| { |
| "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": 1745788946355, |
| "eval_result_metrics": { |
| "mean": { |
| "mean_absorption_fraction_score": 0.12822587064301322, |
| "mean_full_absorption_score": 0.06927551272584227, |
| "mean_num_split_features": 1.4230769230769231, |
| "std_dev_absorption_fraction_score": 0.12088385768549416, |
| "std_dev_full_absorption_score": 0.0884381471712126, |
| "std_dev_num_split_features": 0.6433087546786176 |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "first_letter": "a", |
| "mean_absorption_fraction": 0.06305141828683253, |
| "full_absorption_rate": 0.03353658536585366, |
| "num_full_absorption": 88, |
| "num_probe_true_positives": 2624, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "b", |
| "mean_absorption_fraction": 0.005521077898885913, |
| "full_absorption_rate": 0.006369426751592357, |
| "num_full_absorption": 10, |
| "num_probe_true_positives": 1570, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "c", |
| "mean_absorption_fraction": 0.002498460676244265, |
| "full_absorption_rate": 0.007650273224043716, |
| "num_full_absorption": 21, |
| "num_probe_true_positives": 2745, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "d", |
| "mean_absorption_fraction": 0.024653066513312515, |
| "full_absorption_rate": 0.01464307504575961, |
| "num_full_absorption": 24, |
| "num_probe_true_positives": 1639, |
| "num_split_features": 3 |
| }, |
| { |
| "first_letter": "e", |
| "mean_absorption_fraction": 0.12575011195417476, |
| "full_absorption_rate": 0.08044554455445545, |
| "num_full_absorption": 130, |
| "num_probe_true_positives": 1616, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "f", |
| "mean_absorption_fraction": 0.20668002379891073, |
| "full_absorption_rate": 0.13257243195785778, |
| "num_full_absorption": 151, |
| "num_probe_true_positives": 1139, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "g", |
| "mean_absorption_fraction": 0.507928222376814, |
| "full_absorption_rate": 0.3305084745762712, |
| "num_full_absorption": 390, |
| "num_probe_true_positives": 1180, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "h", |
| "mean_absorption_fraction": 0.045398167063210766, |
| "full_absorption_rate": 0.027384324834749764, |
| "num_full_absorption": 29, |
| "num_probe_true_positives": 1059, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "i", |
| "mean_absorption_fraction": 0.2899860704317634, |
| "full_absorption_rate": 0.17153067301965455, |
| "num_full_absorption": 288, |
| "num_probe_true_positives": 1679, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "j", |
| "mean_absorption_fraction": 0.0868958918215932, |
| "full_absorption_rate": 0.011655011655011656, |
| "num_full_absorption": 5, |
| "num_probe_true_positives": 429, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "k", |
| "mean_absorption_fraction": 0.17611995859054244, |
| "full_absorption_rate": 0.060029282576866766, |
| "num_full_absorption": 41, |
| "num_probe_true_positives": 683, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "l", |
| "mean_absorption_fraction": 0.16221972834485193, |
| "full_absorption_rate": 0.06845003933910307, |
| "num_full_absorption": 87, |
| "num_probe_true_positives": 1271, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "m", |
| "mean_absorption_fraction": 0.02495034338795541, |
| "full_absorption_rate": 0.007972665148063782, |
| "num_full_absorption": 14, |
| "num_probe_true_positives": 1756, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "n", |
| "mean_absorption_fraction": 0.16732994430549802, |
| "full_absorption_rate": 0.07279236276849642, |
| "num_full_absorption": 61, |
| "num_probe_true_positives": 838, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "o", |
| "mean_absorption_fraction": 0.17612145890053107, |
| "full_absorption_rate": 0.09541627689429373, |
| "num_full_absorption": 102, |
| "num_probe_true_positives": 1069, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "p", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0028594771241830064, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 2448, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "q", |
| "mean_absorption_fraction": 0.005924277265558018, |
| "full_absorption_rate": 0.0053475935828877, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 187, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "r", |
| "mean_absorption_fraction": 0.14295658120933505, |
| "full_absorption_rate": 0.07221542227662178, |
| "num_full_absorption": 118, |
| "num_probe_true_positives": 1634, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "s", |
| "mean_absorption_fraction": 0.38510893164025983, |
| "full_absorption_rate": 0.3370634113258732, |
| "num_full_absorption": 994, |
| "num_probe_true_positives": 2949, |
| "num_split_features": 3 |
| }, |
| { |
| "first_letter": "t", |
| "mean_absorption_fraction": 0.10542675152429691, |
| "full_absorption_rate": 0.06776664702416028, |
| "num_full_absorption": 115, |
| "num_probe_true_positives": 1697, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "u", |
| "mean_absorption_fraction": 0.10771569631823806, |
| "full_absorption_rate": 0.047182175622542594, |
| "num_full_absorption": 36, |
| "num_probe_true_positives": 763, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "v", |
| "mean_absorption_fraction": 0.11569975095381459, |
| "full_absorption_rate": 0.056511056511056514, |
| "num_full_absorption": 46, |
| "num_probe_true_positives": 814, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "w", |
| "mean_absorption_fraction": 0.09800221402576327, |
| "full_absorption_rate": 0.048295454545454544, |
| "num_full_absorption": 34, |
| "num_probe_true_positives": 704, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "x", |
| "mean_absorption_fraction": 0.14995182832350804, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 98, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "y", |
| "mean_absorption_fraction": 0.15494858117312987, |
| "full_absorption_rate": 0.03550295857988166, |
| "num_full_absorption": 6, |
| "num_probe_true_positives": 169, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "z", |
| "mean_absorption_fraction": 0.00303407993331871, |
| "full_absorption_rate": 0.007462686567164179, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 268, |
| "num_split_features": 2 |
| } |
| ], |
| "sae_bench_commit_hash": "Unknown", |
| "sae_lens_id": "blocks.1.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.1.hook_resid_post", |
| "hook_layer": 1, |
| "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 |
| } |
