| { |
| "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": 1745790312572, |
| "eval_result_metrics": { |
| "mean": { |
| "mean_absorption_fraction_score": 0.028922321159520942, |
| "mean_full_absorption_score": 0.008418731636494817, |
| "mean_num_split_features": 1.1153846153846154, |
| "std_dev_absorption_fraction_score": 0.07276556379168583, |
| "std_dev_full_absorption_score": 0.016881431034780812, |
| "std_dev_num_split_features": 0.4314554973040049 |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "first_letter": "a", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.005436893203883495, |
| "num_full_absorption": 14, |
| "num_probe_true_positives": 2575, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "b", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0006093845216331506, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 1641, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "c", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 2810, |
| "num_split_features": 3 |
| }, |
| { |
| "first_letter": "d", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0023894862604540022, |
| "num_full_absorption": 4, |
| "num_probe_true_positives": 1674, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "e", |
| "mean_absorption_fraction": 0.0006199628022318661, |
| "full_absorption_rate": 0.0024798512089274642, |
| "num_full_absorption": 4, |
| "num_probe_true_positives": 1613, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "f", |
| "mean_absorption_fraction": 0.0002242419032101037, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1233, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "g", |
| "mean_absorption_fraction": 0.07104513439969147, |
| "full_absorption_rate": 0.03295750216825672, |
| "num_full_absorption": 38, |
| "num_probe_true_positives": 1153, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "h", |
| "mean_absorption_fraction": 0.0012143927337092796, |
| "full_absorption_rate": 0.0019267822736030828, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1038, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "i", |
| "mean_absorption_fraction": 0.016950678015300244, |
| "full_absorption_rate": 0.007986309184255563, |
| "num_full_absorption": 14, |
| "num_probe_true_positives": 1753, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "j", |
| "mean_absorption_fraction": 0.0027721177083826947, |
| "full_absorption_rate": 0.0022271714922048997, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 449, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "k", |
| "mean_absorption_fraction": 0.0004085789609711224, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 692, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "l", |
| "mean_absorption_fraction": 0.0008278145695364238, |
| "full_absorption_rate": 0.0016556291390728477, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1208, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "m", |
| "mean_absorption_fraction": 0.00021969080063217948, |
| "full_absorption_rate": 0.0011467889908256881, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1744, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "n", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 853, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "o", |
| "mean_absorption_fraction": 0.2596632633275213, |
| "full_absorption_rate": 0.07621671258034894, |
| "num_full_absorption": 83, |
| "num_probe_true_positives": 1089, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "p", |
| "mean_absorption_fraction": 0.000317967921429193, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 2351, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "q", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 190, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "r", |
| "mean_absorption_fraction": 0.0009382711087891736, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1633, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "s", |
| "mean_absorption_fraction": 0.0020414337761666887, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 2815, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "t", |
| "mean_absorption_fraction": 3.448365519713449e-05, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1669, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "u", |
| "mean_absorption_fraction": 0.09877487655241585, |
| "full_absorption_rate": 0.03259452411994785, |
| "num_full_absorption": 25, |
| "num_probe_true_positives": 767, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "v", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0012531328320802004, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 798, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "w", |
| "mean_absorption_fraction": 0.012216565410305706, |
| "full_absorption_rate": 0.020618556701030927, |
| "num_full_absorption": 14, |
| "num_probe_true_positives": 679, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "x", |
| "mean_absorption_fraction": 0.016452318658276028, |
| "full_absorption_rate": 0.010638297872340425, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 94, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "y", |
| "mean_absorption_fraction": 0.26677269716722374, |
| "full_absorption_rate": 0.01875, |
| "num_full_absorption": 3, |
| "num_probe_true_positives": 160, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "z", |
| "mean_absorption_fraction": 0.0004858606765542463, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 260, |
| "num_split_features": 1 |
| } |
| ], |
| "sae_bench_commit_hash": "Unknown", |
| "sae_lens_id": "blocks.4.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.4.hook_resid_post", |
| "hook_layer": 4, |
| "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 |
| } |
