| { |
| "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": "7dac3d47-5282-48db-a705-23f393c4c7d6", |
| "datetime_epoch_millis": 1745716276080, |
| "eval_result_metrics": { |
| "mean": { |
| "mean_absorption_fraction_score": 0.009665541385659575, |
| "mean_full_absorption_score": 0.0039931709319401395, |
| "mean_num_split_features": 1.0769230769230769, |
| "std_dev_absorption_fraction_score": 0.0172017685171009, |
| "std_dev_full_absorption_score": 0.007227095384732639, |
| "std_dev_num_split_features": 0.271746488194703 |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "first_letter": "a", |
| "mean_absorption_fraction": 0.00133968812090033, |
| "full_absorption_rate": 0.0042884990253411305, |
| "num_full_absorption": 11, |
| "num_probe_true_positives": 2565, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "b", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0006353240152477764, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 1574, |
| "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": 2824, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "d", |
| "mean_absorption_fraction": 0.001223133484162896, |
| "full_absorption_rate": 0.004273504273504274, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 1638, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "e", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0012763241863433313, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1567, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "f", |
| "mean_absorption_fraction": 0.0009425948412100823, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1271, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "g", |
| "mean_absorption_fraction": 0.04500370764988289, |
| "full_absorption_rate": 0.011494252873563218, |
| "num_full_absorption": 13, |
| "num_probe_true_positives": 1131, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "h", |
| "mean_absorption_fraction": 0.0013868011426805046, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1028, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "i", |
| "mean_absorption_fraction": 0.0053024231452206945, |
| "full_absorption_rate": 0.004139562389118865, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 1691, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "j", |
| "mean_absorption_fraction": 0.0013942248292305476, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 457, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "k", |
| "mean_absorption_fraction": 0.019929101906275545, |
| "full_absorption_rate": 0.012430939226519336, |
| "num_full_absorption": 9, |
| "num_probe_true_positives": 724, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "l", |
| "mean_absorption_fraction": 0.0002673083517990925, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1215, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "m", |
| "mean_absorption_fraction": 0.0011564088092700054, |
| "full_absorption_rate": 0.004988913525498891, |
| "num_full_absorption": 9, |
| "num_probe_true_positives": 1804, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "n", |
| "mean_absorption_fraction": 0.0004708634224627405, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 879, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "o", |
| "mean_absorption_fraction": 0.059274200348401, |
| "full_absorption_rate": 0.030977734753146177, |
| "num_full_absorption": 32, |
| "num_probe_true_positives": 1033, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "p", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0004178854993731718, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 2393, |
| "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": 183, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "r", |
| "mean_absorption_fraction": 0.004679395472931736, |
| "full_absorption_rate": 0.000578368999421631, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 1729, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "s", |
| "mean_absorption_fraction": 0.0013022813230770304, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 2794, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "t", |
| "mean_absorption_fraction": 0.0026578798763838103, |
| "full_absorption_rate": 0.0006053268765133172, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 1652, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "u", |
| "mean_absorption_fraction": 0.023502639373924047, |
| "full_absorption_rate": 0.008706467661691543, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 804, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "v", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0012578616352201257, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 795, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "w", |
| "mean_absorption_fraction": 0.012226853237001789, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 650, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "x", |
| "mean_absorption_fraction": 0.01074709366786789, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 92, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "y", |
| "mean_absorption_fraction": 0.05468277438228415, |
| "full_absorption_rate": 0.01775147928994083, |
| "num_full_absorption": 3, |
| "num_probe_true_positives": 169, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "z", |
| "mean_absorption_fraction": 0.0038147026421821873, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 257, |
| "num_split_features": 1 |
| } |
| ], |
| "sae_bench_commit_hash": "Unknown", |
| "sae_lens_id": "blocks.5.hook_resid_post", |
| "sae_lens_release_id": "gemma-2-2b-res-snap-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.5.hook_resid_post", |
| "hook_layer": 5, |
| "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 |
| } |
