| { |
| "eval_type_id": "absorption_first_letter", |
| "eval_config": { |
| "model_name": "gemma-2-2b", |
| "random_seed": 42, |
| "f1_jump_threshold": 0.03, |
| "max_k_value": 10, |
| "prompt_template": "{word} has the first letter:", |
| "prompt_token_pos": -6, |
| "llm_batch_size": 32, |
| "llm_dtype": "bfloat16", |
| "k_sparse_probe_l1_decay": 0.01, |
| "k_sparse_probe_batch_size": 4096, |
| "k_sparse_probe_num_epochs": 50, |
| "min_GT_probe_f1": 0.6, |
| "min_feats_for_eval": 20 |
| }, |
| "eval_id": "a15aaf76-1ebc-4334-8744-86b1c8367c42", |
| "datetime_epoch_millis": 1745788497924, |
| "eval_result_metrics": { |
| "mean": { |
| "mean_absorption_fraction_score": 0.2260316648063969, |
| "mean_full_absorption_score": 0.13278002787230536, |
| "mean_num_split_features": 1.1923076923076923, |
| "std_dev_absorption_fraction_score": 0.20590129279781358, |
| "std_dev_full_absorption_score": 0.14410153428657962, |
| "std_dev_num_split_features": 0.4019184762342502 |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "first_letter": "a", |
| "mean_absorption_fraction": 0.22514921276779798, |
| "full_absorption_rate": 0.08057054741711642, |
| "num_full_absorption": 209, |
| "num_probe_true_positives": 2594, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "b", |
| "mean_absorption_fraction": 0.2277279475982472, |
| "full_absorption_rate": 0.1568505889646621, |
| "num_full_absorption": 253, |
| "num_probe_true_positives": 1613, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "c", |
| "mean_absorption_fraction": 0.011814578412448167, |
| "full_absorption_rate": 0.0021156558533145277, |
| "num_full_absorption": 6, |
| "num_probe_true_positives": 2836, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "d", |
| "mean_absorption_fraction": 0.03679793400241639, |
| "full_absorption_rate": 0.021711366538952746, |
| "num_full_absorption": 34, |
| "num_probe_true_positives": 1566, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "e", |
| "mean_absorption_fraction": 0.7646689149429188, |
| "full_absorption_rate": 0.49209361163820364, |
| "num_full_absorption": 778, |
| "num_probe_true_positives": 1581, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "f", |
| "mean_absorption_fraction": 0.10120346478510622, |
| "full_absorption_rate": 0.0825, |
| "num_full_absorption": 99, |
| "num_probe_true_positives": 1200, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "g", |
| "mean_absorption_fraction": 0.23002042191653788, |
| "full_absorption_rate": 0.13848920863309352, |
| "num_full_absorption": 154, |
| "num_probe_true_positives": 1112, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "h", |
| "mean_absorption_fraction": 0.04832866591113763, |
| "full_absorption_rate": 0.02213666987487969, |
| "num_full_absorption": 23, |
| "num_probe_true_positives": 1039, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "i", |
| "mean_absorption_fraction": 0.1669916872337898, |
| "full_absorption_rate": 0.09781357882623705, |
| "num_full_absorption": 170, |
| "num_probe_true_positives": 1738, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "j", |
| "mean_absorption_fraction": 0.03470937748836968, |
| "full_absorption_rate": 0.020737327188940093, |
| "num_full_absorption": 9, |
| "num_probe_true_positives": 434, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "k", |
| "mean_absorption_fraction": 0.05059134305567325, |
| "full_absorption_rate": 0.015942028985507246, |
| "num_full_absorption": 11, |
| "num_probe_true_positives": 690, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "l", |
| "mean_absorption_fraction": 0.49009290319513177, |
| "full_absorption_rate": 0.17087542087542087, |
| "num_full_absorption": 203, |
| "num_probe_true_positives": 1188, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "m", |
| "mean_absorption_fraction": 0.150902833096729, |
| "full_absorption_rate": 0.09877913429522753, |
| "num_full_absorption": 178, |
| "num_probe_true_positives": 1802, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "n", |
| "mean_absorption_fraction": 0.17090269948864342, |
| "full_absorption_rate": 0.08675263774912075, |
| "num_full_absorption": 74, |
| "num_probe_true_positives": 853, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "o", |
| "mean_absorption_fraction": 0.8418210875108045, |
| "full_absorption_rate": 0.586046511627907, |
| "num_full_absorption": 630, |
| "num_probe_true_positives": 1075, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "p", |
| "mean_absorption_fraction": 0.21381634890348136, |
| "full_absorption_rate": 0.24552429667519182, |
| "num_full_absorption": 576, |
| "num_probe_true_positives": 2346, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "q", |
| "mean_absorption_fraction": 0.003247505032563331, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 181, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "r", |
| "mean_absorption_fraction": 0.2316999431395606, |
| "full_absorption_rate": 0.16723744292237444, |
| "num_full_absorption": 293, |
| "num_probe_true_positives": 1752, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "s", |
| "mean_absorption_fraction": 0.34066669738538546, |
| "full_absorption_rate": 0.3405947441217151, |
| "num_full_absorption": 985, |
| "num_probe_true_positives": 2892, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "t", |
| "mean_absorption_fraction": 0.17312080660525656, |
| "full_absorption_rate": 0.11402985074626866, |
| "num_full_absorption": 191, |
| "num_probe_true_positives": 1675, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "u", |
| "mean_absorption_fraction": 0.3163494409878043, |
| "full_absorption_rate": 0.16934487021013597, |
| "num_full_absorption": 137, |
| "num_probe_true_positives": 809, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "v", |
| "mean_absorption_fraction": 0.16677238420706927, |
| "full_absorption_rate": 0.05463182897862233, |
| "num_full_absorption": 46, |
| "num_probe_true_positives": 842, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "w", |
| "mean_absorption_fraction": 0.12255263018322858, |
| "full_absorption_rate": 0.08175182481751825, |
| "num_full_absorption": 56, |
| "num_probe_true_positives": 685, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "x", |
| "mean_absorption_fraction": 0.14203166174446147, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 102, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "y", |
| "mean_absorption_fraction": 0.26291835609236486, |
| "full_absorption_rate": 0.08928571428571429, |
| "num_full_absorption": 15, |
| "num_probe_true_positives": 168, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "z", |
| "mean_absorption_fraction": 0.3519244392793917, |
| "full_absorption_rate": 0.11646586345381527, |
| "num_full_absorption": 29, |
| "num_probe_true_positives": 249, |
| "num_split_features": 1 |
| } |
| ], |
| "sae_bench_commit_hash": "Unknown", |
| "sae_lens_id": "blocks.0.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.0.hook_resid_post", |
| "hook_layer": 0, |
| "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": null |
| } |
