| { |
| "eval_type_id": "sparse_probing", |
| "eval_config": { |
| "random_seed": 42, |
| "dataset_names": [ |
| "LabHC/bias_in_bios_class_set1", |
| "LabHC/bias_in_bios_class_set2", |
| "LabHC/bias_in_bios_class_set3", |
| "canrager/amazon_reviews_mcauley_1and5", |
| "canrager/amazon_reviews_mcauley_1and5_sentiment", |
| "codeparrot/github-code", |
| "fancyzhx/ag_news", |
| "Helsinki-NLP/europarl" |
| ], |
| "probe_train_set_size": 4000, |
| "probe_test_set_size": 1000, |
| "context_length": 128, |
| "sae_batch_size": 125, |
| "llm_batch_size": 32, |
| "llm_dtype": "bfloat16", |
| "model_name": "gemma-2-2b", |
| "k_values": [ |
| 1, |
| 2, |
| 5 |
| ], |
| "lower_vram_usage": false |
| }, |
| "eval_id": "23b811f8-60d2-4669-9569-74fb61ee47a0", |
| "datetime_epoch_millis": 1745754994538, |
| "eval_result_metrics": { |
| "llm": { |
| "llm_test_accuracy": 0.9602875433862209, |
| "llm_top_1_test_accuracy": 0.7107374999999999, |
| "llm_top_2_test_accuracy": 0.76, |
| "llm_top_5_test_accuracy": 0.8195374999999999, |
| "llm_top_10_test_accuracy": null, |
| "llm_top_20_test_accuracy": null, |
| "llm_top_50_test_accuracy": null, |
| "llm_top_100_test_accuracy": null |
| }, |
| "sae": { |
| "sae_test_accuracy": 0.9573000423610211, |
| "sae_top_1_test_accuracy": 0.8020249999999998, |
| "sae_top_2_test_accuracy": 0.8709812499999998, |
| "sae_top_5_test_accuracy": 0.9142062500000001, |
| "sae_top_10_test_accuracy": null, |
| "sae_top_20_test_accuracy": null, |
| "sae_top_50_test_accuracy": null, |
| "sae_top_100_test_accuracy": null |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "dataset_name": "LabHC/bias_in_bios_class_set1_results", |
| "llm_test_accuracy": 0.9718000411987304, |
| "llm_top_1_test_accuracy": 0.6761999999999999, |
| "llm_top_2_test_accuracy": 0.7253999999999999, |
| "llm_top_5_test_accuracy": 0.8114000000000001, |
| "llm_top_10_test_accuracy": null, |
| "llm_top_20_test_accuracy": null, |
| "llm_top_50_test_accuracy": null, |
| "llm_top_100_test_accuracy": null, |
| "sae_test_accuracy": 0.9644000411033631, |
| "sae_top_1_test_accuracy": 0.8402000000000001, |
| "sae_top_2_test_accuracy": 0.9006000000000001, |
| "sae_top_5_test_accuracy": 0.9334, |
| "sae_top_10_test_accuracy": null, |
| "sae_top_20_test_accuracy": null, |
| "sae_top_50_test_accuracy": null, |
| "sae_top_100_test_accuracy": null |
| }, |
| { |
| "dataset_name": "LabHC/bias_in_bios_class_set2_results", |
| "llm_test_accuracy": 0.9574000477790833, |
| "llm_top_1_test_accuracy": 0.6768, |
| "llm_top_2_test_accuracy": 0.6898000000000001, |
| "llm_top_5_test_accuracy": 0.747, |
| "llm_top_10_test_accuracy": null, |
| "llm_top_20_test_accuracy": null, |
| "llm_top_50_test_accuracy": null, |
| "llm_top_100_test_accuracy": null, |
| "sae_test_accuracy": 0.9534000515937805, |
| "sae_top_1_test_accuracy": 0.8, |
| "sae_top_2_test_accuracy": 0.8700000000000001, |
| "sae_top_5_test_accuracy": 0.9122, |
| "sae_top_10_test_accuracy": null, |
| "sae_top_20_test_accuracy": null, |
| "sae_top_50_test_accuracy": null, |
| "sae_top_100_test_accuracy": null |
| }, |
| { |
| "dataset_name": "LabHC/bias_in_bios_class_set3_results", |
| "llm_test_accuracy": 0.9314000487327576, |
| "llm_top_1_test_accuracy": 0.7056, |
| "llm_top_2_test_accuracy": 0.7236, |
| "llm_top_5_test_accuracy": 0.7936000000000001, |
| "llm_top_10_test_accuracy": null, |
| "llm_top_20_test_accuracy": null, |
| "llm_top_50_test_accuracy": null, |
| "llm_top_100_test_accuracy": null, |
| "sae_test_accuracy": 0.9292000293731689, |
| "sae_top_1_test_accuracy": 0.8150000000000001, |
| "sae_top_2_test_accuracy": 0.8454, |
| "sae_top_5_test_accuracy": 0.8720000000000001, |
| "sae_top_10_test_accuracy": null, |
| "sae_top_20_test_accuracy": null, |
| "sae_top_50_test_accuracy": null, |
| "sae_top_100_test_accuracy": null |
| }, |
| { |
| "dataset_name": "canrager/amazon_reviews_mcauley_1and5_results", |
| "llm_test_accuracy": 0.9300000429153442, |
| "llm_top_1_test_accuracy": 0.6326, |
| "llm_top_2_test_accuracy": 0.7304, |
| "llm_top_5_test_accuracy": 0.7866, |
| "llm_top_10_test_accuracy": null, |
| "llm_top_20_test_accuracy": null, |
| "llm_top_50_test_accuracy": null, |
| "llm_top_100_test_accuracy": null, |
| "sae_test_accuracy": 0.9328000545501709, |
| "sae_top_1_test_accuracy": 0.7617999999999998, |
| "sae_top_2_test_accuracy": 0.8116, |
| "sae_top_5_test_accuracy": 0.8502000000000001, |
| "sae_top_10_test_accuracy": null, |
| "sae_top_20_test_accuracy": null, |
| "sae_top_50_test_accuracy": null, |
| "sae_top_100_test_accuracy": null |
| }, |
| { |
| "dataset_name": "canrager/amazon_reviews_mcauley_1and5_sentiment_results", |
| "llm_test_accuracy": 0.9720000624656677, |
| "llm_top_1_test_accuracy": 0.692, |
| "llm_top_2_test_accuracy": 0.744, |
| "llm_top_5_test_accuracy": 0.782, |
| "llm_top_10_test_accuracy": null, |
| "llm_top_20_test_accuracy": null, |
| "llm_top_50_test_accuracy": null, |
| "llm_top_100_test_accuracy": null, |
| "sae_test_accuracy": 0.963500052690506, |
| "sae_top_1_test_accuracy": 0.611, |
| "sae_top_2_test_accuracy": 0.808, |
| "sae_top_5_test_accuracy": 0.916, |
| "sae_top_10_test_accuracy": null, |
| "sae_top_20_test_accuracy": null, |
| "sae_top_50_test_accuracy": null, |
| "sae_top_100_test_accuracy": null |
| }, |
| { |
| "dataset_name": "codeparrot/github-code_results", |
| "llm_test_accuracy": 0.9694000482559204, |
| "llm_top_1_test_accuracy": 0.6288, |
| "llm_top_2_test_accuracy": 0.6922, |
| "llm_top_5_test_accuracy": 0.8058, |
| "llm_top_10_test_accuracy": null, |
| "llm_top_20_test_accuracy": null, |
| "llm_top_50_test_accuracy": null, |
| "llm_top_100_test_accuracy": null, |
| "sae_test_accuracy": 0.9682000398635864, |
| "sae_top_1_test_accuracy": 0.8103999999999999, |
| "sae_top_2_test_accuracy": 0.8629999999999999, |
| "sae_top_5_test_accuracy": 0.9362, |
| "sae_top_10_test_accuracy": null, |
| "sae_top_20_test_accuracy": null, |
| "sae_top_50_test_accuracy": null, |
| "sae_top_100_test_accuracy": null |
| }, |
| { |
| "dataset_name": "fancyzhx/ag_news_results", |
| "llm_test_accuracy": 0.9505000412464142, |
| "llm_top_1_test_accuracy": 0.6905, |
| "llm_top_2_test_accuracy": 0.783, |
| "llm_top_5_test_accuracy": 0.8305, |
| "llm_top_10_test_accuracy": null, |
| "llm_top_20_test_accuracy": null, |
| "llm_top_50_test_accuracy": null, |
| "llm_top_100_test_accuracy": null, |
| "sae_test_accuracy": 0.9475000500679016, |
| "sae_top_1_test_accuracy": 0.7989999999999999, |
| "sae_top_2_test_accuracy": 0.8752499999999999, |
| "sae_top_5_test_accuracy": 0.89425, |
| "sae_top_10_test_accuracy": null, |
| "sae_top_20_test_accuracy": null, |
| "sae_top_50_test_accuracy": null, |
| "sae_top_100_test_accuracy": null |
| }, |
| { |
| "dataset_name": "Helsinki-NLP/europarl_results", |
| "llm_test_accuracy": 0.9998000144958497, |
| "llm_top_1_test_accuracy": 0.9833999999999999, |
| "llm_top_2_test_accuracy": 0.9916, |
| "llm_top_5_test_accuracy": 0.9994, |
| "llm_top_10_test_accuracy": null, |
| "llm_top_20_test_accuracy": null, |
| "llm_top_50_test_accuracy": null, |
| "llm_top_100_test_accuracy": null, |
| "sae_test_accuracy": 0.9994000196456909, |
| "sae_top_1_test_accuracy": 0.9788, |
| "sae_top_2_test_accuracy": 0.9940000000000001, |
| "sae_top_5_test_accuracy": 0.9994, |
| "sae_top_10_test_accuracy": null, |
| "sae_top_20_test_accuracy": null, |
| "sae_top_50_test_accuracy": null, |
| "sae_top_100_test_accuracy": null |
| } |
| ], |
| "sae_bench_commit_hash": "Unknown", |
| "sae_lens_id": "blocks.20.hook_resid_post", |
| "sae_lens_release_id": "gemma-2-2b-res-matryoshka-dc", |
| "sae_lens_version": "5.9.1", |
| "sae_cfg_dict": { |
| "architecture": "jumprelu", |
| "d_in": 2304, |
| "d_sae": 32768, |
| "activation_fn_str": "relu", |
| "apply_b_dec_to_input": true, |
| "finetuning_scaling_factor": false, |
| "context_size": 1024, |
| "model_name": "gemma-2-2b", |
| "hook_name": "blocks.20.hook_resid_post", |
| "hook_layer": 20, |
| "hook_head_index": null, |
| "prepend_bos": true, |
| "dataset_path": "chanind/pile-uncopyrighted-gemma-1024-abbrv-1B", |
| "dataset_trust_remote_code": true, |
| "normalize_activations": "none", |
| "dtype": "torch.bfloat16", |
| "device": "cuda", |
| "sae_lens_training_version": "5.5.1", |
| "activation_fn_kwargs": { |
| "k": 40 |
| }, |
| "neuronpedia_id": null, |
| "model_from_pretrained_kwargs": { |
| "center_writing_weights": false |
| }, |
| "seqpos_slice": [ |
| null |
| ] |
| }, |
| "eval_result_unstructured": { |
| "LabHC/bias_in_bios_class_set1_results": { |
| "sae_test_accuracy": { |
| "0": 0.9440000653266907, |
| "1": 0.9650000333786011, |
| "2": 0.9460000395774841, |
| "6": 0.9890000224113464, |
| "9": 0.9780000448226929 |
| }, |
| "llm_test_accuracy": { |
| "0": 0.9540000557899475, |
| "1": 0.968000054359436, |
| "2": 0.9600000381469727, |
| "6": 0.9930000305175781, |
| "9": 0.984000027179718 |
| }, |
| "llm_top_1_test_accuracy": { |
| "0": 0.575, |
| "1": 0.656, |
| "2": 0.674, |
| "6": 0.772, |
| "9": 0.704 |
| }, |
| "llm_top_2_test_accuracy": { |
| "0": 0.591, |
| "1": 0.663, |
| "2": 0.79, |
| "6": 0.859, |
| "9": 0.724 |
| }, |
| "llm_top_5_test_accuracy": { |
| "0": 0.706, |
| "1": 0.716, |
| "2": 0.821, |
| "6": 0.901, |
| "9": 0.913 |
| }, |
| "sae_top_1_test_accuracy": { |
| "0": 0.584, |
| "1": 0.805, |
| "2": 0.885, |
| "6": 0.987, |
| "9": 0.94 |
| }, |
| "sae_top_2_test_accuracy": { |
| "0": 0.879, |
| "1": 0.812, |
| "2": 0.877, |
| "6": 0.988, |
| "9": 0.947 |
| }, |
| "sae_top_5_test_accuracy": { |
| "0": 0.891, |
| "1": 0.938, |
| "2": 0.899, |
| "6": 0.991, |
| "9": 0.948 |
| } |
| }, |
| "LabHC/bias_in_bios_class_set2_results": { |
| "sae_test_accuracy": { |
| "11": 0.9630000591278076, |
| "13": 0.9590000510215759, |
| "14": 0.9530000686645508, |
| "18": 0.9330000281333923, |
| "19": 0.9590000510215759 |
| }, |
| "llm_test_accuracy": { |
| "11": 0.956000030040741, |
| "13": 0.9480000734329224, |
| "14": 0.9690000414848328, |
| "18": 0.9510000348091125, |
| "19": 0.9630000591278076 |
| }, |
| "llm_top_1_test_accuracy": { |
| "11": 0.567, |
| "13": 0.677, |
| "14": 0.637, |
| "18": 0.706, |
| "19": 0.797 |
| }, |
| "llm_top_2_test_accuracy": { |
| "11": 0.648, |
| "13": 0.681, |
| "14": 0.633, |
| "18": 0.726, |
| "19": 0.761 |
| }, |
| "llm_top_5_test_accuracy": { |
| "11": 0.799, |
| "13": 0.725, |
| "14": 0.661, |
| "18": 0.743, |
| "19": 0.807 |
| }, |
| "sae_top_1_test_accuracy": { |
| "11": 0.856, |
| "13": 0.668, |
| "14": 0.912, |
| "18": 0.703, |
| "19": 0.861 |
| }, |
| "sae_top_2_test_accuracy": { |
| "11": 0.87, |
| "13": 0.775, |
| "14": 0.914, |
| "18": 0.915, |
| "19": 0.876 |
| }, |
| "sae_top_5_test_accuracy": { |
| "11": 0.938, |
| "13": 0.888, |
| "14": 0.895, |
| "18": 0.92, |
| "19": 0.92 |
| } |
| }, |
| "LabHC/bias_in_bios_class_set3_results": { |
| "sae_test_accuracy": { |
| "20": 0.956000030040741, |
| "21": 0.9190000295639038, |
| "22": 0.9200000166893005, |
| "25": 0.9540000557899475, |
| "26": 0.8970000147819519 |
| }, |
| "llm_test_accuracy": { |
| "20": 0.9670000672340393, |
| "21": 0.9180000424385071, |
| "22": 0.9100000262260437, |
| "25": 0.9600000381469727, |
| "26": 0.9020000696182251 |
| }, |
| "llm_top_1_test_accuracy": { |
| "20": 0.752, |
| "21": 0.748, |
| "22": 0.66, |
| "25": 0.715, |
| "26": 0.653 |
| }, |
| "llm_top_2_test_accuracy": { |
| "20": 0.81, |
| "21": 0.764, |
| "22": 0.689, |
| "25": 0.717, |
| "26": 0.638 |
| }, |
| "llm_top_5_test_accuracy": { |
| "20": 0.826, |
| "21": 0.803, |
| "22": 0.792, |
| "25": 0.804, |
| "26": 0.743 |
| }, |
| "sae_top_1_test_accuracy": { |
| "20": 0.895, |
| "21": 0.818, |
| "22": 0.838, |
| "25": 0.876, |
| "26": 0.648 |
| }, |
| "sae_top_2_test_accuracy": { |
| "20": 0.924, |
| "21": 0.816, |
| "22": 0.859, |
| "25": 0.869, |
| "26": 0.759 |
| }, |
| "sae_top_5_test_accuracy": { |
| "20": 0.922, |
| "21": 0.835, |
| "22": 0.868, |
| "25": 0.921, |
| "26": 0.814 |
| } |
| }, |
| "canrager/amazon_reviews_mcauley_1and5_results": { |
| "sae_test_accuracy": { |
| "1": 0.9580000638961792, |
| "2": 0.940000057220459, |
| "3": 0.9440000653266907, |
| "5": 0.9290000200271606, |
| "6": 0.893000066280365 |
| }, |
| "llm_test_accuracy": { |
| "1": 0.9590000510215759, |
| "2": 0.9440000653266907, |
| "3": 0.9330000281333923, |
| "5": 0.9270000457763672, |
| "6": 0.8870000243186951 |
| }, |
| "llm_top_1_test_accuracy": { |
| "1": 0.689, |
| "2": 0.616, |
| "3": 0.595, |
| "5": 0.584, |
| "6": 0.679 |
| }, |
| "llm_top_2_test_accuracy": { |
| "1": 0.785, |
| "2": 0.797, |
| "3": 0.665, |
| "5": 0.708, |
| "6": 0.697 |
| }, |
| "llm_top_5_test_accuracy": { |
| "1": 0.829, |
| "2": 0.838, |
| "3": 0.773, |
| "5": 0.803, |
| "6": 0.69 |
| }, |
| "sae_top_1_test_accuracy": { |
| "1": 0.838, |
| "2": 0.877, |
| "3": 0.688, |
| "5": 0.787, |
| "6": 0.619 |
| }, |
| "sae_top_2_test_accuracy": { |
| "1": 0.853, |
| "2": 0.871, |
| "3": 0.719, |
| "5": 0.842, |
| "6": 0.773 |
| }, |
| "sae_top_5_test_accuracy": { |
| "1": 0.91, |
| "2": 0.884, |
| "3": 0.79, |
| "5": 0.898, |
| "6": 0.769 |
| } |
| }, |
| "canrager/amazon_reviews_mcauley_1and5_sentiment_results": { |
| "sae_test_accuracy": { |
| "1.0": 0.9630000591278076, |
| "5.0": 0.9640000462532043 |
| }, |
| "llm_test_accuracy": { |
| "1.0": 0.971000075340271, |
| "5.0": 0.9730000495910645 |
| }, |
| "llm_top_1_test_accuracy": { |
| "1.0": 0.692, |
| "5.0": 0.692 |
| }, |
| "llm_top_2_test_accuracy": { |
| "1.0": 0.744, |
| "5.0": 0.744 |
| }, |
| "llm_top_5_test_accuracy": { |
| "1.0": 0.782, |
| "5.0": 0.782 |
| }, |
| "sae_top_1_test_accuracy": { |
| "1.0": 0.611, |
| "5.0": 0.611 |
| }, |
| "sae_top_2_test_accuracy": { |
| "1.0": 0.808, |
| "5.0": 0.808 |
| }, |
| "sae_top_5_test_accuracy": { |
| "1.0": 0.916, |
| "5.0": 0.916 |
| } |
| }, |
| "codeparrot/github-code_results": { |
| "sae_test_accuracy": { |
| "C": 0.9520000219345093, |
| "Python": 0.987000048160553, |
| "HTML": 0.9800000190734863, |
| "Java": 0.9640000462532043, |
| "PHP": 0.9580000638961792 |
| }, |
| "llm_test_accuracy": { |
| "C": 0.9550000429153442, |
| "Python": 0.9900000691413879, |
| "HTML": 0.9890000224113464, |
| "Java": 0.9590000510215759, |
| "PHP": 0.9540000557899475 |
| }, |
| "llm_top_1_test_accuracy": { |
| "C": 0.561, |
| "Python": 0.604, |
| "HTML": 0.813, |
| "Java": 0.599, |
| "PHP": 0.567 |
| }, |
| "llm_top_2_test_accuracy": { |
| "C": 0.656, |
| "Python": 0.673, |
| "HTML": 0.862, |
| "Java": 0.627, |
| "PHP": 0.643 |
| }, |
| "llm_top_5_test_accuracy": { |
| "C": 0.853, |
| "Python": 0.693, |
| "HTML": 0.908, |
| "Java": 0.793, |
| "PHP": 0.782 |
| }, |
| "sae_top_1_test_accuracy": { |
| "C": 0.875, |
| "Python": 0.938, |
| "HTML": 0.697, |
| "Java": 0.609, |
| "PHP": 0.933 |
| }, |
| "sae_top_2_test_accuracy": { |
| "C": 0.9, |
| "Python": 0.94, |
| "HTML": 0.893, |
| "Java": 0.647, |
| "PHP": 0.935 |
| }, |
| "sae_top_5_test_accuracy": { |
| "C": 0.927, |
| "Python": 0.96, |
| "HTML": 0.955, |
| "Java": 0.896, |
| "PHP": 0.943 |
| } |
| }, |
| "fancyzhx/ag_news_results": { |
| "sae_test_accuracy": { |
| "0": 0.9360000491142273, |
| "1": 0.9810000658035278, |
| "2": 0.9310000538825989, |
| "3": 0.9420000314712524 |
| }, |
| "llm_test_accuracy": { |
| "0": 0.9320000410079956, |
| "1": 0.987000048160553, |
| "2": 0.9320000410079956, |
| "3": 0.9510000348091125 |
| }, |
| "llm_top_1_test_accuracy": { |
| "0": 0.695, |
| "1": 0.618, |
| "2": 0.708, |
| "3": 0.741 |
| }, |
| "llm_top_2_test_accuracy": { |
| "0": 0.798, |
| "1": 0.786, |
| "2": 0.769, |
| "3": 0.779 |
| }, |
| "llm_top_5_test_accuracy": { |
| "0": 0.833, |
| "1": 0.875, |
| "2": 0.808, |
| "3": 0.806 |
| }, |
| "sae_top_1_test_accuracy": { |
| "0": 0.785, |
| "1": 0.951, |
| "2": 0.729, |
| "3": 0.731 |
| }, |
| "sae_top_2_test_accuracy": { |
| "0": 0.877, |
| "1": 0.951, |
| "2": 0.848, |
| "3": 0.825 |
| }, |
| "sae_top_5_test_accuracy": { |
| "0": 0.89, |
| "1": 0.964, |
| "2": 0.837, |
| "3": 0.886 |
| } |
| }, |
| "Helsinki-NLP/europarl_results": { |
| "sae_test_accuracy": { |
| "en": 0.9980000257492065, |
| "fr": 0.999000072479248, |
| "de": 1.0, |
| "es": 1.0, |
| "nl": 1.0 |
| }, |
| "llm_test_accuracy": { |
| "en": 1.0, |
| "fr": 1.0, |
| "de": 1.0, |
| "es": 0.999000072479248, |
| "nl": 1.0 |
| }, |
| "llm_top_1_test_accuracy": { |
| "en": 0.941, |
| "fr": 0.996, |
| "de": 1.0, |
| "es": 0.98, |
| "nl": 1.0 |
| }, |
| "llm_top_2_test_accuracy": { |
| "en": 0.964, |
| "fr": 0.997, |
| "de": 0.999, |
| "es": 0.998, |
| "nl": 1.0 |
| }, |
| "llm_top_5_test_accuracy": { |
| "en": 0.999, |
| "fr": 0.998, |
| "de": 1.0, |
| "es": 1.0, |
| "nl": 1.0 |
| }, |
| "sae_top_1_test_accuracy": { |
| "en": 0.953, |
| "fr": 0.998, |
| "de": 1.0, |
| "es": 0.943, |
| "nl": 1.0 |
| }, |
| "sae_top_2_test_accuracy": { |
| "en": 0.973, |
| "fr": 0.998, |
| "de": 1.0, |
| "es": 0.999, |
| "nl": 1.0 |
| }, |
| "sae_top_5_test_accuracy": { |
| "en": 1.0, |
| "fr": 0.998, |
| "de": 1.0, |
| "es": 0.999, |
| "nl": 1.0 |
| } |
| } |
| } |
| } |
