ] + ], + [ + [ + -12.585088729858398, + false + ] + ], + [ + [ + -15.464625358581543, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.865713119506836, + false + ], + [ + -12.490418434143066, + false + ], + [ + -12.585088729858398, + false + ], + [ + -15.464625358581543, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3fabb95c875a224c6b8a5289f4928512d9403c0f1ade52711fe13600462d2cd1", + "prompt_hash": "cc058b02a36c03400b43d613b77eec781e329a7dc4c8e0e5588cf7397820cf16", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 20, + "doc": { + "question": "What is the function of the model referred to as G_RM?", + "choices": [ + "To generate completions based on user prompts", + "To provide safety specifications for model training", + "To judge completions for correctness, helpfulness, and compliance", + "To encode the chain-of-thought during inference" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the function of the model referred to as G_RM?\nAnswer:", + " To generate completions based on user prompts" + ], + [ + "Question:What is the function of the model referred to as G_RM?\nAnswer:", + " To provide safety specifications for model training" + ], + [ + "Question:What is the function of the model referred to as G_RM?\nAnswer:", + " To judge completions for correctness, helpfulness, and compliance" + ], + [ + "Question:What is the function of the model referred to as G_RM?\nAnswer:", + " To encode the chain-of-thought during inference" + ] + ], + "resps": [ + [ + [ + -28.619800567626953, + false + ] + ], + [ + [ + -35.57646560668945, + false + ] + ], + [ + [ + -51.08036804199219, + false + ] + ], + [ + [ + -38.449432373046875, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.619800567626953, + false + ], + [ + -35.57646560668945, + false + ], + [ + -51.08036804199219, + false + ], + [ + -38.449432373046875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "76cc915c4efc02cf6d849bb601b3d2006b33d12d3b4f4892d0c94aa43042494c", + "prompt_hash": "e15cc7d6ef7856d5aa45d2fc0c5805fcab80ecf2b6ffc22d82642c782e849a7d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 21, + "doc": { + "question": "What are the three labels used to classify user requests based on content policy?", + "choices": [ + "Accept, Reject, Moderate", + "Allow, Deny, Defer", + "Allowed, Disallowed, Requires Safe Completion", + "Safe, Unsafe, Context-Dependent" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What are the three labels used to classify user requests based on content policy?\nAnswer:", + " Accept, Reject, Moderate" + ], + [ + "Question:What are the three labels used to classify user requests based on content policy?\nAnswer:", + " Allow, Deny, Defer" + ], + [ + "Question:What are the three labels used to classify user requests based on content policy?\nAnswer:", + " Allowed, Disallowed, Requires Safe Completion" + ], + [ + "Question:What are the three labels used to classify user requests based on content policy?\nAnswer:", + " Safe, Unsafe, Context-Dependent" + ] + ], + "resps": [ + [ + [ + -17.829608917236328, + false + ] + ], + [ + [ + -20.263399124145508, + false + ] + ], + [ + [ + -43.32582473754883, + false + ] + ], + [ + [ + -25.447376251220703, + false + ] + ] + ], + "filtered_resps": [ + [ + -17.829608917236328, + false + ], + [ + -20.263399124145508, + false + ], + [ + -43.32582473754883, + false + ], + [ + -25.447376251220703, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "756aef5e8cc9ae0ae256cd9578807a8e0142956531fd021cbde0219bca1220be", + "prompt_hash": "7021a82e6a2233212c747912d12da88d433e42f877665a22d6da5907bc0825d1", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 22, + "doc": { + "question": "Which safety category allows educational discussions about suicide but requires a safe-completion for ideation or methods?", + "choices": [ + "Extremism", + "Self-harm", + "Harassment", + "Violence" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which safety category allows educational discussions about suicide but requires a safe-completion for ideation or methods?\nAnswer:", + " Extremism" + ], + [ + "Question:Which safety category allows educational discussions about suicide but requires a safe-completion for ideation or methods?\nAnswer:", + " Self-harm" + ], + [ + "Question:Which safety category allows educational discussions about suicide but requires a safe-completion for ideation or methods?\nAnswer:", + " Harassment" + ], + [ + "Question:Which safety category allows educational discussions about suicide but requires a safe-completion for ideation or methods?\nAnswer:", + " Violence" + ] + ], + "resps": [ + [ + [ + -11.61020565032959, + false + ] + ], + [ + [ + -9.390151023864746, + false + ] + ], + [ + [ + -10.084878921508789, + false + ] + ], + [ + [ + -7.957781791687012, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.61020565032959, + false + ], + [ + -9.390151023864746, + false + ], + [ + -10.084878921508789, + false + ], + [ + -7.957781791687012, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8e13a825edb967f944b382519cbe119e167c8083076f40e27ed48a979af5f556", + "prompt_hash": "2760df879ddc37afd924dd9c38a63395dec97002b56fb0f3bbd691751e7d282a", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 23, + "doc": { + "question": "What does the refusal style guideline explicitly prohibit?", + "choices": [ + "Using abstract references to requests", + "Mentioning specific criminal acts", + "Encouraging users to consult professionals", + "Providing brief apologies" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What does the refusal style guideline explicitly prohibit?\nAnswer:", + " Using abstract references to requests" + ], + [ + "Question:What does the refusal style guideline explicitly prohibit?\nAnswer:", + " Mentioning specific criminal acts" + ], + [ + "Question:What does the refusal style guideline explicitly prohibit?\nAnswer:", + " Encouraging users to consult professionals" + ], + [ + "Question:What does the refusal style guideline explicitly prohibit?\nAnswer:", + " Providing brief apologies" + ] + ], + "resps": [ + [ + [ + -31.08670425415039, + false + ] + ], + [ + [ + -24.397045135498047, + false + ] + ], + [ + [ + -29.132076263427734, + false + ] + ], + [ + [ + -22.21463394165039, + false + ] + ] + ], + "filtered_resps": [ + [ + -31.08670425415039, + false + ], + [ + -24.397045135498047, + false + ], + [ + -29.132076263427734, + false + ], + [ + -22.21463394165039, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6fb9927c8cc4285578785a6dd33dd9c2aa08ec5fd0ba60b17966e1321d6caa0b", + "prompt_hash": "5e1e243af4670b0db136a0863bb51d6893576e2ab2d2eba42416bf75235420a2", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 24, + "doc": { + "question": "What type of training information is excluded from prompts during SFT to encourage model recall of safety policies?", + "choices": [ + "Chain-of-thought references", + "Category-specific specifications", + "Final answer text", + "Safety categories" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What type of training information is excluded from prompts during SFT to encourage model recall of safety policies?\nAnswer:", + " Chain-of-thought references" + ], + [ + "Question:What type of training information is excluded from prompts during SFT to encourage model recall of safety policies?\nAnswer:", + " Category-specific specifications" + ], + [ + "Question:What type of training information is excluded from prompts during SFT to encourage model recall of safety policies?\nAnswer:", + " Final answer text" + ], + [ + "Question:What type of training information is excluded from prompts during SFT to encourage model recall of safety policies?\nAnswer:", + " Safety categories" + ] + ], + "resps": [ + [ + [ + -19.58326530456543, + false + ] + ], + [ + [ + -21.516347885131836, + false + ] + ], + [ + [ + -22.331558227539062, + false + ] + ], + [ + [ + -15.906003952026367, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.58326530456543, + false + ], + [ + -21.516347885131836, + false + ], + [ + -22.331558227539062, + false + ], + [ + -15.906003952026367, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4120363f1bef97c9f0e1c4e575fd00747e559372e35a5088d1b0c464f076421a", + "prompt_hash": "cce0d69851e1b0c4a24df4edfaace2832a3fedfaf5d1bca4cdf8e8849f4c503e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 25, + "doc": { + "question": "What symbol does Search-o1 use to mark the beginning of a search query?", + "choices": [ + "<|begin_search_query|>", + "<|start_search_query|>", + "<|search_query_begin|>", + "<|open_search_query|>" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What symbol does Search-o1 use to mark the beginning of a search query?\nAnswer:", + " <|begin_search_query|>" + ], + [ + "Question:What symbol does Search-o1 use to mark the beginning of a search query?\nAnswer:", + " <|start_search_query|>" + ], + [ + "Question:What symbol does Search-o1 use to mark the beginning of a search query?\nAnswer:", + " <|search_query_begin|>" + ], + [ + "Question:What symbol does Search-o1 use to mark the beginning of a search query?\nAnswer:", + " <|open_search_query|>" + ] + ], + "resps": [ + [ + [ + -23.955759048461914, + false + ] + ], + [ + [ + -23.02724838256836, + false + ] + ], + [ + [ + -24.06060791015625, + false + ] + ], + [ + [ + -28.29389190673828, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.955759048461914, + false + ], + [ + -23.02724838256836, + false + ], + [ + -24.06060791015625, + false + ], + [ + -28.29389190673828, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "266ffc54ecf6947b9dbe06a2643ec3a9a78751d0b06809c042ca50fa25ebdd80", + "prompt_hash": "bcc7f84f3b6b3f32231f935019674ec6e3e6e8383565d32e5385bb91eb20ed20", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 26, + "doc": { + "question": "Which API does Search-o1 use to fetch web content for given URLs?", + "choices": [ + "Jina Reader API", + "OpenAI Retrieval API", + "Bing Content API", + "Doc2Text API" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which API does Search-o1 use to fetch web content for given URLs?\nAnswer:", + " Jina Reader API" + ], + [ + "Question:Which API does Search-o1 use to fetch web content for given URLs?\nAnswer:", + " OpenAI Retrieval API" + ], + [ + "Question:Which API does Search-o1 use to fetch web content for given URLs?\nAnswer:", + " Bing Content API" + ], + [ + "Question:Which API does Search-o1 use to fetch web content for given URLs?\nAnswer:", + " Doc2Text API" + ] + ], + "resps": [ + [ + [ + -31.21390724182129, + false + ] + ], + [ + [ + -21.821231842041016, + false + ] + ], + [ + [ + -15.548657417297363, + false + ] + ], + [ + [ + -21.298263549804688, + false + ] + ] + ], + "filtered_resps": [ + [ + -31.21390724182129, + false + ], + [ + -21.821231842041016, + false + ], + [ + -15.548657417297363, + false + ], + [ + -21.298263549804688, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "12b97b0d53975bf2ea01b8de1d4f72c94139f2d040997485fc5b631ae309bdc4", + "prompt_hash": "cd7f2ae9f1e2f4ed786f45762308172a6b4c6791d7fcc0971b432ac17897e6cb", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 27, + "doc": { + "question": "What is the purpose of the Reason-in-Documents module in the Search-o1 framework?", + "choices": [ + "To analyze retrieved documents and extract relevant information for reasoning", + "To generate random search queries when needed", + "To replace the original reasoning model with a distilled one", + "To summarize the question before answering" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the purpose of the Reason-in-Documents module in the Search-o1 framework?\nAnswer:", + " To analyze retrieved documents and extract relevant information for reasoning" + ], + [ + "Question:What is the purpose of the Reason-in-Documents module in the Search-o1 framework?\nAnswer:", + " To generate random search queries when needed" + ], + [ + "Question:What is the purpose of the Reason-in-Documents module in the Search-o1 framework?\nAnswer:", + " To replace the original reasoning model with a distilled one" + ], + [ + "Question:What is the purpose of the Reason-in-Documents module in the Search-o1 framework?\nAnswer:", + " To summarize the question before answering" + ] + ], + "resps": [ + [ + [ + -25.887866973876953, + false + ] + ], + [ + [ + -29.12590980529785, + false + ] + ], + [ + [ + -37.679595947265625, + false + ] + ], + [ + [ + -23.72551155090332, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.887866973876953, + false + ], + [ + -29.12590980529785, + false + ], + [ + -37.679595947265625, + false + ], + [ + -23.72551155090332, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "52afe4206ea3aa17a25d1a11ac1089b750f7309cae24882d6ed7eb1a59c03fdd", + "prompt_hash": "7a2712b8c180e61fc344ce7e93ff1255af68261f77c18a4d32a2b2d43729afec", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 28, + "doc": { + "question": "What does the GPQA dataset evaluate?", + "choices": [ + "PhD-level science multiple-choice QA", + "Elementary arithmetic reasoning", + "General commonsense QA", + "Code generation tasks" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What does the GPQA dataset evaluate?\nAnswer:", + " PhD-level science multiple-choice QA" + ], + [ + "Question:What does the GPQA dataset evaluate?\nAnswer:", + " Elementary arithmetic reasoning" + ], + [ + "Question:What does the GPQA dataset evaluate?\nAnswer:", + " General commonsense QA" + ], + [ + "Question:What does the GPQA dataset evaluate?\nAnswer:", + " Code generation tasks" + ] + ], + "resps": [ + [ + [ + -38.47972869873047, + false + ] + ], + [ + [ + -19.329471588134766, + false + ] + ], + [ + [ + -19.353612899780273, + false + ] + ], + [ + [ + -13.192042350769043, + false + ] + ] + ], + "filtered_resps": [ + [ + -38.47972869873047, + false + ], + [ + -19.329471588134766, + false + ], + [ + -19.353612899780273, + false + ], + [ + -13.192042350769043, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "561a101e49c5d7e509296df702d6c922e782fa6fbdbdde6a232e1c11981f122e", + "prompt_hash": "947b58fbbac3e915e72713db7825fd176a0104e4b240ab4c19f1a9f4a3312ebb", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 29, + "doc": { + "question": "What function is triggered when a new search query is detected in Search-o1?", + "choices": [ + "Search", + "RetrieveDocs", + "FetchKnowledge", + "ExternalCall" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What function is triggered when a new search query is detected in Search-o1?\nAnswer:", + " Search" + ], + [ + "Question:What function is triggered when a new search query is detected in Search-o1?\nAnswer:", + " RetrieveDocs" + ], + [ + "Question:What function is triggered when a new search query is detected in Search-o1?\nAnswer:", + " FetchKnowledge" + ], + [ + "Question:What function is triggered when a new search query is detected in Search-o1?\nAnswer:", + " ExternalCall" + ] + ], + "resps": [ + [ + [ + -3.8662726879119873, + false + ] + ], + [ + [ + -16.96810531616211, + false + ] + ], + [ + [ + -16.960493087768555, + false + ] + ], + [ + [ + -18.54381561279297, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.8662726879119873, + false + ], + [ + -16.96810531616211, + false + ], + [ + -16.960493087768555, + false + ], + [ + -18.54381561279297, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a24d136e0962c5f560ab2ebf59f1bababedb02eb2e3cc801d2c3ea0679e0d3bb", + "prompt_hash": "d422059119c74c5c9293ac200c32bcfd09919427a52e44b1e1a52b27aaa55aee", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 30, + "doc": { + "question": "What is the main computational advantage of Mamba over Transformer models?", + "choices": [ + "It uses quadratic complexity with respect to sequence length.", + "It requires less training data.", + "It maintains linear computational complexity with respect to sequence length.", + "It ignores spatial features to save computation." + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the main computational advantage of Mamba over Transformer models?\nAnswer:", + " It uses quadratic complexity with respect to sequence length." + ], + [ + "Question:What is the main computational advantage of Mamba over Transformer models?\nAnswer:", + " It requires less training data." + ], + [ + "Question:What is the main computational advantage of Mamba over Transformer models?\nAnswer:", + " It maintains linear computational complexity with respect to sequence length." + ], + [ + "Question:What is the main computational advantage of Mamba over Transformer models?\nAnswer:", + " It ignores spatial features to save computation." + ] + ], + "resps": [ + [ + [ + -28.060897827148438, + false + ] + ], + [ + [ + -19.728771209716797, + false + ] + ], + [ + [ + -23.50119400024414, + false + ] + ], + [ + [ + -40.33341979980469, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.060897827148438, + false + ], + [ + -19.728771209716797, + false + ], + [ + -23.50119400024414, + false + ], + [ + -40.33341979980469, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d14123ec3f8bb1e10aa06ff5803bfe8a8f2865a7e274a4e0332261e5e4f2c5f0", + "prompt_hash": "4b86acb7e70681cd4346ecce9c8ae6978cca29f5f4293186c1c15a10a933061e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 31, + "doc": { + "question": "What are the three core components of the MambaHSI encoder?", + "choices": [ + "Convolutional block, spectral transformer, normalization layer", + "Spectral Mamba block, residual learning block, loss function", + "Spatial Mamba block, spectral Mamba block, spatial-spectral fusion module", + "Embedding layer, classification head, dropout block" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What are the three core components of the MambaHSI encoder?\nAnswer:", + " Convolutional block, spectral transformer, normalization layer" + ], + [ + "Question:What are the three core components of the MambaHSI encoder?\nAnswer:", + " Spectral Mamba block, residual learning block, loss function" + ], + [ + "Question:What are the three core components of the MambaHSI encoder?\nAnswer:", + " Spatial Mamba block, spectral Mamba block, spatial-spectral fusion module" + ], + [ + "Question:What are the three core components of the MambaHSI encoder?\nAnswer:", + " Embedding layer, classification head, dropout block" + ] + ], + "resps": [ + [ + [ + -34.9793701171875, + false + ] + ], + [ + [ + -41.64816665649414, + false + ] + ], + [ + [ + -31.879371643066406, + false + ] + ], + [ + [ + -37.53800964355469, + false + ] + ] + ], + "filtered_resps": [ + [ + -34.9793701171875, + false + ], + [ + -41.64816665649414, + false + ], + [ + -31.879371643066406, + false + ], + [ + -37.53800964355469, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4a50923d09b4ff0d230771ad164546c156529be98d7cc504b58f53010ce32086", + "prompt_hash": "bb33e867da7f83f6e54e7e6ab57cbcc4a685b9497075b9f6bb67fc487c7f7af6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 32, + "doc": { + "question": "What is the purpose of the spatial-spectral fusion module in MambaHSI?", + "choices": [ + "To apply dropout regularization across the spectral bands", + "To reduce spatial resolution for faster inference", + "To adaptively fuse the spatial and spectral information based on their importance", + "To convert hyperspectral images into RGB for classification" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the purpose of the spatial-spectral fusion module in MambaHSI?\nAnswer:", + " To apply dropout regularization across the spectral bands" + ], + [ + "Question:What is the purpose of the spatial-spectral fusion module in MambaHSI?\nAnswer:", + " To reduce spatial resolution for faster inference" + ], + [ + "Question:What is the purpose of the spatial-spectral fusion module in MambaHSI?\nAnswer:", + " To adaptively fuse the spatial and spectral information based on their importance" + ], + [ + "Question:What is the purpose of the spatial-spectral fusion module in MambaHSI?\nAnswer:", + " To convert hyperspectral images into RGB for classification" + ] + ], + "resps": [ + [ + [ + -39.88627624511719, + false + ] + ], + [ + [ + -23.449623107910156, + false + ] + ], + [ + [ + -24.270509719848633, + false + ] + ], + [ + [ + -33.8906135559082, + false + ] + ] + ], + "filtered_resps": [ + [ + -39.88627624511719, + false + ], + [ + -23.449623107910156, + false + ], + [ + -24.270509719848633, + false + ], + [ + -33.8906135559082, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d133d375fb6230f9b0602ad1df59198919b09e3b6d3fc6d82ce0c93e04cfc676", + "prompt_hash": "f9b6db19174384edcfb225554ab41f5a024edbaf13203dc62a2cb366c55f82c4", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 33, + "doc": { + "question": "Which dataset used in the experiments was acquired using a Headwall Nano-Hyperspec imaging sensor on a DJI Matrice 600 Pro UAV?", + "choices": [ + "Pavia University", + "Houston", + "WHU-Hi-HanChuan", + "WHU-Hi-HongHu" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:Which dataset used in the experiments was acquired using a Headwall Nano-Hyperspec imaging sensor on a DJI Matrice 600 Pro UAV?\nAnswer:", + " Pavia University" + ], + [ + "Question:Which dataset used in the experiments was acquired using a Headwall Nano-Hyperspec imaging sensor on a DJI Matrice 600 Pro UAV?\nAnswer:", + " Houston" + ], + [ + "Question:Which dataset used in the experiments was acquired using a Headwall Nano-Hyperspec imaging sensor on a DJI Matrice 600 Pro UAV?\nAnswer:", + " WHU-Hi-HanChuan" + ], + [ + "Question:Which dataset used in the experiments was acquired using a Headwall Nano-Hyperspec imaging sensor on a DJI Matrice 600 Pro UAV?\nAnswer:", + " WHU-Hi-HongHu" + ] + ], + "resps": [ + [ + [ + -9.408138275146484, + false + ] + ], + [ + [ + -10.154462814331055, + false + ] + ], + [ + [ + -36.05073547363281, + false + ] + ], + [ + [ + -41.54401397705078, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.408138275146484, + false + ], + [ + -10.154462814331055, + false + ], + [ + -36.05073547363281, + false + ], + [ + -41.54401397705078, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e2885bbce1d81fa6aab9a9a80d3d189bde323125c6c02015d225850009820d4c", + "prompt_hash": "946ca21eb0121687533e64a30df6b4138c9f817ed48c0d5c0f1d846ada337424", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 34, + "doc": { + "question": "How many classes are included in the Houston hyperspectral dataset used in the experiments?", + "choices": [ + "9", + "16", + "22", + "14" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many classes are included in the Houston hyperspectral dataset used in the experiments?\nAnswer:", + " 9" + ], + [ + "Question:How many classes are included in the Houston hyperspectral dataset used in the experiments?\nAnswer:", + " 16" + ], + [ + "Question:How many classes are included in the Houston hyperspectral dataset used in the experiments?\nAnswer:", + " 22" + ], + [ + "Question:How many classes are included in the Houston hyperspectral dataset used in the experiments?\nAnswer:", + " 14" + ] + ], + "resps": [ + [ + [ + -4.74687385559082, + false + ] + ], + [ + [ + -5.32486629486084, + false + ] + ], + [ + [ + -5.936315059661865, + false + ] + ], + [ + [ + -5.394577980041504, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.74687385559082, + false + ], + [ + -5.32486629486084, + false + ], + [ + -5.936315059661865, + false + ], + [ + -5.394577980041504, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "1838ae62bb0ff5951b7b7f3c1eba72e304cdf3ff32578d7c1f9669a1277d8422", + "prompt_hash": "f04139ad7bf7ffeb0eb8aa5ec4a9cdd207bda35e0da604aa82a145eb493c43c6", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 35, + "doc": { + "question": "What is the total area covered by the DES Y6 Gold wide-area survey footprint?", + "choices": [ + "Approximately 5000 square degrees", + "Approximately 7000 square degrees", + "Approximately 10000 square degrees", + "Approximately 3500 square degrees" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the total area covered by the DES Y6 Gold wide-area survey footprint?\nAnswer:", + " Approximately 5000 square degrees" + ], + [ + "Question:What is the total area covered by the DES Y6 Gold wide-area survey footprint?\nAnswer:", + " Approximately 7000 square degrees" + ], + [ + "Question:What is the total area covered by the DES Y6 Gold wide-area survey footprint?\nAnswer:", + " Approximately 10000 square degrees" + ], + [ + "Question:What is the total area covered by the DES Y6 Gold wide-area survey footprint?\nAnswer:", + " Approximately 3500 square degrees" + ] + ], + "resps": [ + [ + [ + -12.233139038085938, + false + ] + ], + [ + [ + -12.975166320800781, + false + ] + ], + [ + [ + -13.492626190185547, + false + ] + ], + [ + [ + -12.901947021484375, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.233139038085938, + false + ], + [ + -12.975166320800781, + false + ], + [ + -13.492626190185547, + false + ], + [ + -12.901947021484375, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "760366bb1c0e2c5d0f1ba382b67ddbcf2a9bee94a3cdf7a390c2168007e141dd", + "prompt_hash": "1276376381c3a6685b55c9cfbb5cfb01cb4b34d3d78df37c762e9d5288d62dd4", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 36, + "doc": { + "question": "Which magnitude correction coefficient is used for the 'g' band when correcting for interstellar extinction in DES Y6 Gold?", + "choices": [ + "3.186", + "2.140", + "1.569", + "1.048" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which magnitude correction coefficient is used for the 'g' band when correcting for interstellar extinction in DES Y6 Gold?\nAnswer:", + " 3.186" + ], + [ + "Question:Which magnitude correction coefficient is used for the 'g' band when correcting for interstellar extinction in DES Y6 Gold?\nAnswer:", + " 2.140" + ], + [ + "Question:Which magnitude correction coefficient is used for the 'g' band when correcting for interstellar extinction in DES Y6 Gold?\nAnswer:", + " 1.569" + ], + [ + "Question:Which magnitude correction coefficient is used for the 'g' band when correcting for interstellar extinction in DES Y6 Gold?\nAnswer:", + " 1.048" + ] + ], + "resps": [ + [ + [ + -12.510847091674805, + false + ] + ], + [ + [ + -12.662933349609375, + false + ] + ], + [ + [ + -12.489203453063965, + false + ] + ], + [ + [ + -11.289423942565918, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.510847091674805, + false + ], + [ + -12.662933349609375, + false + ], + [ + -12.489203453063965, + false + ], + [ + -11.289423942565918, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fb823d5546e6d00cd71d5cf3002eb41d44998ad3f4f9b222a7a5e8bdf61c4b99", + "prompt_hash": "aceccefb48255c4b8fe744cf7e506e79afcac7d34251224bfc68a004861928f5", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 37, + "doc": { + "question": "What modeling method is used for PSF photometry in the DES Y6 Gold \\u0003fitvd algorithm?", + "choices": [ + "Zero-size point-source model", + "Extended Sersic profile with variable index", + "Gaussian Mixture Model with ten components", + "Power-law profile with variable slope" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What modeling method is used for PSF photometry in the DES Y6 Gold \\u0003fitvd algorithm?\nAnswer:", + " Zero-size point-source model" + ], + [ + "Question:What modeling method is used for PSF photometry in the DES Y6 Gold \\u0003fitvd algorithm?\nAnswer:", + " Extended Sersic profile with variable index" + ], + [ + "Question:What modeling method is used for PSF photometry in the DES Y6 Gold \\u0003fitvd algorithm?\nAnswer:", + " Gaussian Mixture Model with ten components" + ], + [ + "Question:What modeling method is used for PSF photometry in the DES Y6 Gold \\u0003fitvd algorithm?\nAnswer:", + " Power-law profile with variable slope" + ] + ], + "resps": [ + [ + [ + -28.418212890625, + false + ] + ], + [ + [ + -34.0587043762207, + false + ] + ], + [ + [ + -24.72544288635254, + false + ] + ], + [ + [ + -25.261842727661133, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.418212890625, + false + ], + [ + -34.0587043762207, + false + ], + [ + -24.72544288635254, + false + ], + [ + -25.261842727661133, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6211b34489eb7b1c5bd5f5bb696bf2a489773e69dd4526077ec1ff7f3b85f80e", + "prompt_hash": "7d8efd9da91b78776afe4f53f803b44181d00ba7eafc228ca3e91925ae02037b", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 38, + "doc": { + "question": "Which component is used as the photometric calibration reference star in DES Y6 Gold?", + "choices": [ + "Hubble Space Telescope CalSpec standard star C26202", + "Gaia DR3 standard star G18392", + "SDSS Stripe 82 reference star SDSSJ014", + "VISTA Hemisphere Survey star VHS001" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which component is used as the photometric calibration reference star in DES Y6 Gold?\nAnswer:", + " Hubble Space Telescope CalSpec standard star C26202" + ], + [ + "Question:Which component is used as the photometric calibration reference star in DES Y6 Gold?\nAnswer:", + " Gaia DR3 standard star G18392" + ], + [ + "Question:Which component is used as the photometric calibration reference star in DES Y6 Gold?\nAnswer:", + " SDSS Stripe 82 reference star SDSSJ014" + ], + [ + "Question:Which component is used as the photometric calibration reference star in DES Y6 Gold?\nAnswer:", + " VISTA Hemisphere Survey star VHS001" + ] + ], + "resps": [ + [ + [ + -53.36168670654297, + false + ] + ], + [ + [ + -27.988567352294922, + false + ] + ], + [ + [ + -37.09762191772461, + false + ] + ], + [ + [ + -36.82593536376953, + false + ] + ] + ], + "filtered_resps": [ + [ + -53.36168670654297, + false + ], + [ + -27.988567352294922, + false + ], + [ + -37.09762191772461, + false + ], + [ + -36.82593536376953, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "aae765a7d4300bc0b98db398dececf7d045e8a52d324494c18b70af0fc524d02", + "prompt_hash": "90fa208f136060df152c4e678c37834c4714a6b2624d6526562b9974edae8487", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 39, + "doc": { + "question": "What is the total number of DECam exposures included in DES DR2 and Y6 Gold?", + "choices": [ + "72,217", + "84,932", + "45,318", + "60,104" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the total number of DECam exposures included in DES DR2 and Y6 Gold?\nAnswer:", + " 72,217" + ], + [ + "Question:What is the total number of DECam exposures included in DES DR2 and Y6 Gold?\nAnswer:", + " 84,932" + ], + [ + "Question:What is the total number of DECam exposures included in DES DR2 and Y6 Gold?\nAnswer:", + " 45,318" + ], + [ + "Question:What is the total number of DECam exposures included in DES DR2 and Y6 Gold?\nAnswer:", + " 60,104" + ] + ], + "resps": [ + [ + [ + -15.384516716003418, + false + ] + ], + [ + [ + -15.420585632324219, + false + ] + ], + [ + [ + -14.602078437805176, + false + ] + ], + [ + [ + -14.678556442260742, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.384516716003418, + false + ], + [ + -15.420585632324219, + false + ], + [ + -14.602078437805176, + false + ], + [ + -14.678556442260742, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4d1f785b545ac35d5410aa223b3aae264bc492e4fd1ee0b3e0d6517afc30f1a2", + "prompt_hash": "1e570ae93663e6215b95ced6fad37af432a866317248a803a19c9f0da6c06190", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 40, + "doc": { + "question": "What is the name of the benchmark proposed to evaluate multimodal multi-step reasoning tasks?", + "choices": [ + "VQA-Chain", + "VLM-Reason", + "VRC-Bench", + "MultiStepQA" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the name of the benchmark proposed to evaluate multimodal multi-step reasoning tasks?\nAnswer:", + " VQA-Chain" + ], + [ + "Question:What is the name of the benchmark proposed to evaluate multimodal multi-step reasoning tasks?\nAnswer:", + " VLM-Reason" + ], + [ + "Question:What is the name of the benchmark proposed to evaluate multimodal multi-step reasoning tasks?\nAnswer:", + " VRC-Bench" + ], + [ + "Question:What is the name of the benchmark proposed to evaluate multimodal multi-step reasoning tasks?\nAnswer:", + " MultiStepQA" + ] + ], + "resps": [ + [ + [ + -16.848678588867188, + false + ] + ], + [ + [ + -13.863715171813965, + false + ] + ], + [ + [ + -17.058414459228516, + false + ] + ], + [ + [ + -6.643422603607178, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.848678588867188, + false + ], + [ + -13.863715171813965, + false + ], + [ + -17.058414459228516, + false + ], + [ + -6.643422603607178, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f49d9074782d55a4cb7470bafb5c6fd059a2e0b5d3debacab64bae5906b61854", + "prompt_hash": "4a5c0bd5a79a3f897b9cb74ebb8744267e39db166fdbb8d41e60a4fa71925a5d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 41, + "doc": { + "question": "How many manually verified reasoning steps are included in VRC-Bench?", + "choices": [ + "3,500", + "4,173", + "5,200", + "3,998" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many manually verified reasoning steps are included in VRC-Bench?\nAnswer:", + " 3,500" + ], + [ + "Question:How many manually verified reasoning steps are included in VRC-Bench?\nAnswer:", + " 4,173" + ], + [ + "Question:How many manually verified reasoning steps are included in VRC-Bench?\nAnswer:", + " 5,200" + ], + [ + "Question:How many manually verified reasoning steps are included in VRC-Bench?\nAnswer:", + " 3,998" + ] + ], + "resps": [ + [ + [ + -12.997629165649414, + false + ] + ], + [ + [ + -15.097672462463379, + false + ] + ], + [ + [ + -13.911409378051758, + false + ] + ], + [ + [ + -15.563308715820312, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.997629165649414, + false + ], + [ + -15.097672462463379, + false + ], + [ + -13.911409378051758, + false + ], + [ + -15.563308715820312, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ab029bc7ad364193651f7eafa1a76ce79ef4169c5ca128f8c50ee4c768f63ed5", + "prompt_hash": "aaf9e4ab60af7e986dec67e442603b09898c067e4dc1a58a24f24fd83579e7f2", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 42, + "doc": { + "question": "What is the base model used for training LlamaV-o1?", + "choices": [ + "GPT-4o-mini", + "Llava-CoT", + "Llama-3.2-11B-Vision-Instruct", + "Gemini-1.5-Pro" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the base model used for training LlamaV-o1?\nAnswer:", + " GPT-4o-mini" + ], + [ + "Question:What is the base model used for training LlamaV-o1?\nAnswer:", + " Llava-CoT" + ], + [ + "Question:What is the base model used for training LlamaV-o1?\nAnswer:", + " Llama-3.2-11B-Vision-Instruct" + ], + [ + "Question:What is the base model used for training LlamaV-o1?\nAnswer:", + " Gemini-1.5-Pro" + ] + ], + "resps": [ + [ + [ + -15.453899383544922, + false + ] + ], + [ + [ + -22.55254554748535, + false + ] + ], + [ + [ + -40.223777770996094, + false + ] + ], + [ + [ + -11.26479434967041, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.453899383544922, + false + ], + [ + -22.55254554748535, + false + ], + [ + -40.223777770996094, + false + ], + [ + -11.26479434967041, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3794d4c9d0dcc848c6bab66c1636d1cd1ce54b5faa82b7d2367ac43f2f2c0e05", + "prompt_hash": "eee8678d456145a2b8c216698eb4fb00373a3903482962a866cb58ababc6b09a", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 43, + "doc": { + "question": "Which metric checks for irrelevant or fabricated reasoning steps?", + "choices": [ + "Informativeness-Step", + "Hallucination", + "Semantic Coverage", + "Reasoning Alignment" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which metric checks for irrelevant or fabricated reasoning steps?\nAnswer:", + " Informativeness-Step" + ], + [ + "Question:Which metric checks for irrelevant or fabricated reasoning steps?\nAnswer:", + " Hallucination" + ], + [ + "Question:Which metric checks for irrelevant or fabricated reasoning steps?\nAnswer:", + " Semantic Coverage" + ], + [ + "Question:Which metric checks for irrelevant or fabricated reasoning steps?\nAnswer:", + " Reasoning Alignment" + ] + ], + "resps": [ + [ + [ + -23.075510025024414, + false + ] + ], + [ + [ + -9.403883934020996, + false + ] + ], + [ + [ + -17.089900970458984, + false + ] + ], + [ + [ + -13.730018615722656, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.075510025024414, + false + ], + [ + -9.403883934020996, + false + ], + [ + -17.089900970458984, + false + ], + [ + -13.730018615722656, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "208558ab3d63cf67ef2b796ee97528ff8bd9af26bb61f291dadc8607768c025e", + "prompt_hash": "38553bafe3fe4bbd2d61b2e2bfc7c60d6803c445de407b64e062033f2c8cbf4c", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 44, + "doc": { + "question": "What type of learning strategy is used to train LlamaV-o1 progressively from simple to complex tasks?", + "choices": [ + "Reinforcement Learning", + "Contrastive Learning", + "Curriculum Learning", + "Unsupervised Pretraining" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What type of learning strategy is used to train LlamaV-o1 progressively from simple to complex tasks?\nAnswer:", + " Reinforcement Learning" + ], + [ + "Question:What type of learning strategy is used to train LlamaV-o1 progressively from simple to complex tasks?\nAnswer:", + " Contrastive Learning" + ], + [ + "Question:What type of learning strategy is used to train LlamaV-o1 progressively from simple to complex tasks?\nAnswer:", + " Curriculum Learning" + ], + [ + "Question:What type of learning strategy is used to train LlamaV-o1 progressively from simple to complex tasks?\nAnswer:", + " Unsupervised Pretraining" + ] + ], + "resps": [ + [ + [ + -4.469661235809326, + false + ] + ], + [ + [ + -9.512903213500977, + false + ] + ], + [ + [ + -9.978853225708008, + false + ] + ], + [ + [ + -17.42882537841797, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.469661235809326, + false + ], + [ + -9.512903213500977, + false + ], + [ + -9.978853225708008, + false + ], + [ + -17.42882537841797, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "cc742e6917db8a54e9e779f37121f0dd4a291bac16e92b3d504ad95b69bb36cd", + "prompt_hash": "83742b8f079b8e3873d047b3d288f0cc351323444b85ee9e14c69e1436287314", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 45, + "doc": { + "question": "Which parameterization type is commonly used for image generation in diffusion models?", + "choices": [ + "A. x₀-prediction", + "B. ε-prediction", + "C. v-prediction", + "D. zₜ-prediction" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which parameterization type is commonly used for image generation in diffusion models?\nAnswer:", + " A. x₀-prediction" + ], + [ + "Question:Which parameterization type is commonly used for image generation in diffusion models?\nAnswer:", + " B. ε-prediction" + ], + [ + "Question:Which parameterization type is commonly used for image generation in diffusion models?\nAnswer:", + " C. v-prediction" + ], + [ + "Question:Which parameterization type is commonly used for image generation in diffusion models?\nAnswer:", + " D. zₜ-prediction" + ] + ], + "resps": [ + [ + [ + -40.44757080078125, + false + ] + ], + [ + [ + -33.222816467285156, + false + ] + ], + [ + [ + -27.685667037963867, + false + ] + ], + [ + [ + -49.870819091796875, + false + ] + ] + ], + "filtered_resps": [ + [ + -40.44757080078125, + false + ], + [ + -33.222816467285156, + false + ], + [ + -27.685667037963867, + false + ], + [ + -49.870819091796875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "bdcb2a5306cac39bfd15c10f8bdfcd3eb0e3235eded767ddec9791d1ead52847", + "prompt_hash": "5c780888e8ff88fecc3944fcbca387d3da2260d994c71a1e74dd526b8fadb85b", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 46, + "doc": { + "question": "How many training samples does Lotus use to achieve its reported performance?", + "choices": [ + "A. 59K", + "B. 1.2M", + "C. 250K", + "D. 62.6M" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:How many training samples does Lotus use to achieve its reported performance?\nAnswer:", + " A. 59K" + ], + [ + "Question:How many training samples does Lotus use to achieve its reported performance?\nAnswer:", + " B. 1.2M" + ], + [ + "Question:How many training samples does Lotus use to achieve its reported performance?\nAnswer:", + " C. 250K" + ], + [ + "Question:How many training samples does Lotus use to achieve its reported performance?\nAnswer:", + " D. 62.6M" + ] + ], + "resps": [ + [ + [ + -20.41183853149414, + false + ] + ], + [ + [ + -24.378026962280273, + false + ] + ], + [ + [ + -22.16771125793457, + false + ] + ], + [ + [ + -32.99885559082031, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.41183853149414, + false + ], + [ + -24.378026962280273, + false + ], + [ + -22.16771125793457, + false + ], + [ + -32.99885559082031, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "88ff9460fbcea5468e805f29dea11ddcf9ef68bf5e10dda39ddea33508d5d380", + "prompt_hash": "844a065d7c38f0c7413afe8111feef8efd06cc9c9d71f6bde03819e7d54b6473", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 47, + "doc": { + "question": "What type of prediction is used in the discriminative version of Lotus (Lotus-D)?", + "choices": [ + "A. Uses ε-prediction with noise", + "B. Uses x₀-prediction without noise", + "C. Uses v-prediction with noise", + "D. Uses latent-space image conditioning only" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What type of prediction is used in the discriminative version of Lotus (Lotus-D)?\nAnswer:", + " A. Uses ε-prediction with noise" + ], + [ + "Question:What type of prediction is used in the discriminative version of Lotus (Lotus-D)?\nAnswer:", + " B. Uses x₀-prediction without noise" + ], + [ + "Question:What type of prediction is used in the discriminative version of Lotus (Lotus-D)?\nAnswer:", + " C. Uses v-prediction with noise" + ], + [ + "Question:What type of prediction is used in the discriminative version of Lotus (Lotus-D)?\nAnswer:", + " D. Uses latent-space image conditioning only" + ] + ], + "resps": [ + [ + [ + -48.533790588378906, + false + ] + ], + [ + [ + -68.318603515625, + false + ] + ], + [ + [ + -48.11347198486328, + false + ] + ], + [ + [ + -51.81948471069336, + false + ] + ] + ], + "filtered_resps": [ + [ + -48.533790588378906, + false + ], + [ + -68.318603515625, + false + ], + [ + -48.11347198486328, + false + ], + [ + -51.81948471069336, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "dd9d203f53f3b8e692b1909ce6796379f9b1791e6dd80cbea978d40fc469027b", + "prompt_hash": "edcd9a17425e90b0a319a6e586bc5db38481b7a415ee9ec0a3551484f4492a30", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 48, + "doc": { + "question": "What is the primary purpose of the detail preserver module in Lotus?", + "choices": [ + "A. To improve memory efficiency during training", + "B. To enable dual-modality input", + "C. To preserve fine-grained details in dense annotations", + "D. To increase the number of diffusion steps" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the primary purpose of the detail preserver module in Lotus?\nAnswer:", + " A. To improve memory efficiency during training" + ], + [ + "Question:What is the primary purpose of the detail preserver module in Lotus?\nAnswer:", + " B. To enable dual-modality input" + ], + [ + "Question:What is the primary purpose of the detail preserver module in Lotus?\nAnswer:", + " C. To preserve fine-grained details in dense annotations" + ], + [ + "Question:What is the primary purpose of the detail preserver module in Lotus?\nAnswer:", + " D. To increase the number of diffusion steps" + ] + ], + "resps": [ + [ + [ + -26.969858169555664, + false + ] + ], + [ + [ + -36.691349029541016, + false + ] + ], + [ + [ + -41.21318435668945, + false + ] + ], + [ + [ + -36.86516571044922, + false + ] + ] + ], + "filtered_resps": [ + [ + -26.969858169555664, + false + ], + [ + -36.691349029541016, + false + ], + [ + -41.21318435668945, + false + ], + [ + -36.86516571044922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4a7ae6d28946cceb57aa90a9c6cc416901faaa37a1b16dcdb7fa7287614dfdfd", + "prompt_hash": "a1d8f4d91b8b49a00e7e4671a1f365b13ae751c0ce0344060659f55ddbd7b4d9", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 49, + "doc": { + "question": "What happens when ε-prediction is used at the initial denoising steps?", + "choices": [ + "A. The model produces less variance and more stable predictions", + "B. The predicted variance is amplified and propagated", + "C. The model skips early steps to reduce variance", + "D. The denoising step converges faster" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What happens when ε-prediction is used at the initial denoising steps?\nAnswer:", + " A. The model produces less variance and more stable predictions" + ], + [ + "Question:What happens when ε-prediction is used at the initial denoising steps?\nAnswer:", + " B. The predicted variance is amplified and propagated" + ], + [ + "Question:What happens when ε-prediction is used at the initial denoising steps?\nAnswer:", + " C. The model skips early steps to reduce variance" + ], + [ + "Question:What happens when ε-prediction is used at the initial denoising steps?\nAnswer:", + " D. The denoising step converges faster" + ] + ], + "resps": [ + [ + [ + -38.00743103027344, + false + ] + ], + [ + [ + -42.99990463256836, + false + ] + ], + [ + [ + -40.484310150146484, + false + ] + ], + [ + [ + -27.896730422973633, + false + ] + ] + ], + "filtered_resps": [ + [ + -38.00743103027344, + false + ], + [ + -42.99990463256836, + false + ], + [ + -40.484310150146484, + false + ], + [ + -27.896730422973633, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "3786df5cda14a24d6fdf6e60196d1064ffee000819bd029a45cc8d463520b3d9", + "prompt_hash": "d0ba1a9675287af2ea58b7c36baa8d8b2ae830046e092fe8f8b835e59119bf37", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 50, + "doc": { + "question": "What is the default maximum tree depth used in the self-evolution rounds for math problem solving?", + "choices": [ + "12", + "16", + "20", + "24" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the default maximum tree depth used in the self-evolution rounds for math problem solving?\nAnswer:", + " 12" + ], + [ + "Question:What is the default maximum tree depth used in the self-evolution rounds for math problem solving?\nAnswer:", + " 16" + ], + [ + "Question:What is the default maximum tree depth used in the self-evolution rounds for math problem solving?\nAnswer:", + " 20" + ], + [ + "Question:What is the default maximum tree depth used in the self-evolution rounds for math problem solving?\nAnswer:", + " 24" + ] + ], + "resps": [ + [ + [ + -6.0869550704956055, + false + ] + ], + [ + [ + -6.953391075134277, + false + ] + ], + [ + [ + -5.210151672363281, + false + ] + ], + [ + [ + -7.983536243438721, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.0869550704956055, + false + ], + [ + -6.953391075134277, + false + ], + [ + -5.210151672363281, + false + ], + [ + -7.983536243438721, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "800c7f0b32cd7ad8b229bc3280e1de23af72e70bb58dfd1a9409a0dad352f3e1", + "prompt_hash": "8e56d42f1452ec9c6b466ea3808d628d7487ca361c30e83e8777908118e35346", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 51, + "doc": { + "question": "What learning rate is used for fine-tuning Qwen models in this study?", + "choices": [ + "5e-6", + "6e-6", + "7e-6", + "1e-5" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What learning rate is used for fine-tuning Qwen models in this study?\nAnswer:", + " 5e-6" + ], + [ + "Question:What learning rate is used for fine-tuning Qwen models in this study?\nAnswer:", + " 6e-6" + ], + [ + "Question:What learning rate is used for fine-tuning Qwen models in this study?\nAnswer:", + " 7e-6" + ], + [ + "Question:What learning rate is used for fine-tuning Qwen models in this study?\nAnswer:", + " 1e-5" + ] + ], + "resps": [ + [ + [ + -7.968042850494385, + false + ] + ], + [ + [ + -10.571170806884766, + false + ] + ], + [ + [ + -10.910438537597656, + false + ] + ], + [ + [ + -6.601953506469727, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.968042850494385, + false + ], + [ + -10.571170806884766, + false + ], + [ + -10.910438537597656, + false + ], + [ + -6.601953506469727, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "54afa871339ebb9b9c3054b084eb9c1363694a2ef93b2d3668a4005217c21c11", + "prompt_hash": "65d5f04256731d866019a2a49ee95c22d0fd40b44af6335b4e9f2d22b0da4791", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 52, + "doc": { + "question": "Which reward model uses pairwise ranking loss in its training?", + "choices": [ + "PPM", + "PQM", + "ORM", + "SLM" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which reward model uses pairwise ranking loss in its training?\nAnswer:", + " PPM" + ], + [ + "Question:Which reward model uses pairwise ranking loss in its training?\nAnswer:", + " PQM" + ], + [ + "Question:Which reward model uses pairwise ranking loss in its training?\nAnswer:", + " ORM" + ], + [ + "Question:Which reward model uses pairwise ranking loss in its training?\nAnswer:", + " SLM" + ] + ], + "resps": [ + [ + [ + -14.611196517944336, + false + ] + ], + [ + [ + -15.033802032470703, + false + ] + ], + [ + [ + -14.595401763916016, + false + ] + ], + [ + [ + -13.409589767456055, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.611196517944336, + false + ], + [ + -15.033802032470703, + false + ], + [ + -14.595401763916016, + false + ], + [ + -13.409589767456055, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "49c8a7815d94169375ffb8484392b18bbdfa3b9405ce36f0aee8fe96010acfb9", + "prompt_hash": "49dde6a7ddf4c265f80016da4ff826649fbb6d58a6a1303de1db86e62f3314ec", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 53, + "doc": { + "question": "What was the total GPU setup for the bootstrap round using DeepSeek-Coder-v2-Instruct?", + "choices": [ + "5 nodes of 4×80GB", + "10 nodes of 8×80GB", + "15 nodes of 4×40GB", + "12 nodes of 8×40GB" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What was the total GPU setup for the bootstrap round using DeepSeek-Coder-v2-Instruct?\nAnswer:", + " 5 nodes of 4×80GB" + ], + [ + "Question:What was the total GPU setup for the bootstrap round using DeepSeek-Coder-v2-Instruct?\nAnswer:", + " 10 nodes of 8×80GB" + ], + [ + "Question:What was the total GPU setup for the bootstrap round using DeepSeek-Coder-v2-Instruct?\nAnswer:", + " 15 nodes of 4×40GB" + ], + [ + "Question:What was the total GPU setup for the bootstrap round using DeepSeek-Coder-v2-Instruct?\nAnswer:", + " 12 nodes of 8×40GB" + ] + ], + "resps": [ + [ + [ + -29.78156089782715, + false + ] + ], + [ + [ + -29.15294075012207, + false + ] + ], + [ + [ + -32.71136474609375, + false + ] + ], + [ + [ + -29.856426239013672, + false + ] + ] + ], + "filtered_resps": [ + [ + -29.78156089782715, + false + ], + [ + -29.15294075012207, + false + ], + [ + -32.71136474609375, + false + ], + [ + -29.856426239013672, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4d6cca403c78982916e8fcbbd83f5e540b683e23b4a63fe462e5d3ecd944277b", + "prompt_hash": "807ec96a1d77362088976d759e387ea45c93d3c3811b14b0f08df4cbe0a729b1", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 54, + "doc": { + "question": "Which benchmark required the highest average number of tokens during inference, as shown in the paper?", + "choices": [ + "AIME 2024", + "MATH", + "Olympiad Bench", + "AMC 2023" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which benchmark required the highest average number of tokens during inference, as shown in the paper?\nAnswer:", + " AIME 2024" + ], + [ + "Question:Which benchmark required the highest average number of tokens during inference, as shown in the paper?\nAnswer:", + " MATH" + ], + [ + "Question:Which benchmark required the highest average number of tokens during inference, as shown in the paper?\nAnswer:", + " Olympiad Bench" + ], + [ + "Question:Which benchmark required the highest average number of tokens during inference, as shown in the paper?\nAnswer:", + " AMC 2023" + ] + ], + "resps": [ + [ + [ + -21.254226684570312, + false + ] + ], + [ + [ + -9.348733901977539, + false + ] + ], + [ + [ + -22.08643341064453, + false + ] + ], + [ + [ + -21.39059829711914, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.254226684570312, + false + ], + [ + -9.348733901977539, + false + ], + [ + -22.08643341064453, + false + ], + [ + -21.39059829711914, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f8bf812cf01b159edc3acbd279319be5e5b3446f4f103ceeb836283b365baab7", + "prompt_hash": "76cb13e84f4c67a72a12d0297e71d697f0ac713ebf7dc305bd7e46b85a436ce0", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 55, + "doc": { + "question": "What benchmark did the paper identify as being mostly solved by models like GPT-4o and Claude?", + "choices": [ + "GSM8k", + "Hendrycks MATH Levels 1-3", + "Omni-MATH", + "HARP" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What benchmark did the paper identify as being mostly solved by models like GPT-4o and Claude?\nAnswer:", + " GSM8k" + ], + [ + "Question:What benchmark did the paper identify as being mostly solved by models like GPT-4o and Claude?\nAnswer:", + " Hendrycks MATH Levels 1-3" + ], + [ + "Question:What benchmark did the paper identify as being mostly solved by models like GPT-4o and Claude?\nAnswer:", + " Omni-MATH" + ], + [ + "Question:What benchmark did the paper identify as being mostly solved by models like GPT-4o and Claude?\nAnswer:", + " HARP" + ] + ], + "resps": [ + [ + [ + -9.384391784667969, + false + ] + ], + [ + [ + -42.24859619140625, + false + ] + ], + [ + [ + -26.777854919433594, + false + ] + ], + [ + [ + -15.270162582397461, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.384391784667969, + false + ], + [ + -42.24859619140625, + false + ], + [ + -26.777854919433594, + false + ], + [ + -15.270162582397461, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b7607a1f58dd0ff61eaad155541032acca0fe44a2832e5f5ec691925e40af001", + "prompt_hash": "fcf8808346cefa46360238ef422ec517fb052d71605a6a782284cf9a13a5b97f", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 56, + "doc": { + "question": "What concept is introduced as a generalization of Chain-of-Thought for modeling latent reasoning?", + "choices": [ + "System 2 CoT", + "Hierarchical Prompting", + "Meta-CoT", + "Reflexive Reasoning" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What concept is introduced as a generalization of Chain-of-Thought for modeling latent reasoning?\nAnswer:", + " System 2 CoT" + ], + [ + "Question:What concept is introduced as a generalization of Chain-of-Thought for modeling latent reasoning?\nAnswer:", + " Hierarchical Prompting" + ], + [ + "Question:What concept is introduced as a generalization of Chain-of-Thought for modeling latent reasoning?\nAnswer:", + " Meta-CoT" + ], + [ + "Question:What concept is introduced as a generalization of Chain-of-Thought for modeling latent reasoning?\nAnswer:", + " Reflexive Reasoning" + ] + ], + "resps": [ + [ + [ + -20.35006332397461, + false + ] + ], + [ + [ + -14.312450408935547, + false + ] + ], + [ + [ + -14.631093978881836, + false + ] + ], + [ + [ + -10.159168243408203, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.35006332397461, + false + ], + [ + -14.312450408935547, + false + ], + [ + -14.631093978881836, + false + ], + [ + -10.159168243408203, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d82135d6cba018718c9aeaaf8d337401289d8e5b8c3acc4289de4105f50dae27", + "prompt_hash": "36c13037c05d4f84c743d6955a6d83065962bef50e3857b2f40727e8f593883d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 57, + "doc": { + "question": "Which theoretical framework does Meta-CoT draw inspiration from?", + "choices": [ + "Dual-process theory from Cognitive Science", + "Theory of Computation", + "Game Theory", + "Information Theory" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which theoretical framework does Meta-CoT draw inspiration from?\nAnswer:", + " Dual-process theory from Cognitive Science" + ], + [ + "Question:Which theoretical framework does Meta-CoT draw inspiration from?\nAnswer:", + " Theory of Computation" + ], + [ + "Question:Which theoretical framework does Meta-CoT draw inspiration from?\nAnswer:", + " Game Theory" + ], + [ + "Question:Which theoretical framework does Meta-CoT draw inspiration from?\nAnswer:", + " Information Theory" + ] + ], + "resps": [ + [ + [ + -22.927265167236328, + false + ] + ], + [ + [ + -12.723213195800781, + false + ] + ], + [ + [ + -14.483650207519531, + false + ] + ], + [ + [ + -11.9994478225708, + false + ] + ] + ], + "filtered_resps": [ + [ + -22.927265167236328, + false + ], + [ + -12.723213195800781, + false + ], + [ + -14.483650207519531, + false + ], + [ + -11.9994478225708, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "22c72c05e6d3068fa31abb6d36600f1f8a22f6bba52be443b92256419c2368eb", + "prompt_hash": "75cd67bb7118c915a89d85a02cfb0fbd36a8cc06fd89f0da96b1a4531da8cfad", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 58, + "doc": { + "question": "What is the reward value assigned to a correct solution in the MDP formulation described in the paper?", + "choices": [ + "-1", + "0", + "0.5", + "1" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the reward value assigned to a correct solution in the MDP formulation described in the paper?\nAnswer:", + " -1" + ], + [ + "Question:What is the reward value assigned to a correct solution in the MDP formulation described in the paper?\nAnswer:", + " 0" + ], + [ + "Question:What is the reward value assigned to a correct solution in the MDP formulation described in the paper?\nAnswer:", + " 0.5" + ], + [ + "Question:What is the reward value assigned to a correct solution in the MDP formulation described in the paper?\nAnswer:", + " 1" + ] + ], + "resps": [ + [ + [ + -6.114413261413574, + false + ] + ], + [ + [ + -3.182579517364502, + false + ] + ], + [ + [ + -6.184215068817139, + false + ] + ], + [ + [ + -2.7441458702087402, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.114413261413574, + false + ], + [ + -3.182579517364502, + false + ], + [ + -6.184215068817139, + false + ], + [ + -2.7441458702087402, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "daccd138152fa3660e17d99531ed57ab16f78931bdd8ab5f9f484da037d258db", + "prompt_hash": "c73c924fbb4119640eacaaf7fc282e21f239c938ad8bd82d524b66bd340efd24", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 59, + "doc": { + "question": "According to the scaling law hypothesis in the paper, what three components govern performance on complex reasoning tasks?", + "choices": [ + "Model size, dataset quality, and pretraining loss", + "Training data, inference-time compute, and instruction tuning", + "Model size, training data (compute), and inference-time compute", + "Prompt quality, alignment tuning, and data diversity" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:According to the scaling law hypothesis in the paper, what three components govern performance on complex reasoning tasks?\nAnswer:", + " Model size, dataset quality, and pretraining loss" + ], + [ + "Question:According to the scaling law hypothesis in the paper, what three components govern performance on complex reasoning tasks?\nAnswer:", + " Training data, inference-time compute, and instruction tuning" + ], + [ + "Question:According to the scaling law hypothesis in the paper, what three components govern performance on complex reasoning tasks?\nAnswer:", + " Model size, training data (compute), and inference-time compute" + ], + [ + "Question:According to the scaling law hypothesis in the paper, what three components govern performance on complex reasoning tasks?\nAnswer:", + " Prompt quality, alignment tuning, and data diversity" + ] + ], + "resps": [ + [ + [ + -32.06525421142578, + false + ] + ], + [ + [ + -33.734771728515625, + false + ] + ], + [ + [ + -47.407806396484375, + false + ] + ], + [ + [ + -39.53954315185547, + false + ] + ] + ], + "filtered_resps": [ + [ + -32.06525421142578, + false + ], + [ + -33.734771728515625, + false + ], + [ + -47.407806396484375, + false + ], + [ + -39.53954315185547, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "159836df9c4f73cd3009a29ec4b8941010c8a5ebcfa4dff1037083f566dc7353", + "prompt_hash": "5d7b828f6fa4137c020ef5362dcc6a07274c8fa64e6abd14d5a70074996a69a5", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 60, + "doc": { + "question": "What sampling resolution is used by PixArt-Σ in text-conditioned generation?", + "choices": [ + "256", + "512", + "768", + "1024" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What sampling resolution is used by PixArt-Σ in text-conditioned generation?\nAnswer:", + " 256" + ], + [ + "Question:What sampling resolution is used by PixArt-Σ in text-conditioned generation?\nAnswer:", + " 512" + ], + [ + "Question:What sampling resolution is used by PixArt-Σ in text-conditioned generation?\nAnswer:", + " 768" + ], + [ + "Question:What sampling resolution is used by PixArt-Σ in text-conditioned generation?\nAnswer:", + " 1024" + ] + ], + "resps": [ + [ + [ + -4.888120651245117, + false + ] + ], + [ + [ + -5.437291622161865, + false + ] + ], + [ + [ + -8.039185523986816, + false + ] + ], + [ + [ + -5.633706092834473, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.888120651245117, + false + ], + [ + -5.437291622161865, + false + ], + [ + -8.039185523986816, + false + ], + [ + -5.633706092834473, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "97970f6d8f55383bf53b7ea1c2a7f6598e49b38490cb21931965128928d99680", + "prompt_hash": "8d1205772c1e4887ae735da29fb965b85694e64a5b3d66d7d983d2039f1e9d7b", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 61, + "doc": { + "question": "What search algorithm uses a pivot noise and samples in its neighborhood without gradient information?", + "choices": [ + "Random Search", + "Zero-Order Search", + "First-Order Search", + "Greedy Search" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What search algorithm uses a pivot noise and samples in its neighborhood without gradient information?\nAnswer:", + " Random Search" + ], + [ + "Question:What search algorithm uses a pivot noise and samples in its neighborhood without gradient information?\nAnswer:", + " Zero-Order Search" + ], + [ + "Question:What search algorithm uses a pivot noise and samples in its neighborhood without gradient information?\nAnswer:", + " First-Order Search" + ], + [ + "Question:What search algorithm uses a pivot noise and samples in its neighborhood without gradient information?\nAnswer:", + " Greedy Search" + ] + ], + "resps": [ + [ + [ + -7.393685340881348, + false + ] + ], + [ + [ + -16.32645606994629, + false + ] + ], + [ + [ + -17.336776733398438, + false + ] + ], + [ + [ + -9.410146713256836, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.393685340881348, + false + ], + [ + -16.32645606994629, + false + ], + [ + -17.336776733398438, + false + ], + [ + -9.410146713256836, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "357a3b5c14062a599863360c517c0c36af21071861d1d30d8642005017926ee0", + "prompt_hash": "3a6b0cd595528e1087d58a3792b07f8603504acbe6876de34117950e5d8447f3", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 62, + "doc": { + "question": "Which verifier computes scores by comparing the cosine similarity of visual and text features, and rescales them?", + "choices": [ + "Aesthetic", + "CLIPScore", + "ImageReward", + "Verifier Ensemble" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which verifier computes scores by comparing the cosine similarity of visual and text features, and rescales them?\nAnswer:", + " Aesthetic" + ], + [ + "Question:Which verifier computes scores by comparing the cosine similarity of visual and text features, and rescales them?\nAnswer:", + " CLIPScore" + ], + [ + "Question:Which verifier computes scores by comparing the cosine similarity of visual and text features, and rescales them?\nAnswer:", + " ImageReward" + ], + [ + "Question:Which verifier computes scores by comparing the cosine similarity of visual and text features, and rescales them?\nAnswer:", + " Verifier Ensemble" + ] + ], + "resps": [ + [ + [ + -12.820409774780273, + false + ] + ], + [ + [ + -9.78781795501709, + false + ] + ], + [ + [ + -13.850737571716309, + false + ] + ], + [ + [ + -17.897645950317383, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.820409774780273, + false + ], + [ + -9.78781795501709, + false + ], + [ + -13.850737571716309, + false + ], + [ + -17.897645950317383, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d943fede6f4223c92f39b95e19cabe457a468968f4425bb33b07615bf0615b05", + "prompt_hash": "72550d78187553582db0694b36f24c2cf00fa605ba169bbb8f1383db3f4687bc", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 63, + "doc": { + "question": "What is the main drawback of excessive search against a verifier like DINO or CLIP on ImageNet?", + "choices": [ + "It increases diversity but lowers sample quality", + "It causes mode collapse and increases FID", + "It improves all metrics uniformly", + "It reduces the need for classifier-free guidance" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main drawback of excessive search against a verifier like DINO or CLIP on ImageNet?\nAnswer:", + " It increases diversity but lowers sample quality" + ], + [ + "Question:What is the main drawback of excessive search against a verifier like DINO or CLIP on ImageNet?\nAnswer:", + " It causes mode collapse and increases FID" + ], + [ + "Question:What is the main drawback of excessive search against a verifier like DINO or CLIP on ImageNet?\nAnswer:", + " It improves all metrics uniformly" + ], + [ + "Question:What is the main drawback of excessive search against a verifier like DINO or CLIP on ImageNet?\nAnswer:", + " It reduces the need for classifier-free guidance" + ] + ], + "resps": [ + [ + [ + -28.893138885498047, + false + ] + ], + [ + [ + -33.32884979248047, + false + ] + ], + [ + [ + -29.780744552612305, + false + ] + ], + [ + [ + -24.765605926513672, + false + ] + ] + ], + "filtered_resps": [ + [ + -28.893138885498047, + false + ], + [ + -33.32884979248047, + false + ], + [ + -29.780744552612305, + false + ], + [ + -24.765605926513672, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "66cb58bfc739e7209395b48f1f60b50c7cd403a232ed87e6b21afe306162572e", + "prompt_hash": "bffa743545eee44620a51d1937ccd6a7a38c78ee5e35b3afb4289e5623c9f710", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 64, + "doc": { + "question": "Which method enables gradient descent on noise during sampling by reducing memory cost?", + "choices": [ + "Random Search", + "Search Over Paths", + "Gradient Checkpointing", + "Classifier-Free Guidance" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which method enables gradient descent on noise during sampling by reducing memory cost?\nAnswer:", + " Random Search" + ], + [ + "Question:Which method enables gradient descent on noise during sampling by reducing memory cost?\nAnswer:", + " Search Over Paths" + ], + [ + "Question:Which method enables gradient descent on noise during sampling by reducing memory cost?\nAnswer:", + " Gradient Checkpointing" + ], + [ + "Question:Which method enables gradient descent on noise during sampling by reducing memory cost?\nAnswer:", + " Classifier-Free Guidance" + ] + ], + "resps": [ + [ + [ + -11.641936302185059, + false + ] + ], + [ + [ + -25.17833709716797, + false + ] + ], + [ + [ + -12.970060348510742, + false + ] + ], + [ + [ + -15.819764137268066, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.641936302185059, + false + ], + [ + -25.17833709716797, + false + ], + [ + -12.970060348510742, + false + ], + [ + -15.819764137268066, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ccb6fb769cd18ca7b53dae771a5727a281fe06ce1d2f430a267405324a0f9c91", + "prompt_hash": "2e1663f5c49e461573230553cb5b952dc9680ce3d0736bd8847c48a1ef72c907", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 65, + "doc": { + "question": "Which model achieves 2,029 Elo rating on Codeforces, outperforming 96.3% of human participants?", + "choices": [ + "OpenAI-o1-1217", + "DeepSeek-V3", + "DeepSeek-R1", + "Claude-Sonnet-3.5" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model achieves 2,029 Elo rating on Codeforces, outperforming 96.3% of human participants?\nAnswer:", + " OpenAI-o1-1217" + ], + [ + "Question:Which model achieves 2,029 Elo rating on Codeforces, outperforming 96.3% of human participants?\nAnswer:", + " DeepSeek-V3" + ], + [ + "Question:Which model achieves 2,029 Elo rating on Codeforces, outperforming 96.3% of human participants?\nAnswer:", + " DeepSeek-R1" + ], + [ + "Question:Which model achieves 2,029 Elo rating on Codeforces, outperforming 96.3% of human participants?\nAnswer:", + " Claude-Sonnet-3.5" + ] + ], + "resps": [ + [ + [ + -41.88161849975586, + false + ] + ], + [ + [ + -18.81478500366211, + false + ] + ], + [ + [ + -25.116798400878906, + false + ] + ], + [ + [ + -23.355180740356445, + false + ] + ] + ], + "filtered_resps": [ + [ + -41.88161849975586, + false + ], + [ + -18.81478500366211, + false + ], + [ + -25.116798400878906, + false + ], + [ + -23.355180740356445, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "12f635897625f78e42722d6e00b292a1329bb5992e8237414d1c800e2c747446", + "prompt_hash": "dc10c75cd76d40ab92ce60278ccb889c7f3c23e805ff127830e5a6d7a91352c5", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 66, + "doc": { + "question": "What base model is used to develop DeepSeek-R1?", + "choices": [ + "Qwen2.5-32B", + "DeepSeek-V3-Base", + "GPT-4o", + "LLaMA-3" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What base model is used to develop DeepSeek-R1?\nAnswer:", + " Qwen2.5-32B" + ], + [ + "Question:What base model is used to develop DeepSeek-R1?\nAnswer:", + " DeepSeek-V3-Base" + ], + [ + "Question:What base model is used to develop DeepSeek-R1?\nAnswer:", + " GPT-4o" + ], + [ + "Question:What base model is used to develop DeepSeek-R1?\nAnswer:", + " LLaMA-3" + ] + ], + "resps": [ + [ + [ + -23.438631057739258, + false + ] + ], + [ + [ + -12.684707641601562, + false + ] + ], + [ + [ + -9.185466766357422, + false + ] + ], + [ + [ + -10.656867027282715, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.438631057739258, + false + ], + [ + -12.684707641601562, + false + ], + [ + -9.185466766357422, + false + ], + [ + -10.656867027282715, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "830b53bf691311178035303428a1bc6422c64e1b305e3ccff7febbc55cc5a938", + "prompt_hash": "1c46372fd028656f0d94cd8b52585ed208858a18344c1e95f3855aec6cc90230", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 67, + "doc": { + "question": "What method is used in DeepSeek-R1 to estimate the baseline without a critic model?", + "choices": [ + "PPO", + "KL divergence", + "Group Relative Policy Optimization (GRPO)", + "Monte Carlo Tree Search (MCTS)" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What method is used in DeepSeek-R1 to estimate the baseline without a critic model?\nAnswer:", + " PPO" + ], + [ + "Question:What method is used in DeepSeek-R1 to estimate the baseline without a critic model?\nAnswer:", + " KL divergence" + ], + [ + "Question:What method is used in DeepSeek-R1 to estimate the baseline without a critic model?\nAnswer:", + " Group Relative Policy Optimization (GRPO)" + ], + [ + "Question:What method is used in DeepSeek-R1 to estimate the baseline without a critic model?\nAnswer:", + " Monte Carlo Tree Search (MCTS)" + ] + ], + "resps": [ + [ + [ + -6.6992340087890625, + false + ] + ], + [ + [ + -8.444036483764648, + false + ] + ], + [ + [ + -26.221853256225586, + false + ] + ], + [ + [ + -15.762158393859863, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.6992340087890625, + false + ], + [ + -8.444036483764648, + false + ], + [ + -26.221853256225586, + false + ], + [ + -15.762158393859863, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b2da82c32ba8294928839d6e6c8d24f391d9b9539a533e50ab6f9e67898edcec", + "prompt_hash": "9835b440f352329d56947d3df92af42b7df5090cff13c18de40dd239152b41ac", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 68, + "doc": { + "question": "Which distilled model achieves 72.6% on AIME 2024?", + "choices": [ + "DeepSeek-R1-Distill-Qwen-7B", + "DeepSeek-R1-Distill-Qwen-14B", + "DeepSeek-R1-Distill-Qwen-32B", + "DeepSeek-R1-Distill-Qwen-1.5B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which distilled model achieves 72.6% on AIME 2024?\nAnswer:", + " DeepSeek-R1-Distill-Qwen-7B" + ], + [ + "Question:Which distilled model achieves 72.6% on AIME 2024?\nAnswer:", + " DeepSeek-R1-Distill-Qwen-14B" + ], + [ + "Question:Which distilled model achieves 72.6% on AIME 2024?\nAnswer:", + " DeepSeek-R1-Distill-Qwen-32B" + ], + [ + "Question:Which distilled model achieves 72.6% on AIME 2024?\nAnswer:", + " DeepSeek-R1-Distill-Qwen-1.5B" + ] + ], + "resps": [ + [ + [ + -36.866294860839844, + false + ] + ], + [ + [ + -39.773773193359375, + false + ] + ], + [ + [ + -42.2109375, + false + ] + ], + [ + [ + -45.068634033203125, + false + ] + ] + ], + "filtered_resps": [ + [ + -36.866294860839844, + false + ], + [ + -39.773773193359375, + false + ], + [ + -42.2109375, + false + ], + [ + -45.068634033203125, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "94fbc30c8d5f15538ae62b1b0b1d19ed80bdd98942645eb62fd8aefdbb0d8c64", + "prompt_hash": "08b5360c1d40bcc884fedc3389ee606454b71235337d0a0d7af2a0bac79ade0e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 69, + "doc": { + "question": "Which token format does DeepSeek-R1 use to enclose the reasoning process during training?", + "choices": [ + "[REASON] and [/REASON]", + " and ", + " and ", + "[THINK] and [/THINK]" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which token format does DeepSeek-R1 use to enclose the reasoning process during training?\nAnswer:", + " [REASON] and [/REASON]" + ], + [ + "Question:Which token format does DeepSeek-R1 use to enclose the reasoning process during training?\nAnswer:", + " and " + ], + [ + "Question:Which token format does DeepSeek-R1 use to enclose the reasoning process during training?\nAnswer:", + " and " + ], + [ + "Question:Which token format does DeepSeek-R1 use to enclose the reasoning process during training?\nAnswer:", + " [THINK] and [/THINK]" + ] + ], + "resps": [ + [ + [ + -35.707176208496094, + false + ] + ], + [ + [ + -25.014892578125, + false + ] + ], + [ + [ + -28.014991760253906, + false + ] + ], + [ + [ + -32.550506591796875, + false + ] + ] + ], + "filtered_resps": [ + [ + -35.707176208496094, + false + ], + [ + -25.014892578125, + false + ], + [ + -28.014991760253906, + false + ], + [ + -32.550506591796875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "34f46444412900bed38bc52bb80ad9ccfa5f235245b958c94c219766c3bd820c", + "prompt_hash": "59420b6f9ad70812e0273f197b5a199808b0a46ccbeba15fd6136f297f35a279", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 70, + "doc": { + "question": "What score did Janus-Pro-7B achieve on the MMBench benchmark?", + "choices": [ + "69.4", + "75.2", + "79.2", + "68.9" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What score did Janus-Pro-7B achieve on the MMBench benchmark?\nAnswer:", + " 69.4" + ], + [ + "Question:What score did Janus-Pro-7B achieve on the MMBench benchmark?\nAnswer:", + " 75.2" + ], + [ + "Question:What score did Janus-Pro-7B achieve on the MMBench benchmark?\nAnswer:", + " 79.2" + ], + [ + "Question:What score did Janus-Pro-7B achieve on the MMBench benchmark?\nAnswer:", + " 68.9" + ] + ], + "resps": [ + [ + [ + -8.975870132446289, + false + ] + ], + [ + [ + -8.504131317138672, + false + ] + ], + [ + [ + -8.818510055541992, + false + ] + ], + [ + [ + -8.981317520141602, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.975870132446289, + false + ], + [ + -8.504131317138672, + false + ], + [ + -8.818510055541992, + false + ], + [ + -8.981317520141602, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b2201c41f25906adcc32e575d556ec75c843fd5382a1dbbb4424ede976f2d279", + "prompt_hash": "e4deea529e6cebf2be6a0965ffe863d1cdb6f6c1ee616a49e6d70b1a33221159", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 71, + "doc": { + "question": "What vision encoder is used in Janus-Pro for multimodal understanding tasks?", + "choices": [ + "SigLIP-Large-Patch16-384", + "BLIP-2", + "ViT-B/16", + "Swin Transformer" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What vision encoder is used in Janus-Pro for multimodal understanding tasks?\nAnswer:", + " SigLIP-Large-Patch16-384" + ], + [ + "Question:What vision encoder is used in Janus-Pro for multimodal understanding tasks?\nAnswer:", + " BLIP-2" + ], + [ + "Question:What vision encoder is used in Janus-Pro for multimodal understanding tasks?\nAnswer:", + " ViT-B/16" + ], + [ + "Question:What vision encoder is used in Janus-Pro for multimodal understanding tasks?\nAnswer:", + " Swin Transformer" + ] + ], + "resps": [ + [ + [ + -35.78590393066406, + false + ] + ], + [ + [ + -6.9924750328063965, + false + ] + ], + [ + [ + -9.447872161865234, + false + ] + ], + [ + [ + -11.09534740447998, + false + ] + ] + ], + "filtered_resps": [ + [ + -35.78590393066406, + false + ], + [ + -6.9924750328063965, + false + ], + [ + -9.447872161865234, + false + ], + [ + -11.09534740447998, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e31fb9041b3ed83733d7ab2fef2e800237d29146168e7685a4d0e3b5a2062fd1", + "prompt_hash": "368837545e870bb25b65ae2fb5633fd6c56ec6fc633dde24daa17aa1384e37c0", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 72, + "doc": { + "question": "What is the context window size for both Janus-Pro-1B and Janus-Pro-7B?", + "choices": [ + "2048", + "8192", + "4096", + "1024" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the context window size for both Janus-Pro-1B and Janus-Pro-7B?\nAnswer:", + " 2048" + ], + [ + "Question:What is the context window size for both Janus-Pro-1B and Janus-Pro-7B?\nAnswer:", + " 8192" + ], + [ + "Question:What is the context window size for both Janus-Pro-1B and Janus-Pro-7B?\nAnswer:", + " 4096" + ], + [ + "Question:What is the context window size for both Janus-Pro-1B and Janus-Pro-7B?\nAnswer:", + " 1024" + ] + ], + "resps": [ + [ + [ + -4.942681312561035, + false + ] + ], + [ + [ + -5.659657955169678, + false + ] + ], + [ + [ + -5.5945634841918945, + false + ] + ], + [ + [ + -4.923760414123535, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.942681312561035, + false + ], + [ + -5.659657955169678, + false + ], + [ + -5.5945634841918945, + false + ], + [ + -4.923760414123535, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "81611cf017515f09c3ec9e1e5155afb245d7d7ec751eef1606116a4a7a2c8183", + "prompt_hash": "19f05377c6db8c73c7a629329ce9acb3393ed0f98957bf9f3bb59cf39ecbad18", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 73, + "doc": { + "question": "What is the optimizer used to train Janus-Pro?", + "choices": [ + "SGD", + "Adam", + "RMSProp", + "AdamW" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the optimizer used to train Janus-Pro?\nAnswer:", + " SGD" + ], + [ + "Question:What is the optimizer used to train Janus-Pro?\nAnswer:", + " Adam" + ], + [ + "Question:What is the optimizer used to train Janus-Pro?\nAnswer:", + " RMSProp" + ], + [ + "Question:What is the optimizer used to train Janus-Pro?\nAnswer:", + " AdamW" + ] + ], + "resps": [ + [ + [ + -5.474011421203613, + false + ] + ], + [ + [ + -2.1715750694274902, + false + ] + ], + [ + [ + -7.099437236785889, + false + ] + ], + [ + [ + -3.740978717803955, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.474011421203613, + false + ], + [ + -2.1715750694274902, + false + ], + [ + -7.099437236785889, + false + ], + [ + -3.740978717803955, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f4bbf766db06865279ba3415fed33c24892caf06f882a57bb85fb256ad305048", + "prompt_hash": "8562fd057d4b35d3d349c1d6cab66dd07358bf6535936e8013c49fc2d198d3b1", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 74, + "doc": { + "question": "Which benchmark did Janus-Pro-7B outperform DALL-E 3 and SD3-Medium in?", + "choices": [ + "MMBench", + "GenEval", + "GQA", + "MMMU" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which benchmark did Janus-Pro-7B outperform DALL-E 3 and SD3-Medium in?\nAnswer:", + " MMBench" + ], + [ + "Question:Which benchmark did Janus-Pro-7B outperform DALL-E 3 and SD3-Medium in?\nAnswer:", + " GenEval" + ], + [ + "Question:Which benchmark did Janus-Pro-7B outperform DALL-E 3 and SD3-Medium in?\nAnswer:", + " GQA" + ], + [ + "Question:Which benchmark did Janus-Pro-7B outperform DALL-E 3 and SD3-Medium in?\nAnswer:", + " MMMU" + ] + ], + "resps": [ + [ + [ + -13.650606155395508, + false + ] + ], + [ + [ + -14.764251708984375, + false + ] + ], + [ + [ + -10.909505844116211, + false + ] + ], + [ + [ + -15.465291976928711, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.650606155395508, + false + ], + [ + -14.764251708984375, + false + ], + [ + -10.909505844116211, + false + ], + [ + -15.465291976928711, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "01df0a0c6003a0a1e26d89da2d52bfda816fde996d530f584115625e060569d0", + "prompt_hash": "92b90be18377ba553ebd74eff19379547879463d09ad0346e6a83ed4491ae419", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 75, + "doc": { + "question": "Which tokenization method is used in FAST to compress robot action trajectories?", + "choices": [ + "Principal Component Analysis (PCA)", + "Byte-Pair Encoding (BPE) followed by Discrete Cosine Transform (DCT)", + "Discrete Cosine Transform (DCT) followed by Byte-Pair Encoding (BPE)", + "Vector Quantization followed by Huffman Encoding" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which tokenization method is used in FAST to compress robot action trajectories?\nAnswer:", + " Principal Component Analysis (PCA)" + ], + [ + "Question:Which tokenization method is used in FAST to compress robot action trajectories?\nAnswer:", + " Byte-Pair Encoding (BPE) followed by Discrete Cosine Transform (DCT)" + ], + [ + "Question:Which tokenization method is used in FAST to compress robot action trajectories?\nAnswer:", + " Discrete Cosine Transform (DCT) followed by Byte-Pair Encoding (BPE)" + ], + [ + "Question:Which tokenization method is used in FAST to compress robot action trajectories?\nAnswer:", + " Vector Quantization followed by Huffman Encoding" + ] + ], + "resps": [ + [ + [ + -12.270536422729492, + false + ] + ], + [ + [ + -28.064403533935547, + false + ] + ], + [ + [ + -30.861492156982422, + false + ] + ], + [ + [ + -27.247623443603516, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.270536422729492, + false + ], + [ + -28.064403533935547, + false + ], + [ + -30.861492156982422, + false + ], + [ + -27.247623443603516, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "bc20e847d46c6df16829c8f8f51c270aab55e343930e22ad24b4b80c130e8a89", + "prompt_hash": "6780db88023435c5ac2aeabf4f41c3af941c4a465c712acf25388a7131518804", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 76, + "doc": { + "question": "What is the action chunk duration used for policy training in the described VLA experiments?", + "choices": [ + "0.5 seconds", + "2 seconds", + "5 seconds", + "1 second" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the action chunk duration used for policy training in the described VLA experiments?\nAnswer:", + " 0.5 seconds" + ], + [ + "Question:What is the action chunk duration used for policy training in the described VLA experiments?\nAnswer:", + " 2 seconds" + ], + [ + "Question:What is the action chunk duration used for policy training in the described VLA experiments?\nAnswer:", + " 5 seconds" + ], + [ + "Question:What is the action chunk duration used for policy training in the described VLA experiments?\nAnswer:", + " 1 second" + ] + ], + "resps": [ + [ + [ + -7.42416524887085, + false + ] + ], + [ + [ + -5.825782775878906, + false + ] + ], + [ + [ + -5.891945838928223, + false + ] + ], + [ + [ + -5.117180347442627, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.42416524887085, + false + ], + [ + -5.825782775878906, + false + ], + [ + -5.891945838928223, + false + ], + [ + -5.117180347442627, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b2ab5010ee473215515f17ab098c12cdea5091d6623315912d4d4e695273bb67", + "prompt_hash": "668d931deaf85161766f8c175e973cb02b9540443c0d2c8bbf77c366bf489b08", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 77, + "doc": { + "question": "Which evaluation task uses a bi-manual ARX robot to fold a t-shirt?", + "choices": [ + "Laundry Folding", + "Toast out of Toaster", + "T-Shirt Folding", + "Table Bussing" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which evaluation task uses a bi-manual ARX robot to fold a t-shirt?\nAnswer:", + " Laundry Folding" + ], + [ + "Question:Which evaluation task uses a bi-manual ARX robot to fold a t-shirt?\nAnswer:", + " Toast out of Toaster" + ], + [ + "Question:Which evaluation task uses a bi-manual ARX robot to fold a t-shirt?\nAnswer:", + " T-Shirt Folding" + ], + [ + "Question:Which evaluation task uses a bi-manual ARX robot to fold a t-shirt?\nAnswer:", + " Table Bussing" + ] + ], + "resps": [ + [ + [ + -15.65053653717041, + false + ] + ], + [ + [ + -34.21724319458008, + false + ] + ], + [ + [ + -7.9458465576171875, + false + ] + ], + [ + [ + -26.62312889099121, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.65053653717041, + false + ], + [ + -34.21724319458008, + false + ], + [ + -7.9458465576171875, + false + ], + [ + -26.62312889099121, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "535b5c2b5e85ffd70b07c718bcd48ae81fb7299737cbf4c4346da27a0de5bc0a", + "prompt_hash": "5df4789577c533275bffc59775528b1a167b7932bb27a2967fabdf5117b59865", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 78, + "doc": { + "question": "What optimizer is used for policy training in the VLA experiments?", + "choices": [ + "SGD with Momentum", + "Adam", + "AdamW", + "RMSProp" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What optimizer is used for policy training in the VLA experiments?\nAnswer:", + " SGD with Momentum" + ], + [ + "Question:What optimizer is used for policy training in the VLA experiments?\nAnswer:", + " Adam" + ], + [ + "Question:What optimizer is used for policy training in the VLA experiments?\nAnswer:", + " AdamW" + ], + [ + "Question:What optimizer is used for policy training in the VLA experiments?\nAnswer:", + " RMSProp" + ] + ], + "resps": [ + [ + [ + -13.135553359985352, + false + ] + ], + [ + [ + -3.396702766418457, + false + ] + ], + [ + [ + -5.222354888916016, + false + ] + ], + [ + [ + -6.109254360198975, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.135553359985352, + false + ], + [ + -3.396702766418457, + false + ], + [ + -5.222354888916016, + false + ], + [ + -6.109254360198975, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c6ca2f0960faaebe4c1c1a5ed7c0bb61c4ace30f26b3b586ac0f0c8a3e672e44", + "prompt_hash": "10273564162dc5c045ea3ead115969d25d3a00836aaf9eb8620485f76cacbfe1", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 79, + "doc": { + "question": "What control frequency is associated with the T-Shirt Folding task?", + "choices": [ + "15 Hz", + "20 Hz", + "50 Hz", + "5 Hz" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What control frequency is associated with the T-Shirt Folding task?\nAnswer:", + " 15 Hz" + ], + [ + "Question:What control frequency is associated with the T-Shirt Folding task?\nAnswer:", + " 20 Hz" + ], + [ + "Question:What control frequency is associated with the T-Shirt Folding task?\nAnswer:", + " 50 Hz" + ], + [ + "Question:What control frequency is associated with the T-Shirt Folding task?\nAnswer:", + " 5 Hz" + ] + ], + "resps": [ + [ + [ + -6.0199737548828125, + false + ] + ], + [ + [ + -5.349177360534668, + false + ] + ], + [ + [ + -5.311049461364746, + false + ] + ], + [ + [ + -5.154747009277344, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.0199737548828125, + false + ], + [ + -5.349177360534668, + false + ], + [ + -5.311049461364746, + false + ], + [ + -5.154747009277344, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fc67d63b526c2bbd910fe54a16fbd13bb882cbf2a4281310beb7bc522bbf1400", + "prompt_hash": "3536b38edac5e5232898c82dce006c4b587468dc08c5c806303e7817e30f6ceb", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 80, + "doc": { + "question": "What technique has shown that prompts like 'Let's think step by step' can improve reasoning in language models?", + "choices": [ + "Chain-of-Thought prompting", + "Least-to-Most prompting", + "Plan-and-Solve", + "Tree-of-Thought prompting" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What technique has shown that prompts like 'Let's think step by step' can improve reasoning in language models?\nAnswer:", + " Chain-of-Thought prompting" + ], + [ + "Question:What technique has shown that prompts like 'Let's think step by step' can improve reasoning in language models?\nAnswer:", + " Least-to-Most prompting" + ], + [ + "Question:What technique has shown that prompts like 'Let's think step by step' can improve reasoning in language models?\nAnswer:", + " Plan-and-Solve" + ], + [ + "Question:What technique has shown that prompts like 'Let's think step by step' can improve reasoning in language models?\nAnswer:", + " Tree-of-Thought prompting" + ] + ], + "resps": [ + [ + [ + -7.43093204498291, + false + ] + ], + [ + [ + -16.591632843017578, + false + ] + ], + [ + [ + -15.584571838378906, + false + ] + ], + [ + [ + -10.290903091430664, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.43093204498291, + false + ], + [ + -16.591632843017578, + false + ], + [ + -15.584571838378906, + false + ], + [ + -10.290903091430664, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "4895c2f15e6e631e16395be84803e67bf4ec8c1ba2694ccfed56750089722f91", + "prompt_hash": "88de32826fc0329be4c66c083fadb45cbe138d6ebdb91cae39a9799eb3ad610a", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 81, + "doc": { + "question": "What does Direct Preference Optimization (DPO) eliminate in reinforcement learning fine-tuning?", + "choices": [ + "Prompt engineering", + "Reward model training and sampling", + "Token-level scoring", + "Gradient clipping" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What does Direct Preference Optimization (DPO) eliminate in reinforcement learning fine-tuning?\nAnswer:", + " Prompt engineering" + ], + [ + "Question:What does Direct Preference Optimization (DPO) eliminate in reinforcement learning fine-tuning?\nAnswer:", + " Reward model training and sampling" + ], + [ + "Question:What does Direct Preference Optimization (DPO) eliminate in reinforcement learning fine-tuning?\nAnswer:", + " Token-level scoring" + ], + [ + "Question:What does Direct Preference Optimization (DPO) eliminate in reinforcement learning fine-tuning?\nAnswer:", + " Gradient clipping" + ] + ], + "resps": [ + [ + [ + -10.809982299804688, + false + ] + ], + [ + [ + -27.00271224975586, + false + ] + ], + [ + [ + -22.927026748657227, + false + ] + ], + [ + [ + -11.805936813354492, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.809982299804688, + false + ], + [ + -27.00271224975586, + false + ], + [ + -22.927026748657227, + false + ], + [ + -11.805936813354492, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "556e5a4a3ff7ffe7968dd7acadb8155bfec81f188a9c06600fbb9490117cd22b", + "prompt_hash": "df6ee6fd84ed4ca6c319809234c9234f25b815c286da491bfae8916ca174f2e3", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 82, + "doc": { + "question": "What strategy in Monte Carlo Tree Search (MCTS) helps reduce computational overhead in reasoning tasks?", + "choices": [ + "Heuristic initialization", + "Gradient pruning", + "Leaf node sharing of high-quality parent nodes", + "Prompt compression" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What strategy in Monte Carlo Tree Search (MCTS) helps reduce computational overhead in reasoning tasks?\nAnswer:", + " Heuristic initialization" + ], + [ + "Question:What strategy in Monte Carlo Tree Search (MCTS) helps reduce computational overhead in reasoning tasks?\nAnswer:", + " Gradient pruning" + ], + [ + "Question:What strategy in Monte Carlo Tree Search (MCTS) helps reduce computational overhead in reasoning tasks?\nAnswer:", + " Leaf node sharing of high-quality parent nodes" + ], + [ + "Question:What strategy in Monte Carlo Tree Search (MCTS) helps reduce computational overhead in reasoning tasks?\nAnswer:", + " Prompt compression" + ] + ], + "resps": [ + [ + [ + -15.309856414794922, + false + ] + ], + [ + [ + -19.264671325683594, + false + ] + ], + [ + [ + -45.913230895996094, + false + ] + ], + [ + [ + -18.06313133239746, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.309856414794922, + false + ], + [ + -19.264671325683594, + false + ], + [ + -45.913230895996094, + false + ], + [ + -18.06313133239746, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c4f7658a1b0e85d0f29104240811cad7577426cdd20d7c60df3922dc9ce83068", + "prompt_hash": "770d1f160ad2a838c42b8537f5f53b79b454987d2870b0681c43ed3785c539b0", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 83, + "doc": { + "question": "Which reinforcement fine-tuning technique is noted for enabling learning from just dozens of user preference data?", + "choices": [ + "ReFT", + "VinePPO", + "RLHF", + "SVPO" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which reinforcement fine-tuning technique is noted for enabling learning from just dozens of user preference data?\nAnswer:", + " ReFT" + ], + [ + "Question:Which reinforcement fine-tuning technique is noted for enabling learning from just dozens of user preference data?\nAnswer:", + " VinePPO" + ], + [ + "Question:Which reinforcement fine-tuning technique is noted for enabling learning from just dozens of user preference data?\nAnswer:", + " RLHF" + ], + [ + "Question:Which reinforcement fine-tuning technique is noted for enabling learning from just dozens of user preference data?\nAnswer:", + " SVPO" + ] + ], + "resps": [ + [ + [ + -16.842266082763672, + false + ] + ], + [ + [ + -26.258159637451172, + false + ] + ], + [ + [ + -5.796584606170654, + false + ] + ], + [ + [ + -23.100177764892578, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.842266082763672, + false + ], + [ + -26.258159637451172, + false + ], + [ + -5.796584606170654, + false + ], + [ + -23.100177764892578, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "64deeaefa753f9a68d8ab08e24fd3dafcb8db63bf159e920cb613b2491d819b7", + "prompt_hash": "f7b999fc73f2232a7facb7f3e4c02090f7fc41ad7d90b92f7ff00fc5cb9ee231", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 84, + "doc": { + "question": "Which prompting method breaks down complex problems into smaller subproblems for better reasoning?", + "choices": [ + "Tree-of-Thoughts", + "Self-Consistency", + "Least-to-Most Prompting", + "Chain-of-Thought" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which prompting method breaks down complex problems into smaller subproblems for better reasoning?\nAnswer:", + " Tree-of-Thoughts" + ], + [ + "Question:Which prompting method breaks down complex problems into smaller subproblems for better reasoning?\nAnswer:", + " Self-Consistency" + ], + [ + "Question:Which prompting method breaks down complex problems into smaller subproblems for better reasoning?\nAnswer:", + " Least-to-Most Prompting" + ], + [ + "Question:Which prompting method breaks down complex problems into smaller subproblems for better reasoning?\nAnswer:", + " Chain-of-Thought" + ] + ], + "resps": [ + [ + [ + -9.340664863586426, + false + ] + ], + [ + [ + -9.643707275390625, + false + ] + ], + [ + [ + -13.097075462341309, + false + ] + ], + [ + [ + -5.191791534423828, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.340664863586426, + false + ], + [ + -9.643707275390625, + false + ], + [ + -13.097075462341309, + false + ], + [ + -5.191791534423828, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "451ac49f5ff3b62849ec7af0502925f7a3f42595bd827636a337dbcbf8da985f", + "prompt_hash": "1bf69240ab675c5dfb72f2d01b0f633f72f35c2ca71789d9ca2d6052d4c20981", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 85, + "doc": { + "question": "Which activation function is used in the Qwen2.5 dense model architecture?", + "choices": [ + "ReLU", + "SwiGLU", + "GELU", + "SiLU" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which activation function is used in the Qwen2.5 dense model architecture?\nAnswer:", + " ReLU" + ], + [ + "Question:Which activation function is used in the Qwen2.5 dense model architecture?\nAnswer:", + " SwiGLU" + ], + [ + "Question:Which activation function is used in the Qwen2.5 dense model architecture?\nAnswer:", + " GELU" + ], + [ + "Question:Which activation function is used in the Qwen2.5 dense model architecture?\nAnswer:", + " SiLU" + ] + ], + "resps": [ + [ + [ + -3.774085760116577, + false + ] + ], + [ + [ + -8.605820655822754, + false + ] + ], + [ + [ + -5.870603561401367, + false + ] + ], + [ + [ + -7.402235984802246, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.774085760116577, + false + ], + [ + -8.605820655822754, + false + ], + [ + -5.870603561401367, + false + ], + [ + -7.402235984802246, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "6d76752172c847efd8d622077eedb3a62333a62e3f12e9e306ed0795b12dfda1", + "prompt_hash": "9ea9a1067541f5390152093014beb8623cec14591f5b08571de978de4f2e4bd4", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 86, + "doc": { + "question": "What vocabulary size is used in the Qwen2.5 tokenizer?", + "choices": [ + "32,768", + "100,000", + "151,643", + "256,000" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What vocabulary size is used in the Qwen2.5 tokenizer?\nAnswer:", + " 32,768" + ], + [ + "Question:What vocabulary size is used in the Qwen2.5 tokenizer?\nAnswer:", + " 100,000" + ], + [ + "Question:What vocabulary size is used in the Qwen2.5 tokenizer?\nAnswer:", + " 151,643" + ], + [ + "Question:What vocabulary size is used in the Qwen2.5 tokenizer?\nAnswer:", + " 256,000" + ] + ], + "resps": [ + [ + [ + -9.168275833129883, + false + ] + ], + [ + [ + -11.967578887939453, + false + ] + ], + [ + [ + -20.90795135498047, + false + ] + ], + [ + [ + -13.019368171691895, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.168275833129883, + false + ], + [ + -11.967578887939453, + false + ], + [ + -20.90795135498047, + false + ], + [ + -13.019368171691895, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "dbded3b4ad35993632462e288e6826b97d7820468e31f62f47cf06eba28b9d37", + "prompt_hash": "02c92dddcf37f76136a51a2822d6e29407b41630ad7c1b2d39f9b8fc2a60929b", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 87, + "doc": { + "question": "How many control tokens are included in the Qwen2.5 tokenizer?", + "choices": [ + "3", + "10", + "22", + "151,643" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many control tokens are included in the Qwen2.5 tokenizer?\nAnswer:", + " 3" + ], + [ + "Question:How many control tokens are included in the Qwen2.5 tokenizer?\nAnswer:", + " 10" + ], + [ + "Question:How many control tokens are included in the Qwen2.5 tokenizer?\nAnswer:", + " 22" + ], + [ + "Question:How many control tokens are included in the Qwen2.5 tokenizer?\nAnswer:", + " 151,643" + ] + ], + "resps": [ + [ + [ + -3.963737964630127, + false + ] + ], + [ + [ + -4.7243499755859375, + false + ] + ], + [ + [ + -6.240811347961426, + false + ] + ], + [ + [ + -19.046680450439453, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.963737964630127, + false + ], + [ + -4.7243499755859375, + false + ], + [ + -6.240811347961426, + false + ], + [ + -19.046680450439453, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5e6b5263857f6a7b6e410d36c03b0b1716afaf1b54f60c1e9962f2127298ac11", + "prompt_hash": "6305352bd5b8c7676abf7c3a2dfec5fce5b327bfb1c85fdf41f498492917e294", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 88, + "doc": { + "question": "Which benchmark is used to evaluate multilingual understanding in the Qwen2.5 base model?", + "choices": [ + "BELEBELE", + "TruthfulQA", + "LiveBench", + "ARC-C" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which benchmark is used to evaluate multilingual understanding in the Qwen2.5 base model?\nAnswer:", + " BELEBELE" + ], + [ + "Question:Which benchmark is used to evaluate multilingual understanding in the Qwen2.5 base model?\nAnswer:", + " TruthfulQA" + ], + [ + "Question:Which benchmark is used to evaluate multilingual understanding in the Qwen2.5 base model?\nAnswer:", + " LiveBench" + ], + [ + "Question:Which benchmark is used to evaluate multilingual understanding in the Qwen2.5 base model?\nAnswer:", + " ARC-C" + ] + ], + "resps": [ + [ + [ + -20.12228775024414, + false + ] + ], + [ + [ + -11.192878723144531, + false + ] + ], + [ + [ + -16.66100311279297, + false + ] + ], + [ + [ + -12.205469131469727, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.12228775024414, + false + ], + [ + -11.192878723144531, + false + ], + [ + -16.66100311279297, + false + ], + [ + -12.205469131469727, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ddb5b99be432bca64544f9418ed8f3363d61cb4de20929c1fcfa1c57f2fa5247", + "prompt_hash": "c7efca85bb4edd89316ecb42e3ed4f9db0718b5ed440de9ae016fe666f0bec0a", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 89, + "doc": { + "question": "What is the performance of Qwen2.5-72B on the MATH dataset?", + "choices": [ + "50.9", + "62.1", + "64.4", + "83.1" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the performance of Qwen2.5-72B on the MATH dataset?\nAnswer:", + " 50.9" + ], + [ + "Question:What is the performance of Qwen2.5-72B on the MATH dataset?\nAnswer:", + " 62.1" + ], + [ + "Question:What is the performance of Qwen2.5-72B on the MATH dataset?\nAnswer:", + " 64.4" + ], + [ + "Question:What is the performance of Qwen2.5-72B on the MATH dataset?\nAnswer:", + " 83.1" + ] + ], + "resps": [ + [ + [ + -10.716265678405762, + false + ] + ], + [ + [ + -10.533012390136719, + false + ] + ], + [ + [ + -10.480714797973633, + false + ] + ], + [ + [ + -10.151132583618164, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.716265678405762, + false + ], + [ + -10.533012390136719, + false + ], + [ + -10.480714797973633, + false + ], + [ + -10.151132583618164, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b22ab978713dabb7f86c6882088a05c621c77a6637d6387ad21bce8e66b32c11", + "prompt_hash": "7ffd84244888096461c1da3fa98ef3348eac99cf5217ecac6cfcfaf1255e2401", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 90, + "doc": { + "question": "What training loss is used in InternVL 2.5's Stage 1 and Stage 1.5?", + "choices": [ + "Contrastive loss", + "Next Token Prediction loss", + "Cosine similarity loss", + "Cross-entropy loss" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What training loss is used in InternVL 2.5's Stage 1 and Stage 1.5?\nAnswer:", + " Contrastive loss" + ], + [ + "Question:What training loss is used in InternVL 2.5's Stage 1 and Stage 1.5?\nAnswer:", + " Next Token Prediction loss" + ], + [ + "Question:What training loss is used in InternVL 2.5's Stage 1 and Stage 1.5?\nAnswer:", + " Cosine similarity loss" + ], + [ + "Question:What training loss is used in InternVL 2.5's Stage 1 and Stage 1.5?\nAnswer:", + " Cross-entropy loss" + ] + ], + "resps": [ + [ + [ + -8.985716819763184, + false + ] + ], + [ + [ + -20.21367835998535, + false + ] + ], + [ + [ + -14.408019065856934, + false + ] + ], + [ + [ + -7.583483695983887, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.985716819763184, + false + ], + [ + -20.21367835998535, + false + ], + [ + -14.408019065856934, + false + ], + [ + -7.583483695983887, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d73d1b99689947592c163a229b52313719187c2c8c271ab48aeab0cf2b2432e9", + "prompt_hash": "4ff46b1d0c77fa76208ff6114af26f58991e33db92629e1bda64fb4ad5eee115", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 91, + "doc": { + "question": "What is the final number of parameters in InternViT-6B-448px-V2.5?", + "choices": [ + "5.9B", + "6.2B", + "5.5B", + "4.8B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the final number of parameters in InternViT-6B-448px-V2.5?\nAnswer:", + " 5.9B" + ], + [ + "Question:What is the final number of parameters in InternViT-6B-448px-V2.5?\nAnswer:", + " 6.2B" + ], + [ + "Question:What is the final number of parameters in InternViT-6B-448px-V2.5?\nAnswer:", + " 5.5B" + ], + [ + "Question:What is the final number of parameters in InternViT-6B-448px-V2.5?\nAnswer:", + " 4.8B" + ] + ], + "resps": [ + [ + [ + -12.59568977355957, + false + ] + ], + [ + [ + -11.892919540405273, + false + ] + ], + [ + [ + -12.99958610534668, + false + ] + ], + [ + [ + -12.356107711791992, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.59568977355957, + false + ], + [ + -11.892919540405273, + false + ], + [ + -12.99958610534668, + false + ], + [ + -12.356107711791992, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f44a6242fcf0b2ab3a67d2ca668f0122bdc71c0ac3b57d75592826d316614528", + "prompt_hash": "18c04ac86168a8bed0fe625fa0d74a50ffd557102dbda3157b1470b2c6729b22", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 92, + "doc": { + "question": "Which prompt-based technique was shown to improve performance on MMMU in InternVL 2.5?", + "choices": [ + "Zero-shot chain-of-thought", + "Majority voting with CoT", + "Self-consistency decoding", + "Visual instruction tuning" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which prompt-based technique was shown to improve performance on MMMU in InternVL 2.5?\nAnswer:", + " Zero-shot chain-of-thought" + ], + [ + "Question:Which prompt-based technique was shown to improve performance on MMMU in InternVL 2.5?\nAnswer:", + " Majority voting with CoT" + ], + [ + "Question:Which prompt-based technique was shown to improve performance on MMMU in InternVL 2.5?\nAnswer:", + " Self-consistency decoding" + ], + [ + "Question:Which prompt-based technique was shown to improve performance on MMMU in InternVL 2.5?\nAnswer:", + " Visual instruction tuning" + ] + ], + "resps": [ + [ + [ + -14.020514488220215, + false + ] + ], + [ + [ + -23.533023834228516, + false + ] + ], + [ + [ + -19.21437644958496, + false + ] + ], + [ + [ + -12.545539855957031, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.020514488220215, + false + ], + [ + -23.533023834228516, + false + ], + [ + -19.21437644958496, + false + ], + [ + -12.545539855957031, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "37f70303923c80a15f00d15dc63d4dc1dbeb2e45703f039d8140753335cc3f10", + "prompt_hash": "2cd34b7f95101f1507b83dcd748877c9476778ce53e0263ab5512c0c7ec5fb05", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 93, + "doc": { + "question": "What is the resolution size used for each image tile in the InternVL 2.5 architecture?", + "choices": [ + "224×224", + "336×336", + "512×512", + "448×448" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the resolution size used for each image tile in the InternVL 2.5 architecture?\nAnswer:", + " 224×224" + ], + [ + "Question:What is the resolution size used for each image tile in the InternVL 2.5 architecture?\nAnswer:", + " 336×336" + ], + [ + "Question:What is the resolution size used for each image tile in the InternVL 2.5 architecture?\nAnswer:", + " 512×512" + ], + [ + "Question:What is the resolution size used for each image tile in the InternVL 2.5 architecture?\nAnswer:", + " 448×448" + ] + ], + "resps": [ + [ + [ + -8.673851013183594, + false + ] + ], + [ + [ + -10.982954978942871, + false + ] + ], + [ + [ + -7.892268180847168, + false + ] + ], + [ + [ + -9.570035934448242, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.673851013183594, + false + ], + [ + -10.982954978942871, + false + ], + [ + -7.892268180847168, + false + ], + [ + -9.570035934448242, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a28c2abde649e23518ede576c5124c3e89ba96cbedaf97d1660315b44d9cb492", + "prompt_hash": "04cf5c99c4384326cb8493acc380a00f9e8f3123c06cf3a4aad83eae3541e6a5", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 94, + "doc": { + "question": "What is the maximum number of tiles assigned per video sample during training in InternVL 2.5?", + "choices": [ + "1", + "6", + "12", + "24" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the maximum number of tiles assigned per video sample during training in InternVL 2.5?\nAnswer:", + " 1" + ], + [ + "Question:What is the maximum number of tiles assigned per video sample during training in InternVL 2.5?\nAnswer:", + " 6" + ], + [ + "Question:What is the maximum number of tiles assigned per video sample during training in InternVL 2.5?\nAnswer:", + " 12" + ], + [ + "Question:What is the maximum number of tiles assigned per video sample during training in InternVL 2.5?\nAnswer:", + " 24" + ] + ], + "resps": [ + [ + [ + -2.4906997680664062, + false + ] + ], + [ + [ + -3.65911865234375, + false + ] + ], + [ + [ + -4.247614860534668, + false + ] + ], + [ + [ + -5.091933250427246, + false + ] + ] + ], + "filtered_resps": [ + [ + -2.4906997680664062, + false + ], + [ + -3.65911865234375, + false + ], + [ + -4.247614860534668, + false + ], + [ + -5.091933250427246, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e8b5737944f8f32ee3e7b5b6356e280b9018773368427db03e9f6f0290f0670e", + "prompt_hash": "01f6f066cf469892ed01e86213bec73cf971aefa7266e5f0c85f73359ae29b6f", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 95, + "doc": { + "question": "How many models with 1B parameters were trained to evaluate the impact of data mixtures?", + "choices": [ + "32", + "64", + "128", + "256" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many models with 1B parameters were trained to evaluate the impact of data mixtures?\nAnswer:", + " 32" + ], + [ + "Question:How many models with 1B parameters were trained to evaluate the impact of data mixtures?\nAnswer:", + " 64" + ], + [ + "Question:How many models with 1B parameters were trained to evaluate the impact of data mixtures?\nAnswer:", + " 128" + ], + [ + "Question:How many models with 1B parameters were trained to evaluate the impact of data mixtures?\nAnswer:", + " 256" + ] + ], + "resps": [ + [ + [ + -6.174529075622559, + false + ] + ], + [ + [ + -7.080163955688477, + false + ] + ], + [ + [ + -8.164529800415039, + false + ] + ], + [ + [ + -8.490699768066406, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.174529075622559, + false + ], + [ + -7.080163955688477, + false + ], + [ + -8.164529800415039, + false + ], + [ + -8.490699768066406, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "22c302fdda64067f2611d938454e180cc12322d6bb6f233cf4fc7caa396ebc11", + "prompt_hash": "8fe79270102b389c7e664a2ba107cf3f68742e48a2064c5188b5e082cfc5703a", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 96, + "doc": { + "question": "Which dataset showed the strongest correlation with downstream task performance?", + "choices": [ + "Wikipedia (en)", + "WikiText", + "Pile-CC", + "C4" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which dataset showed the strongest correlation with downstream task performance?\nAnswer:", + " Wikipedia (en)" + ], + [ + "Question:Which dataset showed the strongest correlation with downstream task performance?\nAnswer:", + " WikiText" + ], + [ + "Question:Which dataset showed the strongest correlation with downstream task performance?\nAnswer:", + " Pile-CC" + ], + [ + "Question:Which dataset showed the strongest correlation with downstream task performance?\nAnswer:", + " C4" + ] + ], + "resps": [ + [ + [ + -21.24456787109375, + false + ] + ], + [ + [ + -9.600040435791016, + false + ] + ], + [ + [ + -19.288917541503906, + false + ] + ], + [ + [ + -10.923744201660156, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.24456787109375, + false + ], + [ + -9.600040435791016, + false + ], + [ + -19.288917541503906, + false + ], + [ + -10.923744201660156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "279c342e0b86c189b778e77be74e371238a82b037dd8ac7f198bb5601131badc", + "prompt_hash": "5da37f35317d56dc7b0631566fb19ef772cf50ec5acd8dde624b908cb28cb08b", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 97, + "doc": { + "question": "What evaluation metric is used if normalized accuracy is not provided by lm-eval-harness?", + "choices": [ + "BLEU score", + "F1 score", + "Perplexity", + "Regular accuracy" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What evaluation metric is used if normalized accuracy is not provided by lm-eval-harness?\nAnswer:", + " BLEU score" + ], + [ + "Question:What evaluation metric is used if normalized accuracy is not provided by lm-eval-harness?\nAnswer:", + " F1 score" + ], + [ + "Question:What evaluation metric is used if normalized accuracy is not provided by lm-eval-harness?\nAnswer:", + " Perplexity" + ], + [ + "Question:What evaluation metric is used if normalized accuracy is not provided by lm-eval-harness?\nAnswer:", + " Regular accuracy" + ] + ], + "resps": [ + [ + [ + -9.628717422485352, + false + ] + ], + [ + [ + -5.973865032196045, + false + ] + ], + [ + [ + -6.907043933868408, + false + ] + ], + [ + [ + -13.605592727661133, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.628717422485352, + false + ], + [ + -5.973865032196045, + false + ], + [ + -6.907043933868408, + false + ], + [ + -13.605592727661133, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0a7d09c18007aa19218484c38030cc8c137386c6995ff12622f783aedb310279", + "prompt_hash": "abc7194d9f039eb154a8e9c5c668965d30e904867908a5c735dbeebc91e07164", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 98, + "doc": { + "question": "Which benchmark showed the largest performance gap (Δ) between the best and worst models?", + "choices": [ + "PiQA", + "Lambada", + "HellaSwag", + "OpenBookQA" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which benchmark showed the largest performance gap (Δ) between the best and worst models?\nAnswer:", + " PiQA" + ], + [ + "Question:Which benchmark showed the largest performance gap (Δ) between the best and worst models?\nAnswer:", + " Lambada" + ], + [ + "Question:Which benchmark showed the largest performance gap (Δ) between the best and worst models?\nAnswer:", + " HellaSwag" + ], + [ + "Question:Which benchmark showed the largest performance gap (Δ) between the best and worst models?\nAnswer:", + " OpenBookQA" + ] + ], + "resps": [ + [ + [ + -14.476099014282227, + false + ] + ], + [ + [ + -12.178361892700195, + false + ] + ], + [ + [ + -8.666653633117676, + false + ] + ], + [ + [ + -12.985079765319824, + false + ] + ] + ], + "filtered_resps": [ + [ + -14.476099014282227, + false + ], + [ + -12.178361892700195, + false + ], + [ + -8.666653633117676, + false + ], + [ + -12.985079765319824, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "06b48f908f1700e94e22b781c4215de054b9b365f41c55084daedb22d807b5b6", + "prompt_hash": "05aa0d7a5d81aea67d91accbc2cffa2ea224c1d4e1f94c79e834c13a88868ad8", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 99, + "doc": { + "question": "What kind of regression model was used in \\u201courmethod\\u201d to predict the optimal data mixture?", + "choices": [ + "Lasso regression", + "Linear regression", + "LightGBM", + "Random forest" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What kind of regression model was used in \\u201courmethod\\u201d to predict the optimal data mixture?\nAnswer:", + " Lasso regression" + ], + [ + "Question:What kind of regression model was used in \\u201courmethod\\u201d to predict the optimal data mixture?\nAnswer:", + " Linear regression" + ], + [ + "Question:What kind of regression model was used in \\u201courmethod\\u201d to predict the optimal data mixture?\nAnswer:", + " LightGBM" + ], + [ + "Question:What kind of regression model was used in \\u201courmethod\\u201d to predict the optimal data mixture?\nAnswer:", + " Random forest" + ] + ], + "resps": [ + [ + [ + -10.868882179260254, + false + ] + ], + [ + [ + -5.115058898925781, + false + ] + ], + [ + [ + -12.149599075317383, + false + ] + ], + [ + [ + -8.57761001586914, + false + ] + ] + ], + "filtered_resps": [ + [ + -10.868882179260254, + false + ], + [ + -5.115058898925781, + false + ], + [ + -12.149599075317383, + false + ], + [ + -8.57761001586914, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b53157e14dd6ab1308b1db7dd65b121972f82fef1555d1a8b7bb279605a2f2f4", + "prompt_hash": "15d02ff658f84423bb16b693b71e523b96c253fb5f5075947c1008589e38ecbf", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + } + ], + "arxiv_mc_2502": [ + { + "doc_id": 0, + "doc": { + "question": "How many video clips does the UCF-101 dataset contain?", + "choices": [ + "13,320", + "35,000", + "70 million", + "2,265" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:How many video clips does the UCF-101 dataset contain?\nAnswer:", + " 13,320" + ], + [ + "Question:How many video clips does the UCF-101 dataset contain?\nAnswer:", + " 35,000" + ], + [ + "Question:How many video clips does the UCF-101 dataset contain?\nAnswer:", + " 70 million" + ], + [ + "Question:How many video clips does the UCF-101 dataset contain?\nAnswer:", + " 2,265" + ] + ], + "resps": [ + [ + [ + -15.38673210144043, + false + ] + ], + [ + [ + -13.665287017822266, + false + ] + ], + [ + [ + -18.272319793701172, + false + ] + ], + [ + [ + -14.839933395385742, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.38673210144043, + false + ], + [ + -13.665287017822266, + false + ], + [ + -18.272319793701172, + false + ], + [ + -14.839933395385742, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d918d9290ba27f5972ce763142840a34b15fccae886c519a5ff068af2d982b0b", + "prompt_hash": "66610f22552921f853fe59e1ea7704b01e3c69cbffae9d2feca90e1b0a4a4e1e", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 1, + "doc": { + "question": "What is the average video length in the WebVid-10M dataset?", + "choices": [ + "7.2 seconds", + "18.7 seconds", + "11.4 seconds", + "8.5 seconds" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the average video length in the WebVid-10M dataset?\nAnswer:", + " 7.2 seconds" + ], + [ + "Question:What is the average video length in the WebVid-10M dataset?\nAnswer:", + " 18.7 seconds" + ], + [ + "Question:What is the average video length in the WebVid-10M dataset?\nAnswer:", + " 11.4 seconds" + ], + [ + "Question:What is the average video length in the WebVid-10M dataset?\nAnswer:", + " 8.5 seconds" + ] + ], + "resps": [ + [ + [ + -9.703564643859863, + false + ] + ], + [ + [ + -10.211236953735352, + false + ] + ], + [ + [ + -9.757585525512695, + false + ] + ], + [ + [ + -9.347663879394531, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.703564643859863, + false + ], + [ + -10.211236953735352, + false + ], + [ + -9.757585525512695, + false + ], + [ + -9.347663879394531, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0c051f64404c22a332292565deb4bcc0138c49b138d6b18679c18b46e9f01084", + "prompt_hash": "93266376b337b79147290d8bc60a6240e38a5e688ef9c7e34717db3df76c55fe", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 2, + "doc": { + "question": "Which dataset contains 2,265 high-quality time-lapse videos with text descriptions?", + "choices": [ + "Taichi-HD", + "ChronoMagic", + "CelebvHQ", + "Panda-70M" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset contains 2,265 high-quality time-lapse videos with text descriptions?\nAnswer:", + " Taichi-HD" + ], + [ + "Question:Which dataset contains 2,265 high-quality time-lapse videos with text descriptions?\nAnswer:", + " ChronoMagic" + ], + [ + "Question:Which dataset contains 2,265 high-quality time-lapse videos with text descriptions?\nAnswer:", + " CelebvHQ" + ], + [ + "Question:Which dataset contains 2,265 high-quality time-lapse videos with text descriptions?\nAnswer:", + " Panda-70M" + ] + ], + "resps": [ + [ + [ + -26.884653091430664, + false + ] + ], + [ + [ + -30.02543830871582, + false + ] + ], + [ + [ + -24.76498794555664, + false + ] + ], + [ + [ + -22.008342742919922, + false + ] + ] + ], + "filtered_resps": [ + [ + -26.884653091430664, + false + ], + [ + -30.02543830871582, + false + ], + [ + -24.76498794555664, + false + ], + [ + -22.008342742919922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "aea8b1e5f35964a5f6bd80bb657c379c0763ba1b2dd00311ea6d8d1a834f7f6b", + "prompt_hash": "3a25697ba9302d309fefbd23edef13e2233a48d1caa5aa842f6dae2e8c6168cc", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 3, + "doc": { + "question": "How many video clips and hours does the InternVid dataset include?", + "choices": [ + "13,320 clips and 2.7 hours", + "2,265 clips and 7 hours", + "234 million clips and 760.3K hours", + "35,000 clips and 65 hours" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many video clips and hours does the InternVid dataset include?\nAnswer:", + " 13,320 clips and 2.7 hours" + ], + [ + "Question:How many video clips and hours does the InternVid dataset include?\nAnswer:", + " 2,265 clips and 7 hours" + ], + [ + "Question:How many video clips and hours does the InternVid dataset include?\nAnswer:", + " 234 million clips and 760.3K hours" + ], + [ + "Question:How many video clips and hours does the InternVid dataset include?\nAnswer:", + " 35,000 clips and 65 hours" + ] + ], + "resps": [ + [ + [ + -23.848697662353516, + false + ] + ], + [ + [ + -21.163166046142578, + false + ] + ], + [ + [ + -40.58928298950195, + false + ] + ], + [ + [ + -19.135757446289062, + false + ] + ] + ], + "filtered_resps": [ + [ + -23.848697662353516, + false + ], + [ + -21.163166046142578, + false + ], + [ + -40.58928298950195, + false + ], + [ + -19.135757446289062, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7ccc4b4a88a7e240547fd23c9a7bed6ef3a596e323340ff9f56c7e337aeffb38", + "prompt_hash": "6d123085222da44784840c0b919186526daf6426113cc555b4d15ce2daece6af", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 4, + "doc": { + "question": "What resolution do the videos in the \\u201chd-dataset\\u201d (HD version of the new dataset) achieve?", + "choices": [ + "512×512", + "1920×1080", + "640×360", + "1280×720" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What resolution do the videos in the \\u201chd-dataset\\u201d (HD version of the new dataset) achieve?\nAnswer:", + " 512×512" + ], + [ + "Question:What resolution do the videos in the \\u201chd-dataset\\u201d (HD version of the new dataset) achieve?\nAnswer:", + " 1920×1080" + ], + [ + "Question:What resolution do the videos in the \\u201chd-dataset\\u201d (HD version of the new dataset) achieve?\nAnswer:", + " 640×360" + ], + [ + "Question:What resolution do the videos in the \\u201chd-dataset\\u201d (HD version of the new dataset) achieve?\nAnswer:", + " 1280×720" + ] + ], + "resps": [ + [ + [ + -9.11259651184082, + false + ] + ], + [ + [ + -6.933300018310547, + false + ] + ], + [ + [ + -9.924569129943848, + false + ] + ], + [ + [ + -8.05185317993164, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.11259651184082, + false + ], + [ + -6.933300018310547, + false + ], + [ + -9.924569129943848, + false + ], + [ + -8.05185317993164, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "68595c0f4ac35ddb77399e10113b20bcee003b37dadc9b9416f1f86c43cdbf6a", + "prompt_hash": "9ff2435dd70469a91ad50c31a8a09ec1e749efb2b188ee18f7e70ae23854d4f9", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 5, + "doc": { + "question": "Which dataset was used as the retain dataset for harmful request refusal training?", + "choices": [ + "Magpie-Align", + "Wikitext", + "Camel-bio", + "CTFtime" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which dataset was used as the retain dataset for harmful request refusal training?\nAnswer:", + " Magpie-Align" + ], + [ + "Question:Which dataset was used as the retain dataset for harmful request refusal training?\nAnswer:", + " Wikitext" + ], + [ + "Question:Which dataset was used as the retain dataset for harmful request refusal training?\nAnswer:", + " Camel-bio" + ], + [ + "Question:Which dataset was used as the retain dataset for harmful request refusal training?\nAnswer:", + " CTFtime" + ] + ], + "resps": [ + [ + [ + -30.949909210205078, + false + ] + ], + [ + [ + -12.598445892333984, + false + ] + ], + [ + [ + -29.680171966552734, + false + ] + ], + [ + [ + -28.077442169189453, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.949909210205078, + false + ], + [ + -12.598445892333984, + false + ], + [ + -29.680171966552734, + false + ], + [ + -28.077442169189453, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7c7ab492f9d6fb841340d642fa8084c5abd581c141445b0e25816877a193aa7d", + "prompt_hash": "73b6ceeaa2d9583cf65a3e36a39c36cade024ff3a12289eae9f7c7422f18af51", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 6, + "doc": { + "question": "What optimizer was used for train-time adversaries in the Biosecurity weaponization restriction setup?", + "choices": [ + "AdamW", + "SGD with Nesterov Momentum", + "Adadelta", + "Schedule Free AdamW" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What optimizer was used for train-time adversaries in the Biosecurity weaponization restriction setup?\nAnswer:", + " AdamW" + ], + [ + "Question:What optimizer was used for train-time adversaries in the Biosecurity weaponization restriction setup?\nAnswer:", + " SGD with Nesterov Momentum" + ], + [ + "Question:What optimizer was used for train-time adversaries in the Biosecurity weaponization restriction setup?\nAnswer:", + " Adadelta" + ], + [ + "Question:What optimizer was used for train-time adversaries in the Biosecurity weaponization restriction setup?\nAnswer:", + " Schedule Free AdamW" + ] + ], + "resps": [ + [ + [ + -4.982319355010986, + false + ] + ], + [ + [ + -16.44598388671875, + false + ] + ], + [ + [ + -7.690619945526123, + false + ] + ], + [ + [ + -31.53502655029297, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.982319355010986, + false + ], + [ + -16.44598388671875, + false + ], + [ + -7.690619945526123, + false + ], + [ + -31.53502655029297, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f8e9a563d381fb960aa6e786dad81b874336a607396c0768ee26742c75b881d4", + "prompt_hash": "3c661c4dc3ad9162e3993d2761aa2ee773b88ac8b9353087f14c1579f1711327", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 7, + "doc": { + "question": "How many adversary setups were simulated for Cybersecurity weaponization restriction during training?", + "choices": [ + "2", + "4", + "6", + "8" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many adversary setups were simulated for Cybersecurity weaponization restriction during training?\nAnswer:", + " 2" + ], + [ + "Question:How many adversary setups were simulated for Cybersecurity weaponization restriction during training?\nAnswer:", + " 4" + ], + [ + "Question:How many adversary setups were simulated for Cybersecurity weaponization restriction during training?\nAnswer:", + " 6" + ], + [ + "Question:How many adversary setups were simulated for Cybersecurity weaponization restriction during training?\nAnswer:", + " 8" + ] + ], + "resps": [ + [ + [ + -3.0547232627868652, + false + ] + ], + [ + [ + -3.540640354156494, + false + ] + ], + [ + [ + -4.074788570404053, + false + ] + ], + [ + [ + -4.373826503753662, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.0547232627868652, + false + ], + [ + -3.540640354156494, + false + ], + [ + -4.074788570404053, + false + ], + [ + -4.373826503753662, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b5e1ccec8cfe6f15e81a591cf22a589bb992cfee4e754932728adb85e0877637", + "prompt_hash": "64f0040af9d1edaa68dde03333b2d294fd6c95538f6dfd8acfabd6a86b688086", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 8, + "doc": { + "question": "In the additional harmful request refusal results, what was the Post-Attack ASR of the TAR method?", + "choices": [ + "84.8", + "78.3", + "74.5", + "63.9" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:In the additional harmful request refusal results, what was the Post-Attack ASR of the TAR method?\nAnswer:", + " 84.8" + ], + [ + "Question:In the additional harmful request refusal results, what was the Post-Attack ASR of the TAR method?\nAnswer:", + " 78.3" + ], + [ + "Question:In the additional harmful request refusal results, what was the Post-Attack ASR of the TAR method?\nAnswer:", + " 74.5" + ], + [ + "Question:In the additional harmful request refusal results, what was the Post-Attack ASR of the TAR method?\nAnswer:", + " 63.9" + ] + ], + "resps": [ + [ + [ + -9.289198875427246, + false + ] + ], + [ + [ + -9.459244728088379, + false + ] + ], + [ + [ + -9.550447463989258, + false + ] + ], + [ + [ + -9.714252471923828, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.289198875427246, + false + ], + [ + -9.459244728088379, + false + ], + [ + -9.550447463989258, + false + ], + [ + -9.714252471923828, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "785f5608cc184077484fbf0dac845374b6161aa05d6f07c68f282d09fcc69fe8", + "prompt_hash": "42b34dfe5d13cad8f57f2d25f673ef255ed327321f125c23443d82caa6cbdd0f", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 9, + "doc": { + "question": "Which dataset was scraped to construct the Cybersecurity forget dataset?", + "choices": [ + "Camel AI Biology", + "CTFtime writeups", + "WMDP", + "Pile-bio Retain" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which dataset was scraped to construct the Cybersecurity forget dataset?\nAnswer:", + " Camel AI Biology" + ], + [ + "Question:Which dataset was scraped to construct the Cybersecurity forget dataset?\nAnswer:", + " CTFtime writeups" + ], + [ + "Question:Which dataset was scraped to construct the Cybersecurity forget dataset?\nAnswer:", + " WMDP" + ], + [ + "Question:Which dataset was scraped to construct the Cybersecurity forget dataset?\nAnswer:", + " Pile-bio Retain" + ] + ], + "resps": [ + [ + [ + -36.0738410949707, + false + ] + ], + [ + [ + -41.3992919921875, + false + ] + ], + [ + [ + -16.297882080078125, + false + ] + ], + [ + [ + -48.197181701660156, + false + ] + ] + ], + "filtered_resps": [ + [ + -36.0738410949707, + false + ], + [ + -41.3992919921875, + false + ], + [ + -16.297882080078125, + false + ], + [ + -48.197181701660156, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "9b5d1e8c7027ddc744eeace2cc139a04af92ebd1bf2b40858abd6a1e04cbfc6e", + "prompt_hash": "47bbc435192555e617c67175eb3a4de72282016369dd827a64650787898b3e71", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 10, + "doc": { + "question": "In the Last Letter Concatenation task, how many problem queries are generated for each input length?", + "choices": [ + "128", + "350", + "4096", + "7200" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:In the Last Letter Concatenation task, how many problem queries are generated for each input length?\nAnswer:", + " 128" + ], + [ + "Question:In the Last Letter Concatenation task, how many problem queries are generated for each input length?\nAnswer:", + " 350" + ], + [ + "Question:In the Last Letter Concatenation task, how many problem queries are generated for each input length?\nAnswer:", + " 4096" + ], + [ + "Question:In the Last Letter Concatenation task, how many problem queries are generated for each input length?\nAnswer:", + " 7200" + ] + ], + "resps": [ + [ + [ + -8.70600414276123, + false + ] + ], + [ + [ + -10.003148078918457, + false + ] + ], + [ + [ + -10.063257217407227, + false + ] + ], + [ + [ + -13.331168174743652, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.70600414276123, + false + ], + [ + -10.003148078918457, + false + ], + [ + -10.063257217407227, + false + ], + [ + -13.331168174743652, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b48a19f02f6e3781a3960dc5a958e312b6e2457f2179f5a77d9fca4adc6283b7", + "prompt_hash": "39737723d57a40e7a488825a3a9245f925f737b612e3215a668bbb522bc1c2c0", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 11, + "doc": { + "question": "For the Word Sorting task, how many lists of words are generated as problem queries for each input length?", + "choices": [ + "128", + "350", + "4096", + "7200" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:For the Word Sorting task, how many lists of words are generated as problem queries for each input length?\nAnswer:", + " 128" + ], + [ + "Question:For the Word Sorting task, how many lists of words are generated as problem queries for each input length?\nAnswer:", + " 350" + ], + [ + "Question:For the Word Sorting task, how many lists of words are generated as problem queries for each input length?\nAnswer:", + " 4096" + ], + [ + "Question:For the Word Sorting task, how many lists of words are generated as problem queries for each input length?\nAnswer:", + " 7200" + ] + ], + "resps": [ + [ + [ + -9.128795623779297, + false + ] + ], + [ + [ + -10.338838577270508, + false + ] + ], + [ + [ + -10.242339134216309, + false + ] + ], + [ + [ + -13.073063850402832, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.128795623779297, + false + ], + [ + -10.338838577270508, + false + ], + [ + -10.242339134216309, + false + ], + [ + -13.073063850402832, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7d4e66b5e39a6f94c63b31e4990ba314e9cfc5c45bebbb1ac8207ba01fb1b0a3", + "prompt_hash": "8f3960d8166d21235488fa33ac5c46b15b7149d8c82186bf7886d3734e4f2c14", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 12, + "doc": { + "question": "In the Grade School Math dataset, approximately how many problems are used for training?", + "choices": [ + "1.3K", + "128", + "7.2K", + "50K" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:In the Grade School Math dataset, approximately how many problems are used for training?\nAnswer:", + " 1.3K" + ], + [ + "Question:In the Grade School Math dataset, approximately how many problems are used for training?\nAnswer:", + " 128" + ], + [ + "Question:In the Grade School Math dataset, approximately how many problems are used for training?\nAnswer:", + " 7.2K" + ], + [ + "Question:In the Grade School Math dataset, approximately how many problems are used for training?\nAnswer:", + " 50K" + ] + ], + "resps": [ + [ + [ + -11.14500617980957, + false + ] + ], + [ + [ + -7.693936824798584, + false + ] + ], + [ + [ + -12.409635543823242, + false + ] + ], + [ + [ + -8.355731964111328, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.14500617980957, + false + ], + [ + -7.693936824798584, + false + ], + [ + -12.409635543823242, + false + ], + [ + -8.355731964111328, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ddfcc10d4cb72f3178246031e500bd4e67b7661425f254523ce17198c9b51dbf", + "prompt_hash": "4db725a7777f9e61ed2d46829e43cf32250785dd099802a22a78ade8030be237", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 13, + "doc": { + "question": "Which optimizer is used for training Gemma-based generative verifiers?", + "choices": [ + "SGD", + "Adam", + "RMSprop", + "Adagrad" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which optimizer is used for training Gemma-based generative verifiers?\nAnswer:", + " SGD" + ], + [ + "Question:Which optimizer is used for training Gemma-based generative verifiers?\nAnswer:", + " Adam" + ], + [ + "Question:Which optimizer is used for training Gemma-based generative verifiers?\nAnswer:", + " RMSprop" + ], + [ + "Question:Which optimizer is used for training Gemma-based generative verifiers?\nAnswer:", + " Adagrad" + ] + ], + "resps": [ + [ + [ + -7.656381607055664, + false + ] + ], + [ + [ + -4.861548900604248, + false + ] + ], + [ + [ + -10.62295913696289, + false + ] + ], + [ + [ + -10.300954818725586, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.656381607055664, + false + ], + [ + -4.861548900604248, + false + ], + [ + -10.62295913696289, + false + ], + [ + -10.300954818725586, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b4f41f766d1df5f2e48170794a9089c321f14247018cfe937b62066953b701a2", + "prompt_hash": "9363cb97c6af5e29ff7a793c58c6291445cbac5ad99bf7c35222bfdadc47b50d", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 14, + "doc": { + "question": "What learning rate worked best for DPO training in the described experiments?", + "choices": [ + "1e-7", + "5e-7", + "1e-6", + "2e-6" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What learning rate worked best for DPO training in the described experiments?\nAnswer:", + " 1e-7" + ], + [ + "Question:What learning rate worked best for DPO training in the described experiments?\nAnswer:", + " 5e-7" + ], + [ + "Question:What learning rate worked best for DPO training in the described experiments?\nAnswer:", + " 1e-6" + ], + [ + "Question:What learning rate worked best for DPO training in the described experiments?\nAnswer:", + " 2e-6" + ] + ], + "resps": [ + [ + [ + -9.123523712158203, + false + ] + ], + [ + [ + -9.504151344299316, + false + ] + ], + [ + [ + -8.128350257873535, + false + ] + ], + [ + [ + -9.45744514465332, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.123523712158203, + false + ], + [ + -9.504151344299316, + false + ], + [ + -8.128350257873535, + false + ], + [ + -9.45744514465332, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "aa2f1f819abc173eee2d667708f63bbd94114eceb9eb24a99d3ca9a048cac5e7", + "prompt_hash": "afa0657c4d865c63c194aeb264649949e04b109fc2a251da1cbb8288f6d2c3c3", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 15, + "doc": { + "question": "What is the name of the pre-trained visual encoder developed in the Oryx model?", + "choices": [ + "OryxViT", + "CLIP", + "NaViT", + "FlexiViT" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the name of the pre-trained visual encoder developed in the Oryx model?\nAnswer:", + " OryxViT" + ], + [ + "Question:What is the name of the pre-trained visual encoder developed in the Oryx model?\nAnswer:", + " CLIP" + ], + [ + "Question:What is the name of the pre-trained visual encoder developed in the Oryx model?\nAnswer:", + " NaViT" + ], + [ + "Question:What is the name of the pre-trained visual encoder developed in the Oryx model?\nAnswer:", + " FlexiViT" + ] + ], + "resps": [ + [ + [ + -9.803050994873047, + false + ] + ], + [ + [ + -3.1500496864318848, + false + ] + ], + [ + [ + -17.866470336914062, + false + ] + ], + [ + [ + -14.184885025024414, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.803050994873047, + false + ], + [ + -3.1500496864318848, + false + ], + [ + -17.866470336914062, + false + ], + [ + -14.184885025024414, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "156b7b840777a70e8b642ac95c6705845ad2ff391c9a995ca447269ee8435f98", + "prompt_hash": "68d2ed3788c0d39dd7228baf9d11482275475f353ed20d41c46b3cab10d97490", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 16, + "doc": { + "question": "Which dataset was used to prepare long-form temporal training data with 'needle-in-a-haystack' tasks?", + "choices": [ + "MovieNet", + "ScanQA", + "LLaVA-NeXt", + "Cambrian-1" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which dataset was used to prepare long-form temporal training data with 'needle-in-a-haystack' tasks?\nAnswer:", + " MovieNet" + ], + [ + "Question:Which dataset was used to prepare long-form temporal training data with 'needle-in-a-haystack' tasks?\nAnswer:", + " ScanQA" + ], + [ + "Question:Which dataset was used to prepare long-form temporal training data with 'needle-in-a-haystack' tasks?\nAnswer:", + " LLaVA-NeXt" + ], + [ + "Question:Which dataset was used to prepare long-form temporal training data with 'needle-in-a-haystack' tasks?\nAnswer:", + " Cambrian-1" + ] + ], + "resps": [ + [ + [ + -8.061107635498047, + false + ] + ], + [ + [ + -16.233022689819336, + false + ] + ], + [ + [ + -17.29337501525879, + false + ] + ], + [ + [ + -22.008586883544922, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.061107635498047, + false + ], + [ + -16.233022689819336, + false + ], + [ + -17.29337501525879, + false + ], + [ + -22.008586883544922, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "eba15b55f0b7a7094523969a778e1f9adb1ad8119d2cf8253c2134b659baa96a", + "prompt_hash": "849ce8d2eb09cbeba5b325412eb5b336286ebbc1f4864858b8569f2b94124d5a", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 17, + "doc": { + "question": "Which tracking model was used to generate coarse correspondences for the ScanQA training set?", + "choices": [ + "Track-Anything", + "NaViT", + "SigLIP", + "FlexiViT" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which tracking model was used to generate coarse correspondences for the ScanQA training set?\nAnswer:", + " Track-Anything" + ], + [ + "Question:Which tracking model was used to generate coarse correspondences for the ScanQA training set?\nAnswer:", + " NaViT" + ], + [ + "Question:Which tracking model was used to generate coarse correspondences for the ScanQA training set?\nAnswer:", + " SigLIP" + ], + [ + "Question:Which tracking model was used to generate coarse correspondences for the ScanQA training set?\nAnswer:", + " FlexiViT" + ] + ], + "resps": [ + [ + [ + -11.741609573364258, + false + ] + ], + [ + [ + -14.81590747833252, + false + ] + ], + [ + [ + -11.319101333618164, + false + ] + ], + [ + [ + -23.714706420898438, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.741609573364258, + false + ], + [ + -14.81590747833252, + false + ], + [ + -11.319101333618164, + false + ], + [ + -23.714706420898438, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "de765beb9e0c5abd06ea1512e4806499af83248f2e9e33ce3f70050ee8c92677", + "prompt_hash": "b3c3c00f007e36c1b86ce3a18f4adf255d055ea83561ca0de94f5c3f37d1bb27", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 18, + "doc": { + "question": "In Stage 1 of the Oryx training pipeline, approximately how many supervised fine-tuning image-text pairs were collected?", + "choices": [ + "4 million", + "400 thousand", + "40 million", + "1 million" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:In Stage 1 of the Oryx training pipeline, approximately how many supervised fine-tuning image-text pairs were collected?\nAnswer:", + " 4 million" + ], + [ + "Question:In Stage 1 of the Oryx training pipeline, approximately how many supervised fine-tuning image-text pairs were collected?\nAnswer:", + " 400 thousand" + ], + [ + "Question:In Stage 1 of the Oryx training pipeline, approximately how many supervised fine-tuning image-text pairs were collected?\nAnswer:", + " 40 million" + ], + [ + "Question:In Stage 1 of the Oryx training pipeline, approximately how many supervised fine-tuning image-text pairs were collected?\nAnswer:", + " 1 million" + ] + ], + "resps": [ + [ + [ + -9.426055908203125, + false + ] + ], + [ + [ + -13.373111724853516, + false + ] + ], + [ + [ + -9.889270782470703, + false + ] + ], + [ + [ + -8.437106132507324, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.426055908203125, + false + ], + [ + -13.373111724853516, + false + ], + [ + -9.889270782470703, + false + ], + [ + -8.437106132507324, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "566b2aac0784a025c3eb826a286d646e61de8d7cecbede2afd0985094ec48415", + "prompt_hash": "fe109bf7a63cea191cc404f1cfd2cae1cc1277fe4036eaf0e2cb64c20daeab38", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 19, + "doc": { + "question": "Which benchmark did the Oryx-1.5-32B model outperform GPT-4o by 7.7%?", + "choices": [ + "MLVU", + "LongVideoBench", + "NextQA", + "OCRBench" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which benchmark did the Oryx-1.5-32B model outperform GPT-4o by 7.7%?\nAnswer:", + " MLVU" + ], + [ + "Question:Which benchmark did the Oryx-1.5-32B model outperform GPT-4o by 7.7%?\nAnswer:", + " LongVideoBench" + ], + [ + "Question:Which benchmark did the Oryx-1.5-32B model outperform GPT-4o by 7.7%?\nAnswer:", + " NextQA" + ], + [ + "Question:Which benchmark did the Oryx-1.5-32B model outperform GPT-4o by 7.7%?\nAnswer:", + " OCRBench" + ] + ], + "resps": [ + [ + [ + -20.46422576904297, + false + ] + ], + [ + [ + -21.4570255279541, + false + ] + ], + [ + [ + -15.763558387756348, + false + ] + ], + [ + [ + -15.791434288024902, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.46422576904297, + false + ], + [ + -21.4570255279541, + false + ], + [ + -15.763558387756348, + false + ], + [ + -15.791434288024902, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "929aedefb09734275f0ac9ad7b47f7a727f751bc2aed619a363d3479e01e1ac6", + "prompt_hash": "c35ffd77c1af1f3323947385684544d64e854cc292267e354343a7e9cc11c5dd", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 20, + "doc": { + "question": "Which activation function is used in the point-wise tokenization embedding for time series points?", + "choices": [ + "ReLU", + "SwiGLU", + "Tanh", + "Sigmoid" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which activation function is used in the point-wise tokenization embedding for time series points?\nAnswer:", + " ReLU" + ], + [ + "Question:Which activation function is used in the point-wise tokenization embedding for time series points?\nAnswer:", + " SwiGLU" + ], + [ + "Question:Which activation function is used in the point-wise tokenization embedding for time series points?\nAnswer:", + " Tanh" + ], + [ + "Question:Which activation function is used in the point-wise tokenization embedding for time series points?\nAnswer:", + " Sigmoid" + ] + ], + "resps": [ + [ + [ + -3.263493776321411, + false + ] + ], + [ + [ + -7.855595588684082, + false + ] + ], + [ + [ + -4.75715970993042, + false + ] + ], + [ + [ + -4.864443302154541, + false + ] + ] + ], + "filtered_resps": [ + [ + -3.263493776321411, + false + ], + [ + -7.855595588684082, + false + ], + [ + -4.75715970993042, + false + ], + [ + -4.864443302154541, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "dcd247f6043e0192a6a4ab1861e955946c93c6d730b7eb13d08215173969bba0", + "prompt_hash": "1372eb375cb978e2bf04c001594c7612e60f838b77d77a7ae699921ef3940ec3", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 21, + "doc": { + "question": "What type of positional embedding is adopted instead of absolute positional encoding?", + "choices": [ + "Sinusoidal positional encoding", + "Rotary positional embeddings", + "Learned positional embeddings", + "Relative positional embeddings" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What type of positional embedding is adopted instead of absolute positional encoding?\nAnswer:", + " Sinusoidal positional encoding" + ], + [ + "Question:What type of positional embedding is adopted instead of absolute positional encoding?\nAnswer:", + " Rotary positional embeddings" + ], + [ + "Question:What type of positional embedding is adopted instead of absolute positional encoding?\nAnswer:", + " Learned positional embeddings" + ], + [ + "Question:What type of positional embedding is adopted instead of absolute positional encoding?\nAnswer:", + " Relative positional embeddings" + ] + ], + "resps": [ + [ + [ + -12.207653999328613, + false + ] + ], + [ + [ + -13.981444358825684, + false + ] + ], + [ + [ + -14.838812828063965, + false + ] + ], + [ + [ + -9.279513359069824, + false + ] + ] + ], + "filtered_resps": [ + [ + -12.207653999328613, + false + ], + [ + -13.981444358825684, + false + ], + [ + -14.838812828063965, + false + ], + [ + -9.279513359069824, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c0c0237d53201c05ca5446b5272bb12d88287d6e79e64c304a49e708b209080f", + "prompt_hash": "7e0afa984471c4a497ec57593d7a8f8d51751bad8017b324dfc9344edc4d5123", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 22, + "doc": { + "question": "How many time points does the constructed dataset include after processing?", + "choices": [ + "3 billion", + "30 billion", + "300 billion", + "3 trillion" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many time points does the constructed dataset include after processing?\nAnswer:", + " 3 billion" + ], + [ + "Question:How many time points does the constructed dataset include after processing?\nAnswer:", + " 30 billion" + ], + [ + "Question:How many time points does the constructed dataset include after processing?\nAnswer:", + " 300 billion" + ], + [ + "Question:How many time points does the constructed dataset include after processing?\nAnswer:", + " 3 trillion" + ] + ], + "resps": [ + [ + [ + -16.558652877807617, + false + ] + ], + [ + [ + -17.89537811279297, + false + ] + ], + [ + [ + -18.499900817871094, + false + ] + ], + [ + [ + -16.72484016418457, + false + ] + ] + ], + "filtered_resps": [ + [ + -16.558652877807617, + false + ], + [ + -17.89537811279297, + false + ], + [ + -18.499900817871094, + false + ], + [ + -16.72484016418457, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "226e991f49c1d5fd3206a0a0a278cc3dc98554f228cb4e67b2b2ba2fe49d96c5", + "prompt_hash": "eaa8068334e6738a83830dd69c2d18959f2e120641de642a72f068823ca70be6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 23, + "doc": { + "question": "Which loss function is used to improve robustness against outliers in training?", + "choices": [ + "Cross-entropy loss", + "MSE loss", + "Huber loss", + "KL-divergence loss" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which loss function is used to improve robustness against outliers in training?\nAnswer:", + " Cross-entropy loss" + ], + [ + "Question:Which loss function is used to improve robustness against outliers in training?\nAnswer:", + " MSE loss" + ], + [ + "Question:Which loss function is used to improve robustness against outliers in training?\nAnswer:", + " Huber loss" + ], + [ + "Question:Which loss function is used to improve robustness against outliers in training?\nAnswer:", + " KL-divergence loss" + ] + ], + "resps": [ + [ + [ + -7.551655292510986, + false + ] + ], + [ + [ + -9.450379371643066, + false + ] + ], + [ + [ + -10.613362312316895, + false + ] + ], + [ + [ + -11.747757911682129, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.551655292510986, + false + ], + [ + -9.450379371643066, + false + ], + [ + -10.613362312316895, + false + ], + [ + -11.747757911682129, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c528078998f2fb64b3e49b759f52aa67bf869dfc58966f3a6c80b96a53d35824", + "prompt_hash": "228d572ea74fd9a40d2de41cf02d95b737734e128b282417db9a0f171f561923", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 24, + "doc": { + "question": "What is the maximum sequence length used during training?", + "choices": [ + "1024", + "2048", + "4096", + "8192" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the maximum sequence length used during training?\nAnswer:", + " 1024" + ], + [ + "Question:What is the maximum sequence length used during training?\nAnswer:", + " 2048" + ], + [ + "Question:What is the maximum sequence length used during training?\nAnswer:", + " 4096" + ], + [ + "Question:What is the maximum sequence length used during training?\nAnswer:", + " 8192" + ] + ], + "resps": [ + [ + [ + -4.160928726196289, + false + ] + ], + [ + [ + -5.0833048820495605, + false + ] + ], + [ + [ + -5.026993751525879, + false + ] + ], + [ + [ + -5.385982036590576, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.160928726196289, + false + ], + [ + -5.0833048820495605, + false + ], + [ + -5.026993751525879, + false + ], + [ + -5.385982036590576, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "390e1da76812662b20880e121591fd850a1a17f394ad7930d5eb4dff5289720e", + "prompt_hash": "a4432fe7d5f31174a59b3e50b838f427cb5e616f76805fb50d00d7116ba27897", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 25, + "doc": { + "question": "How many Kaggle competitions are included in MLE-bench?", + "choices": [ + "50", + "75", + "100", + "150" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many Kaggle competitions are included in MLE-bench?\nAnswer:", + " 50" + ], + [ + "Question:How many Kaggle competitions are included in MLE-bench?\nAnswer:", + " 75" + ], + [ + "Question:How many Kaggle competitions are included in MLE-bench?\nAnswer:", + " 100" + ], + [ + "Question:How many Kaggle competitions are included in MLE-bench?\nAnswer:", + " 150" + ] + ], + "resps": [ + [ + [ + -6.944413185119629, + false + ] + ], + [ + [ + -8.142876625061035, + false + ] + ], + [ + [ + -7.314369201660156, + false + ] + ], + [ + [ + -9.207300186157227, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.944413185119629, + false + ], + [ + -8.142876625061035, + false + ], + [ + -7.314369201660156, + false + ], + [ + -9.207300186157227, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "052584f68f0b2428e10a00891e5d2883b043bf7950c1da756deb1605ef48fc53", + "prompt_hash": "66e4628c81fbdc9d2ae738e93618091ef84479831f56b4c8e0ed12e03c356ae5", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 26, + "doc": { + "question": "What was the total value of prizes awarded across the 75 competitions in MLE-bench?", + "choices": [ + "$1,000,000", + "$1,948,016", + "$2,500,000", + "$5,000,000" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What was the total value of prizes awarded across the 75 competitions in MLE-bench?\nAnswer:", + " $1,000,000" + ], + [ + "Question:What was the total value of prizes awarded across the 75 competitions in MLE-bench?\nAnswer:", + " $1,948,016" + ], + [ + "Question:What was the total value of prizes awarded across the 75 competitions in MLE-bench?\nAnswer:", + " $2,500,000" + ], + [ + "Question:What was the total value of prizes awarded across the 75 competitions in MLE-bench?\nAnswer:", + " $5,000,000" + ] + ], + "resps": [ + [ + [ + -8.74365234375, + false + ] + ], + [ + [ + -20.456871032714844, + false + ] + ], + [ + [ + -10.962664604187012, + false + ] + ], + [ + [ + -10.215169906616211, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.74365234375, + false + ], + [ + -20.456871032714844, + false + ], + [ + -10.962664604187012, + false + ], + [ + -10.215169906616211, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "86ca1ab5e80d755d678b903acba040813918e585858b1b912f6eedb3a3b34d59", + "prompt_hash": "8b7c80469eced0452b5db403fb4d86ea98b8a11306bd639158ddcc4dfeb3140f", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 27, + "doc": { + "question": "Which model with the AIDE scaffold achieved a medal in 16.9% of competitions on MLE-bench?", + "choices": [ + "GPT-4o", + "Claude 3.5 Sonnet", + "o1-preview", + "Llama 3.1 405B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model with the AIDE scaffold achieved a medal in 16.9% of competitions on MLE-bench?\nAnswer:", + " GPT-4o" + ], + [ + "Question:Which model with the AIDE scaffold achieved a medal in 16.9% of competitions on MLE-bench?\nAnswer:", + " Claude 3.5 Sonnet" + ], + [ + "Question:Which model with the AIDE scaffold achieved a medal in 16.9% of competitions on MLE-bench?\nAnswer:", + " o1-preview" + ], + [ + "Question:Which model with the AIDE scaffold achieved a medal in 16.9% of competitions on MLE-bench?\nAnswer:", + " Llama 3.1 405B" + ] + ], + "resps": [ + [ + [ + -9.823843002319336, + false + ] + ], + [ + [ + -12.214433670043945, + false + ] + ], + [ + [ + -29.94273567199707, + false + ] + ], + [ + [ + -38.722164154052734, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.823843002319336, + false + ], + [ + -12.214433670043945, + false + ], + [ + -29.94273567199707, + false + ], + [ + -38.722164154052734, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "8aa00e51fa2ef433fd813fa8e76efe0379871555e0115b5953bfb69dc498c845", + "prompt_hash": "ae6765ca651d6c2d2afad9f8a23da90dc99b9b8fc63d5780eb2380285524018f", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 28, + "doc": { + "question": "What is the percentage of competitions in MLE-bench categorized as Medium complexity?", + "choices": [ + "20%", + "30%", + "50%", + "70%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the percentage of competitions in MLE-bench categorized as Medium complexity?\nAnswer:", + " 20%" + ], + [ + "Question:What is the percentage of competitions in MLE-bench categorized as Medium complexity?\nAnswer:", + " 30%" + ], + [ + "Question:What is the percentage of competitions in MLE-bench categorized as Medium complexity?\nAnswer:", + " 50%" + ], + [ + "Question:What is the percentage of competitions in MLE-bench categorized as Medium complexity?\nAnswer:", + " 70%" + ] + ], + "resps": [ + [ + [ + -8.905912399291992, + false + ] + ], + [ + [ + -9.422713279724121, + false + ] + ], + [ + [ + -9.525880813598633, + false + ] + ], + [ + [ + -10.475693702697754, + false + ] + ] + ], + "filtered_resps": [ + [ + -8.905912399291992, + false + ], + [ + -9.422713279724121, + false + ], + [ + -9.525880813598633, + false + ], + [ + -10.475693702697754, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ce97fde4984d268d82cf472e616e69071fe67efd92522924494fca607ef02b45", + "prompt_hash": "295e3ced57c305763394eac8e83129b55dbc39b4d42cc18b9ba35bbd27ea2297", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 29, + "doc": { + "question": "What type of hardware setup was provided to agents in the main experiments?", + "choices": [ + "8 vCPUs, 128GB RAM, one Nvidia P100 GPU", + "36 vCPUs, 440GB RAM, one Nvidia A10 GPU", + "64 vCPUs, 1TB RAM, two Nvidia A100 GPUs", + "16 vCPUs, 256GB RAM, CPU-only" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What type of hardware setup was provided to agents in the main experiments?\nAnswer:", + " 8 vCPUs, 128GB RAM, one Nvidia P100 GPU" + ], + [ + "Question:What type of hardware setup was provided to agents in the main experiments?\nAnswer:", + " 36 vCPUs, 440GB RAM, one Nvidia A10 GPU" + ], + [ + "Question:What type of hardware setup was provided to agents in the main experiments?\nAnswer:", + " 64 vCPUs, 1TB RAM, two Nvidia A100 GPUs" + ], + [ + "Question:What type of hardware setup was provided to agents in the main experiments?\nAnswer:", + " 16 vCPUs, 256GB RAM, CPU-only" + ] + ], + "resps": [ + [ + [ + -39.894012451171875, + false + ] + ], + [ + [ + -50.85301971435547, + false + ] + ], + [ + [ + -34.946414947509766, + false + ] + ], + [ + [ + -34.39673614501953, + false + ] + ] + ], + "filtered_resps": [ + [ + -39.894012451171875, + false + ], + [ + -50.85301971435547, + false + ], + [ + -34.946414947509766, + false + ], + [ + -34.39673614501953, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "324c996d7e72ebec5325fd5337852d942c0db98233ed0d4d48ccca45ecb50ec8", + "prompt_hash": "4cda520c3bf17ff9d3431813d3b315605c015fe02763d0c869cc04b255e6719e", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 30, + "doc": { + "question": "How many text tokens can the Long-prompt MetaCLIP model process?", + "choices": [ + "128", + "256", + "512", + "1024" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:How many text tokens can the Long-prompt MetaCLIP model process?\nAnswer:", + " 128" + ], + [ + "Question:How many text tokens can the Long-prompt MetaCLIP model process?\nAnswer:", + " 256" + ], + [ + "Question:How many text tokens can the Long-prompt MetaCLIP model process?\nAnswer:", + " 512" + ], + [ + "Question:How many text tokens can the Long-prompt MetaCLIP model process?\nAnswer:", + " 1024" + ] + ], + "resps": [ + [ + [ + -5.544123649597168, + false + ] + ], + [ + [ + -5.784742832183838, + false + ] + ], + [ + [ + -5.334044933319092, + false + ] + ], + [ + [ + -5.7418904304504395, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.544123649597168, + false + ], + [ + -5.784742832183838, + false + ], + [ + -5.334044933319092, + false + ], + [ + -5.7418904304504395, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d1b2ded393a41d5e4100bc8ae2aae64f9edeb103890d6b8b9ff30606131a999e", + "prompt_hash": "7c40cd813ebb5da0e8c000275aae0577ed7e15a0c8036e8bba23fa8fbe9e72f8", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 31, + "doc": { + "question": "Which text encoder is used to encode the text within quotation marks for visual text generation?", + "choices": [ + "BERT", + "ByT5", + "T5-base", + "GPT-2" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which text encoder is used to encode the text within quotation marks for visual text generation?\nAnswer:", + " BERT" + ], + [ + "Question:Which text encoder is used to encode the text within quotation marks for visual text generation?\nAnswer:", + " ByT5" + ], + [ + "Question:Which text encoder is used to encode the text within quotation marks for visual text generation?\nAnswer:", + " T5-base" + ], + [ + "Question:Which text encoder is used to encode the text within quotation marks for visual text generation?\nAnswer:", + " GPT-2" + ] + ], + "resps": [ + [ + [ + -4.217916011810303, + false + ] + ], + [ + [ + -19.261009216308594, + false + ] + ], + [ + [ + -9.65689754486084, + false + ] + ], + [ + [ + -8.248754501342773, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.217916011810303, + false + ], + [ + -19.261009216308594, + false + ], + [ + -9.65689754486084, + false + ], + [ + -8.248754501342773, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "f27c798ae14ce3089bf0f0c00deae0322b0372fa8f26a19b8f78218d3ff0714a", + "prompt_hash": "13d9d8a856a2a64dcbc496b7df5c45a65221315895eb20b594e88ea16556d4a2", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 32, + "doc": { + "question": "What is the context length of the \\u005cOursVideo model mentioned as being memory capacity constrained?", + "choices": [ + "32K", + "64K", + "73K", + "100K" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the context length of the \\u005cOursVideo model mentioned as being memory capacity constrained?\nAnswer:", + " 32K" + ], + [ + "Question:What is the context length of the \\u005cOursVideo model mentioned as being memory capacity constrained?\nAnswer:", + " 64K" + ], + [ + "Question:What is the context length of the \\u005cOursVideo model mentioned as being memory capacity constrained?\nAnswer:", + " 73K" + ], + [ + "Question:What is the context length of the \\u005cOursVideo model mentioned as being memory capacity constrained?\nAnswer:", + " 100K" + ] + ], + "resps": [ + [ + [ + -6.79046106338501, + false + ] + ], + [ + [ + -6.787683010101318, + false + ] + ], + [ + [ + -13.908809661865234, + false + ] + ], + [ + [ + -8.865966796875, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.79046106338501, + false + ], + [ + -6.787683010101318, + false + ], + [ + -13.908809661865234, + false + ], + [ + -8.865966796875, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "809bdf95a7d2b2e018745fc5b2d5201b7de5809c2a97ce5d890bfeeeed94d5c3", + "prompt_hash": "03e62a79c165dedef8a425e117dcebd7ed4031dd7561a4ae8a557b510b8eeed9", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 33, + "doc": { + "question": "Which optimizer is used for both pre-training and fine-tuning of the 13B parameter model?", + "choices": [ + "SGD", + "Adam", + "AdamW", + "RMSProp" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which optimizer is used for both pre-training and fine-tuning of the 13B parameter model?\nAnswer:", + " SGD" + ], + [ + "Question:Which optimizer is used for both pre-training and fine-tuning of the 13B parameter model?\nAnswer:", + " Adam" + ], + [ + "Question:Which optimizer is used for both pre-training and fine-tuning of the 13B parameter model?\nAnswer:", + " AdamW" + ], + [ + "Question:Which optimizer is used for both pre-training and fine-tuning of the 13B parameter model?\nAnswer:", + " RMSProp" + ] + ], + "resps": [ + [ + [ + -7.288270950317383, + false + ] + ], + [ + [ + -3.32779598236084, + false + ] + ], + [ + [ + -3.610401153564453, + false + ] + ], + [ + [ + -9.468937873840332, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.288270950317383, + false + ], + [ + -3.32779598236084, + false + ], + [ + -3.610401153564453, + false + ], + [ + -9.468937873840332, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "99c196e8e4c7409c2b7aa9f30d47df15ca1b0c7a722cbbc15a613bfd9dd5aef3", + "prompt_hash": "43cdf1515da7e9b1eb00b02198d7d783e08e0881430e1f3c47504dd92e08ae11", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 34, + "doc": { + "question": "What is the audio sampling rate represented in the DAC-VAE latent features?", + "choices": [ + "16kHz", + "24kHz", + "32kHz", + "48kHz" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the audio sampling rate represented in the DAC-VAE latent features?\nAnswer:", + " 16kHz" + ], + [ + "Question:What is the audio sampling rate represented in the DAC-VAE latent features?\nAnswer:", + " 24kHz" + ], + [ + "Question:What is the audio sampling rate represented in the DAC-VAE latent features?\nAnswer:", + " 32kHz" + ], + [ + "Question:What is the audio sampling rate represented in the DAC-VAE latent features?\nAnswer:", + " 48kHz" + ] + ], + "resps": [ + [ + [ + -5.249194145202637, + false + ] + ], + [ + [ + -6.91918420791626, + false + ] + ], + [ + [ + -7.8241071701049805, + false + ] + ], + [ + [ + -5.517940521240234, + false + ] + ] + ], + "filtered_resps": [ + [ + -5.249194145202637, + false + ], + [ + -6.91918420791626, + false + ], + [ + -7.8241071701049805, + false + ], + [ + -5.517940521240234, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "145ccfd7bc07648de249cf0b46076373042acd5df8adc434331dc87be75f34bd", + "prompt_hash": "55567af4aa8ff2edd11efac29095b85b6d1f930db9afcbb8f368d9b15c592253", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 35, + "doc": { + "question": "What does the 'point-wise' input format for LLM-as-a-judge involve?", + "choices": [ + "Comparing multiple candidates simultaneously", + "Evaluating a single candidate in isolation", + "Ranking a list of outputs", + "Generating multiple outputs from one input" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What does the 'point-wise' input format for LLM-as-a-judge involve?\nAnswer:", + " Comparing multiple candidates simultaneously" + ], + [ + "Question:What does the 'point-wise' input format for LLM-as-a-judge involve?\nAnswer:", + " Evaluating a single candidate in isolation" + ], + [ + "Question:What does the 'point-wise' input format for LLM-as-a-judge involve?\nAnswer:", + " Ranking a list of outputs" + ], + [ + "Question:What does the 'point-wise' input format for LLM-as-a-judge involve?\nAnswer:", + " Generating multiple outputs from one input" + ] + ], + "resps": [ + [ + [ + -27.684053421020508, + false + ] + ], + [ + [ + -25.727405548095703, + false + ] + ], + [ + [ + -24.2668399810791, + false + ] + ], + [ + [ + -20.393577575683594, + false + ] + ] + ], + "filtered_resps": [ + [ + -27.684053421020508, + false + ], + [ + -25.727405548095703, + false + ], + [ + -24.2668399810791, + false + ], + [ + -20.393577575683594, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5147b0a7b504381ff4f451965e9cb04f6545f215720ef31c8c0c3afdcd189f9a", + "prompt_hash": "da8f6cedd16b324786b50aa5729cb61fb27c4084eee4aaeb83c4c531386cf529", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 36, + "doc": { + "question": "Which metric is used in JudgeBench to evaluate general performance?", + "choices": [ + "Discernment Score", + "Accuracy and F1", + "Cohen’s kappa and Correlation", + "BrierScore" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which metric is used in JudgeBench to evaluate general performance?\nAnswer:", + " Discernment Score" + ], + [ + "Question:Which metric is used in JudgeBench to evaluate general performance?\nAnswer:", + " Accuracy and F1" + ], + [ + "Question:Which metric is used in JudgeBench to evaluate general performance?\nAnswer:", + " Cohen’s kappa and Correlation" + ], + [ + "Question:Which metric is used in JudgeBench to evaluate general performance?\nAnswer:", + " BrierScore" + ] + ], + "resps": [ + [ + [ + -19.43678855895996, + false + ] + ], + [ + [ + -9.403539657592773, + false + ] + ], + [ + [ + -24.37989044189453, + false + ] + ], + [ + [ + -14.429201126098633, + false + ] + ] + ], + "filtered_resps": [ + [ + -19.43678855895996, + false + ], + [ + -9.403539657592773, + false + ], + [ + -24.37989044189453, + false + ], + [ + -14.429201126098633, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "c853498ca30f12dca012b59e7d6af17ed62c08019674f5677eb5aa99e4380064", + "prompt_hash": "63b034eedcf0a182cce93896fb71599c6b54dafef9ed34606e8cefe024d3461d", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 37, + "doc": { + "question": "What is the main idea behind the 'swapping operation' in LLM-as-a-judge prompting?", + "choices": [ + "Using different LLMs for comparison", + "Generating multiple candidate answers for diversity", + "Swapping the order of candidates to reduce positional bias", + "Embedding rule-based instructions inside prompts" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the main idea behind the 'swapping operation' in LLM-as-a-judge prompting?\nAnswer:", + " Using different LLMs for comparison" + ], + [ + "Question:What is the main idea behind the 'swapping operation' in LLM-as-a-judge prompting?\nAnswer:", + " Generating multiple candidate answers for diversity" + ], + [ + "Question:What is the main idea behind the 'swapping operation' in LLM-as-a-judge prompting?\nAnswer:", + " Swapping the order of candidates to reduce positional bias" + ], + [ + "Question:What is the main idea behind the 'swapping operation' in LLM-as-a-judge prompting?\nAnswer:", + " Embedding rule-based instructions inside prompts" + ] + ], + "resps": [ + [ + [ + -24.063907623291016, + false + ] + ], + [ + [ + -30.256454467773438, + false + ] + ], + [ + [ + -36.72766876220703, + false + ] + ], + [ + [ + -35.9759521484375, + false + ] + ] + ], + "filtered_resps": [ + [ + -24.063907623291016, + false + ], + [ + -30.256454467773438, + false + ], + [ + -36.72766876220703, + false + ], + [ + -35.9759521484375, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "fef50f0d1b80245b2d719dd1ea8316f6f2fa1b8e3658ab61fbc1666932a1dbb7", + "prompt_hash": "c17a750fee73b52ead083b0ff2525ee5d3fe394f6cf9437c474f352bc31ba3c6", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 38, + "doc": { + "question": "Which tuning method is most commonly used to train judge LLMs?", + "choices": [ + "Prompt-tuning", + "Supervised fine-tuning", + "Reinforcement learning", + "Zero-shot learning" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which tuning method is most commonly used to train judge LLMs?\nAnswer:", + " Prompt-tuning" + ], + [ + "Question:Which tuning method is most commonly used to train judge LLMs?\nAnswer:", + " Supervised fine-tuning" + ], + [ + "Question:Which tuning method is most commonly used to train judge LLMs?\nAnswer:", + " Reinforcement learning" + ], + [ + "Question:Which tuning method is most commonly used to train judge LLMs?\nAnswer:", + " Zero-shot learning" + ] + ], + "resps": [ + [ + [ + -11.313997268676758, + false + ] + ], + [ + [ + -7.638260364532471, + false + ] + ], + [ + [ + -7.296088695526123, + false + ] + ], + [ + [ + -8.118300437927246, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.313997268676758, + false + ], + [ + -7.638260364532471, + false + ], + [ + -7.296088695526123, + false + ], + [ + -8.118300437927246, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2d66761160ed9f7951b51315e9b16a0cc365c2b1c8d8acfea1a973d619089c8b", + "prompt_hash": "80ec918f8a736022d5913e2c075a35cec28d2e7842abff0ca5435fe849afa940", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 39, + "doc": { + "question": "Which application domain uses LLM-as-a-judge to filter noisy or irrelevant knowledge in a sub-knowledge graph?", + "choices": [ + "Summarization", + "Dialogue evaluation", + "Retrieval-Augmented Generation", + "Legal reasoning" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which application domain uses LLM-as-a-judge to filter noisy or irrelevant knowledge in a sub-knowledge graph?\nAnswer:", + " Summarization" + ], + [ + "Question:Which application domain uses LLM-as-a-judge to filter noisy or irrelevant knowledge in a sub-knowledge graph?\nAnswer:", + " Dialogue evaluation" + ], + [ + "Question:Which application domain uses LLM-as-a-judge to filter noisy or irrelevant knowledge in a sub-knowledge graph?\nAnswer:", + " Retrieval-Augmented Generation" + ], + [ + "Question:Which application domain uses LLM-as-a-judge to filter noisy or irrelevant knowledge in a sub-knowledge graph?\nAnswer:", + " Legal reasoning" + ] + ], + "resps": [ + [ + [ + -9.570253372192383, + false + ] + ], + [ + [ + -17.797739028930664, + false + ] + ], + [ + [ + -7.8971781730651855, + false + ] + ], + [ + [ + -11.672818183898926, + false + ] + ] + ], + "filtered_resps": [ + [ + -9.570253372192383, + false + ], + [ + -17.797739028930664, + false + ], + [ + -7.8971781730651855, + false + ], + [ + -11.672818183898926, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "51fc77b030408ed6f676f75a626b6f57c63b81eb0a4af3ac720bb5e1cc86a199", + "prompt_hash": "faffe2aa09e2f10dfef227dcff03f205a24d025d756646b3b54740f5aba52296", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 40, + "doc": { + "question": "What is the total number of parameters in DeepSeek-V2-8 (dsviii)?", + "choices": [ + "175B", + "371B", + "671B", + "870B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the total number of parameters in DeepSeek-V2-8 (dsviii)?\nAnswer:", + " 175B" + ], + [ + "Question:What is the total number of parameters in DeepSeek-V2-8 (dsviii)?\nAnswer:", + " 371B" + ], + [ + "Question:What is the total number of parameters in DeepSeek-V2-8 (dsviii)?\nAnswer:", + " 671B" + ], + [ + "Question:What is the total number of parameters in DeepSeek-V2-8 (dsviii)?\nAnswer:", + " 870B" + ] + ], + "resps": [ + [ + [ + -13.916131973266602, + false + ] + ], + [ + [ + -15.847789764404297, + false + ] + ], + [ + [ + -17.21529197692871, + false + ] + ], + [ + [ + -16.59684944152832, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.916131973266602, + false + ], + [ + -15.847789764404297, + false + ], + [ + -17.21529197692871, + false + ], + [ + -16.59684944152832, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7ea93a402030f3c4d9a8bd084d2d28d49a5095123c71e757e71f66ecedf9603c", + "prompt_hash": "b37ad914f970199ece9b60eb37b1e500e5d90430547fb149a786ca0a91a68199", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 41, + "doc": { + "question": "How many activated parameters are used per token in dsviii?", + "choices": [ + "13B", + "25B", + "37B", + "50B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:How many activated parameters are used per token in dsviii?\nAnswer:", + " 13B" + ], + [ + "Question:How many activated parameters are used per token in dsviii?\nAnswer:", + " 25B" + ], + [ + "Question:How many activated parameters are used per token in dsviii?\nAnswer:", + " 37B" + ], + [ + "Question:How many activated parameters are used per token in dsviii?\nAnswer:", + " 50B" + ] + ], + "resps": [ + [ + [ + -13.25839614868164, + false + ] + ], + [ + [ + -14.30405330657959, + false + ] + ], + [ + [ + -14.241345405578613, + false + ] + ], + [ + [ + -14.088238716125488, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.25839614868164, + false + ], + [ + -14.30405330657959, + false + ], + [ + -14.241345405578613, + false + ], + [ + -14.088238716125488, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ad3719a19744495dfb9c5a1ecc6df61a3cde9eb7114c6c9e959570b322ac1353", + "prompt_hash": "1cc12d4fcb1c309957d1f70fb0e1501acb19e2992f602b1462bd4d6a4cc76475", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 42, + "doc": { + "question": "What data format is used to track the first and second moments in the AdamW optimizer in the FP8 training framework?", + "choices": [ + "FP32", + "BF16", + "INT8", + "E5M2" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What data format is used to track the first and second moments in the AdamW optimizer in the FP8 training framework?\nAnswer:", + " FP32" + ], + [ + "Question:What data format is used to track the first and second moments in the AdamW optimizer in the FP8 training framework?\nAnswer:", + " BF16" + ], + [ + "Question:What data format is used to track the first and second moments in the AdamW optimizer in the FP8 training framework?\nAnswer:", + " INT8" + ], + [ + "Question:What data format is used to track the first and second moments in the AdamW optimizer in the FP8 training framework?\nAnswer:", + " E5M2" + ] + ], + "resps": [ + [ + [ + -6.028921127319336, + false + ] + ], + [ + [ + -12.088250160217285, + false + ] + ], + [ + [ + -8.075591087341309, + false + ] + ], + [ + [ + -19.291461944580078, + false + ] + ] + ], + "filtered_resps": [ + [ + -6.028921127319336, + false + ], + [ + -12.088250160217285, + false + ], + [ + -8.075591087341309, + false + ], + [ + -19.291461944580078, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "197e98fe296eb5a9e46206a3581a73f46176e34978688b86540fe1073a41a6bd", + "prompt_hash": "628799f35bbdf1f99ea4cc59b4c1b981c83059334376dc8a4d09a3fe0f1fa7ea", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 43, + "doc": { + "question": "Which components are retained in higher precision for stability in the FP8 mixed precision framework?", + "choices": [ + "GEMM operations, SwiGLU, and MTP modules", + "Embedding module, output head, MoE gating modules, normalization operators, and attention operators", + "Only weight gradients and optimizer states", + "Only the attention operators and SwiGLU" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:Which components are retained in higher precision for stability in the FP8 mixed precision framework?\nAnswer:", + " GEMM operations, SwiGLU, and MTP modules" + ], + [ + "Question:Which components are retained in higher precision for stability in the FP8 mixed precision framework?\nAnswer:", + " Embedding module, output head, MoE gating modules, normalization operators, and attention operators" + ], + [ + "Question:Which components are retained in higher precision for stability in the FP8 mixed precision framework?\nAnswer:", + " Only weight gradients and optimizer states" + ], + [ + "Question:Which components are retained in higher precision for stability in the FP8 mixed precision framework?\nAnswer:", + " Only the attention operators and SwiGLU" + ] + ], + "resps": [ + [ + [ + -47.497535705566406, + false + ] + ], + [ + [ + -75.67779541015625, + false + ] + ], + [ + [ + -19.668460845947266, + false + ] + ], + [ + [ + -29.46173858642578, + false + ] + ] + ], + "filtered_resps": [ + [ + -47.497535705566406, + false + ], + [ + -75.67779541015625, + false + ], + [ + -19.668460845947266, + false + ], + [ + -29.46173858642578, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "324ca988bc896486250ce24ea2d7deaaf1c860ef1c1c34b1197fdbedb6e9355f", + "prompt_hash": "e371a9fdd26fb76793b01d971a20d8b4eb8c18293ae2002a4bdb34f2a2f5ff34", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 44, + "doc": { + "question": "What is the accumulation interval (Nc) set for improving FP8 GEMM precision in the quantization strategy?", + "choices": [ + "32", + "64", + "96", + "128" + ], + "answer": 3 + }, + "target": 3, + "arguments": [ + [ + "Question:What is the accumulation interval (Nc) set for improving FP8 GEMM precision in the quantization strategy?\nAnswer:", + " 32" + ], + [ + "Question:What is the accumulation interval (Nc) set for improving FP8 GEMM precision in the quantization strategy?\nAnswer:", + " 64" + ], + [ + "Question:What is the accumulation interval (Nc) set for improving FP8 GEMM precision in the quantization strategy?\nAnswer:", + " 96" + ], + [ + "Question:What is the accumulation interval (Nc) set for improving FP8 GEMM precision in the quantization strategy?\nAnswer:", + " 128" + ] + ], + "resps": [ + [ + [ + -4.304234027862549, + false + ] + ], + [ + [ + -4.729826927185059, + false + ] + ], + [ + [ + -6.842705249786377, + false + ] + ], + [ + [ + -5.4878153800964355, + false + ] + ] + ], + "filtered_resps": [ + [ + -4.304234027862549, + false + ], + [ + -4.729826927185059, + false + ], + [ + -6.842705249786377, + false + ], + [ + -5.4878153800964355, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5801dde4164274339b30b0390dccd2dedbf00432977312be951e0e4248214ae1", + "prompt_hash": "6e5988a5ae8b3c538c504ef5add7c7e2a8d705df79e155423594dc5f698a7190", + "target_hash": "4e07408562bedb8b60ce05c1decfe3ad16b72230967de01f640b7e4729b49fce", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 45, + "doc": { + "question": "What percentage more tokens did o1-like models consume compared to conventional models to answer 'what is the answer of 2 plus 3?'", + "choices": [ + "1953%", + "953%", + "305%", + "120%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What percentage more tokens did o1-like models consume compared to conventional models to answer 'what is the answer of 2 plus 3?'\nAnswer:", + " 1953%" + ], + [ + "Question:What percentage more tokens did o1-like models consume compared to conventional models to answer 'what is the answer of 2 plus 3?'\nAnswer:", + " 953%" + ], + [ + "Question:What percentage more tokens did o1-like models consume compared to conventional models to answer 'what is the answer of 2 plus 3?'\nAnswer:", + " 305%" + ], + [ + "Question:What percentage more tokens did o1-like models consume compared to conventional models to answer 'what is the answer of 2 plus 3?'\nAnswer:", + " 120%" + ] + ], + "resps": [ + [ + [ + -20.174894332885742, + false + ] + ], + [ + [ + -19.59738540649414, + false + ] + ], + [ + [ + -18.090007781982422, + false + ] + ], + [ + [ + -13.911893844604492, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.174894332885742, + false + ], + [ + -19.59738540649414, + false + ], + [ + -18.090007781982422, + false + ], + [ + -13.911893844604492, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "0b0e6779100e4b9673f74cc0c39a40f404d64451eac6844048e1f6384aa72049", + "prompt_hash": "8f0f3344016101019ae1762b909f868ea45e0cd39f37f0d6bc0e4c46e08c21f8", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 46, + "doc": { + "question": "What is the average number of solutions generated by QwQ-32B-Preview on the ASDIV test set?", + "choices": [ + "3.5", + "3.2", + "4.3", + "2.8" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the average number of solutions generated by QwQ-32B-Preview on the ASDIV test set?\nAnswer:", + " 3.5" + ], + [ + "Question:What is the average number of solutions generated by QwQ-32B-Preview on the ASDIV test set?\nAnswer:", + " 3.2" + ], + [ + "Question:What is the average number of solutions generated by QwQ-32B-Preview on the ASDIV test set?\nAnswer:", + " 4.3" + ], + [ + "Question:What is the average number of solutions generated by QwQ-32B-Preview on the ASDIV test set?\nAnswer:", + " 2.8" + ] + ], + "resps": [ + [ + [ + -7.472996711730957, + false + ] + ], + [ + [ + -7.623888969421387, + false + ] + ], + [ + [ + -8.0897798538208, + false + ] + ], + [ + [ + -7.952129364013672, + false + ] + ] + ], + "filtered_resps": [ + [ + -7.472996711730957, + false + ], + [ + -7.623888969421387, + false + ], + [ + -8.0897798538208, + false + ], + [ + -7.952129364013672, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ec0a559a5ff4a1a43d020e7510901163af232b1965edfa2dbb194944d45948b5", + "prompt_hash": "6e995540d3062269c2768fb3c72f748a83b71ad6886e28f838f025c64b8dccb0", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 47, + "doc": { + "question": "Which test set consists of high school math competition problems with difficulty levels based on AoPS?", + "choices": [ + "MATH500", + "ASDIV", + "GSM8K", + "GPQA" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which test set consists of high school math competition problems with difficulty levels based on AoPS?\nAnswer:", + " MATH500" + ], + [ + "Question:Which test set consists of high school math competition problems with difficulty levels based on AoPS?\nAnswer:", + " ASDIV" + ], + [ + "Question:Which test set consists of high school math competition problems with difficulty levels based on AoPS?\nAnswer:", + " GSM8K" + ], + [ + "Question:Which test set consists of high school math competition problems with difficulty levels based on AoPS?\nAnswer:", + " GPQA" + ] + ], + "resps": [ + [ + [ + -13.954486846923828, + false + ] + ], + [ + [ + -23.078916549682617, + false + ] + ], + [ + [ + -6.7538533210754395, + false + ] + ], + [ + [ + -12.076217651367188, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.954486846923828, + false + ], + [ + -23.078916549682617, + false + ], + [ + -6.7538533210754395, + false + ], + [ + -12.076217651367188, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "b6db1da4a0557810aa7d2562fa4c6005c641bb0b5dfedee7da28d5f2d63a3223", + "prompt_hash": "37cdc3dc374f3260b0a938b7a88d7e61032286a9ad6c759c43f393db2300b36c", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 48, + "doc": { + "question": "What is the process efficiency score of QwQ-32B-Preview on the MATH500 dataset?", + "choices": [ + "71.2%", + "66.5%", + "100%", + "52.3%" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:What is the process efficiency score of QwQ-32B-Preview on the MATH500 dataset?\nAnswer:", + " 71.2%" + ], + [ + "Question:What is the process efficiency score of QwQ-32B-Preview on the MATH500 dataset?\nAnswer:", + " 66.5%" + ], + [ + "Question:What is the process efficiency score of QwQ-32B-Preview on the MATH500 dataset?\nAnswer:", + " 100%" + ], + [ + "Question:What is the process efficiency score of QwQ-32B-Preview on the MATH500 dataset?\nAnswer:", + " 52.3%" + ] + ], + "resps": [ + [ + [ + -15.457481384277344, + false + ] + ], + [ + [ + -15.54973030090332, + false + ] + ], + [ + [ + -12.790609359741211, + false + ] + ], + [ + [ + -15.633562088012695, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.457481384277344, + false + ], + [ + -15.54973030090332, + false + ], + [ + -12.790609359741211, + false + ], + [ + -15.633562088012695, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2ad217e842103818fb058ca928915af562c3d32d97403ba2a58aff078ac2dfc9", + "prompt_hash": "1c05afa023d5534e9a4adb5ed1e77790c1f49f9191b7727c9bc7f3ba92500f1b", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 1.0 + }, + { + "doc_id": 49, + "doc": { + "question": "Which training method modifies the DPO loss by adding a negative log-likelihood term on the preferred response?", + "choices": [ + "Reasoning Preference Optimization", + "Direct Preference Optimization", + "Simple Preference Optimization", + "Supervised Fine-Tuning" + ], + "answer": 0 + }, + "target": 0, + "arguments": [ + [ + "Question:Which training method modifies the DPO loss by adding a negative log-likelihood term on the preferred response?\nAnswer:", + " Reasoning Preference Optimization" + ], + [ + "Question:Which training method modifies the DPO loss by adding a negative log-likelihood term on the preferred response?\nAnswer:", + " Direct Preference Optimization" + ], + [ + "Question:Which training method modifies the DPO loss by adding a negative log-likelihood term on the preferred response?\nAnswer:", + " Simple Preference Optimization" + ], + [ + "Question:Which training method modifies the DPO loss by adding a negative log-likelihood term on the preferred response?\nAnswer:", + " Supervised Fine-Tuning" + ] + ], + "resps": [ + [ + [ + -21.339204788208008, + false + ] + ], + [ + [ + -7.6230621337890625, + false + ] + ], + [ + [ + -15.905888557434082, + false + ] + ], + [ + [ + -11.230256080627441, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.339204788208008, + false + ], + [ + -7.6230621337890625, + false + ], + [ + -15.905888557434082, + false + ], + [ + -11.230256080627441, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ae5156c48bff3c2947d892d0104a321c3b910ae6797d6036673ab01f7b04a690", + "prompt_hash": "0a3df59b66b51a90ee60d83f63e77576fba1d1f9cb6476ca070010f130515783", + "target_hash": "5feceb66ffc86f38d952786c6d696c79c2dbc239dd4e91b46729d73a27fb57e9", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 50, + "doc": { + "question": "What is the main purpose of the underthinking metric introduced in the study?", + "choices": [ + "To evaluate model training speed", + "To measure token efficiency in incorrect responses", + "To calculate the accuracy of correct answers", + "To compare different datasets' difficulty levels" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the main purpose of the underthinking metric introduced in the study?\nAnswer:", + " To evaluate model training speed" + ], + [ + "Question:What is the main purpose of the underthinking metric introduced in the study?\nAnswer:", + " To measure token efficiency in incorrect responses" + ], + [ + "Question:What is the main purpose of the underthinking metric introduced in the study?\nAnswer:", + " To calculate the accuracy of correct answers" + ], + [ + "Question:What is the main purpose of the underthinking metric introduced in the study?\nAnswer:", + " To compare different datasets' difficulty levels" + ] + ], + "resps": [ + [ + [ + -20.352832794189453, + false + ] + ], + [ + [ + -38.05958557128906, + false + ] + ], + [ + [ + -21.12741470336914, + false + ] + ], + [ + [ + -26.977842330932617, + false + ] + ] + ], + "filtered_resps": [ + [ + -20.352832794189453, + false + ], + [ + -38.05958557128906, + false + ], + [ + -21.12741470336914, + false + ], + [ + -26.977842330932617, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "a6f3c5261b5413457e10134ff2c03446b0836d09ab936070b0813af4d21110cc", + "prompt_hash": "b0adf48617c122abcbd12087b5da586115198df7e7317189a6afbd3bafe260aa", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 51, + "doc": { + "question": "Which of the following models was evaluated only on MATH500 and AIME test sets due to daily message limits?", + "choices": [ + "QwQ-32B-Preview", + "DeepSeek-R1-671B", + "DeepSeek-R1-Preview", + "DeepSeek-R1-Distill-Qwen-32B" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which of the following models was evaluated only on MATH500 and AIME test sets due to daily message limits?\nAnswer:", + " QwQ-32B-Preview" + ], + [ + "Question:Which of the following models was evaluated only on MATH500 and AIME test sets due to daily message limits?\nAnswer:", + " DeepSeek-R1-671B" + ], + [ + "Question:Which of the following models was evaluated only on MATH500 and AIME test sets due to daily message limits?\nAnswer:", + " DeepSeek-R1-Preview" + ], + [ + "Question:Which of the following models was evaluated only on MATH500 and AIME test sets due to daily message limits?\nAnswer:", + " DeepSeek-R1-Distill-Qwen-32B" + ] + ], + "resps": [ + [ + [ + -36.87464904785156, + false + ] + ], + [ + [ + -45.31444549560547, + false + ] + ], + [ + [ + -40.90092468261719, + false + ] + ], + [ + [ + -55.350929260253906, + false + ] + ] + ], + "filtered_resps": [ + [ + -36.87464904785156, + false + ], + [ + -45.31444549560547, + false + ], + [ + -40.90092468261719, + false + ], + [ + -55.350929260253906, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "7404b2b8cef2fead98865f72bf451d6f363b7078b7f0400602b9a84af2932597", + "prompt_hash": "45139f45e167354a848f1d29a32181af1ba20827585568989c1f658e076971ea", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 52, + "doc": { + "question": "What is the role of the penalty strength parameter α in the Tip decoding method?", + "choices": [ + "To determine the number of reasoning thoughts", + "To control the penalty duration applied to tokens", + "To reduce the likelihood of thought-switching tokens", + "To adjust the token length for correct answers" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What is the role of the penalty strength parameter α in the Tip decoding method?\nAnswer:", + " To determine the number of reasoning thoughts" + ], + [ + "Question:What is the role of the penalty strength parameter α in the Tip decoding method?\nAnswer:", + " To control the penalty duration applied to tokens" + ], + [ + "Question:What is the role of the penalty strength parameter α in the Tip decoding method?\nAnswer:", + " To reduce the likelihood of thought-switching tokens" + ], + [ + "Question:What is the role of the penalty strength parameter α in the Tip decoding method?\nAnswer:", + " To adjust the token length for correct answers" + ] + ], + "resps": [ + [ + [ + -30.340669631958008, + false + ] + ], + [ + [ + -34.11429214477539, + false + ] + ], + [ + [ + -43.95995330810547, + false + ] + ], + [ + [ + -37.085777282714844, + false + ] + ] + ], + "filtered_resps": [ + [ + -30.340669631958008, + false + ], + [ + -34.11429214477539, + false + ], + [ + -43.95995330810547, + false + ], + [ + -37.085777282714844, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "e43c3398bd1490fd4441f8fea9482409b851f3a08624aa7467c36115685d6565", + "prompt_hash": "6fc20179bce03abad0beb9260800aee68a16fcd61d29bc013e831cf94a82478c", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 53, + "doc": { + "question": "In the underthinking score formula, what does a higher value of ξ_UT indicate?", + "choices": [ + "Greater token efficiency", + "Lower token efficiency", + "Shorter reasoning chains", + "Fewer reasoning thoughts" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:In the underthinking score formula, what does a higher value of ξ_UT indicate?\nAnswer:", + " Greater token efficiency" + ], + [ + "Question:In the underthinking score formula, what does a higher value of ξ_UT indicate?\nAnswer:", + " Lower token efficiency" + ], + [ + "Question:In the underthinking score formula, what does a higher value of ξ_UT indicate?\nAnswer:", + " Shorter reasoning chains" + ], + [ + "Question:In the underthinking score formula, what does a higher value of ξ_UT indicate?\nAnswer:", + " Fewer reasoning thoughts" + ] + ], + "resps": [ + [ + [ + -25.667449951171875, + false + ] + ], + [ + [ + -21.466705322265625, + false + ] + ], + [ + [ + -23.393604278564453, + false + ] + ], + [ + [ + -25.824403762817383, + false + ] + ] + ], + "filtered_resps": [ + [ + -25.667449951171875, + false + ], + [ + -21.466705322265625, + false + ], + [ + -23.393604278564453, + false + ], + [ + -25.824403762817383, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "5a2f46c94c1ba3bcb2613a2f231ef9e9fc8a6ec762be09de290a5cf17cc5b670", + "prompt_hash": "f677ecf5c9cdfa65b5fd4869c3dd0ed700ba59daa188fc96c96db5f1b2b011ac", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 1.0, + "acc_norm": 0.0 + }, + { + "doc_id": 54, + "doc": { + "question": "Which prompt-based strategy showed only modest improvements in comparison to the Tip decoding method?", + "choices": [ + "Laconic Decoding", + "Self-Consistency Sampling", + "Prompt Engineering", + "Token Filtering" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which prompt-based strategy showed only modest improvements in comparison to the Tip decoding method?\nAnswer:", + " Laconic Decoding" + ], + [ + "Question:Which prompt-based strategy showed only modest improvements in comparison to the Tip decoding method?\nAnswer:", + " Self-Consistency Sampling" + ], + [ + "Question:Which prompt-based strategy showed only modest improvements in comparison to the Tip decoding method?\nAnswer:", + " Prompt Engineering" + ], + [ + "Question:Which prompt-based strategy showed only modest improvements in comparison to the Tip decoding method?\nAnswer:", + " Token Filtering" + ] + ], + "resps": [ + [ + [ + -21.468769073486328, + false + ] + ], + [ + [ + -18.396778106689453, + false + ] + ], + [ + [ + -12.113412857055664, + false + ] + ], + [ + [ + -16.382930755615234, + false + ] + ] + ], + "filtered_resps": [ + [ + -21.468769073486328, + false + ], + [ + -18.396778106689453, + false + ], + [ + -12.113412857055664, + false + ], + [ + -16.382930755615234, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "d8d69259421589a9881ad09598da0caf91ccd7c26ec1c2758b0c6c862ae81689", + "prompt_hash": "5a1b3dc3f8c0d7848aef3f588f05c7c36fe4faf51cf3a53465878666ef351a49", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 1.0, + "acc_norm": 1.0 + }, + { + "doc_id": 55, + "doc": { + "question": "What is the reported improvement in sample efficiency when using PRIME compared to RL using only outcome rewards?", + "choices": [ + "1.5×", + "2.5×", + "4.0×", + "6.9×" + ], + "answer": 1 + }, + "target": 1, + "arguments": [ + [ + "Question:What is the reported improvement in sample efficiency when using PRIME compared to RL using only outcome rewards?\nAnswer:", + " 1.5×" + ], + [ + "Question:What is the reported improvement in sample efficiency when using PRIME compared to RL using only outcome rewards?\nAnswer:", + " 2.5×" + ], + [ + "Question:What is the reported improvement in sample efficiency when using PRIME compared to RL using only outcome rewards?\nAnswer:", + " 4.0×" + ], + [ + "Question:What is the reported improvement in sample efficiency when using PRIME compared to RL using only outcome rewards?\nAnswer:", + " 6.9×" + ] + ], + "resps": [ + [ + [ + -13.429561614990234, + false + ] + ], + [ + [ + -14.082334518432617, + false + ] + ], + [ + [ + -15.574737548828125, + false + ] + ], + [ + [ + -17.069332122802734, + false + ] + ] + ], + "filtered_resps": [ + [ + -13.429561614990234, + false + ], + [ + -14.082334518432617, + false + ], + [ + -15.574737548828125, + false + ], + [ + -17.069332122802734, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "ce050f483e2d1900645d7aa208c21d9f86fb6ddef6e214eda733bad133756a9c", + "prompt_hash": "7256d5a3a7e1b82502e80fc9bf58c2bd8a01a4c7a3aa9a50a998d1a882e261a0", + "target_hash": "6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 56, + "doc": { + "question": "Which model serves as the starting point for training in the PRIME framework?", + "choices": [ + "GPT-4o", + "LLaMA-3.1-70B-Instruct", + "Qwen2.5-Math-7B-Base", + "DeepSeek-R1" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:Which model serves as the starting point for training in the PRIME framework?\nAnswer:", + " GPT-4o" + ], + [ + "Question:Which model serves as the starting point for training in the PRIME framework?\nAnswer:", + " LLaMA-3.1-70B-Instruct" + ], + [ + "Question:Which model serves as the starting point for training in the PRIME framework?\nAnswer:", + " Qwen2.5-Math-7B-Base" + ], + [ + "Question:Which model serves as the starting point for training in the PRIME framework?\nAnswer:", + " DeepSeek-R1" + ] + ], + "resps": [ + [ + [ + -11.462921142578125, + false + ] + ], + [ + [ + -31.877954483032227, + false + ] + ], + [ + [ + -40.35797119140625, + false + ] + ], + [ + [ + -25.913959503173828, + false + ] + ] + ], + "filtered_resps": [ + [ + -11.462921142578125, + false + ], + [ + -31.877954483032227, + false + ], + [ + -40.35797119140625, + false + ], + [ + -25.913959503173828, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "783695e7dd24e82177f27d12e81b48b130250480bd512fbc0859b8e43722cd79", + "prompt_hash": "629ea5c8e6c1a46907d0960aa86fcc9c372b377e4dfca59fc6bac98042f1fabe", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 57, + "doc": { + "question": "What evaluation score did Eurus-2-7B-PRIME achieve on AIME 2024?", + "choices": [ + "16.7%", + "20.0%", + "26.7%", + "32.5%" + ], + "answer": 2 + }, + "target": 2, + "arguments": [ + [ + "Question:What evaluation score did Eurus-2-7B-PRIME achieve on AIME 2024?\nAnswer:", + " 16.7%" + ], + [ + "Question:What evaluation score did Eurus-2-7B-PRIME achieve on AIME 2024?\nAnswer:", + " 20.0%" + ], + [ + "Question:What evaluation score did Eurus-2-7B-PRIME achieve on AIME 2024?\nAnswer:", + " 26.7%" + ], + [ + "Question:What evaluation score did Eurus-2-7B-PRIME achieve on AIME 2024?\nAnswer:", + " 32.5%" + ] + ], + "resps": [ + [ + [ + -15.968780517578125, + false + ] + ], + [ + [ + -15.175897598266602, + false + ] + ], + [ + [ + -15.799150466918945, + false + ] + ], + [ + [ + -15.199570655822754, + false + ] + ] + ], + "filtered_resps": [ + [ + -15.968780517578125, + false + ], + [ + -15.175897598266602, + false + ], + [ + -15.799150466918945, + false + ], + [ + -15.199570655822754, + false + ] + ], + "filter": "none", + "metrics": [ + "acc", + "acc_norm" + ], + "doc_hash": "2aa1e3918ba13aed9884bb87c237252fc729d54b11bccfa83da6b42a19648ec4", + "prompt_hash": "09a72f388f08d5eef9b51f61736ac1c5a71e2d72d4f05d16586c8f8350b3c64e", + "target_hash": "d4735e3a265e16eee03f59718b9b5d03019c07d8b6c51f90da3a666eec13ab35", + "acc": 0.0, + "acc_norm": 0.0 + }, + { + "doc_id": 58, + "doc": { + "question": "What is the formula for process reward in the Implicit PRM used by PRIME?", + "choices": [ + "r_\\phi(y_t) := \\alpha \\log \\frac{\\pi_\\phi(y_t)}{\\pi_\\text{ref}(y_t)}", + "r_\\phi(y_t) := \\log \\frac{\\pi_\\text{ref}(y_t|\\mathbf{y}_{