| { |
| "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": 1745789399708, |
| "eval_result_metrics": { |
| "mean": { |
| "mean_absorption_fraction_score": 0.025265417878406127, |
| "mean_full_absorption_score": 0.008238723662459634, |
| "mean_num_split_features": 1.3461538461538463, |
| "std_dev_absorption_fraction_score": 0.06602091706510793, |
| "std_dev_full_absorption_score": 0.020490530365154826, |
| "std_dev_num_split_features": 0.6894814104706634 |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "first_letter": "a", |
| "mean_absorption_fraction": 0.007794429924657954, |
| "full_absorption_rate": 0.008468052347959968, |
| "num_full_absorption": 22, |
| "num_probe_true_positives": 2598, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "b", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.000630119722747322, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 1587, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "c", |
| "mean_absorption_fraction": 8.787964004499438e-06, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 2667, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "d", |
| "mean_absorption_fraction": 0.0008014102081590394, |
| "full_absorption_rate": 0.0012422360248447205, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1610, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "e", |
| "mean_absorption_fraction": 0.014167078069604516, |
| "full_absorption_rate": 0.006819590824550527, |
| "num_full_absorption": 11, |
| "num_probe_true_positives": 1613, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "f", |
| "mean_absorption_fraction": 0.000206893580744046, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1173, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "g", |
| "mean_absorption_fraction": 0.021764797496677216, |
| "full_absorption_rate": 0.0041876046901172526, |
| "num_full_absorption": 5, |
| "num_probe_true_positives": 1194, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "h", |
| "mean_absorption_fraction": 0.00364815257301705, |
| "full_absorption_rate": 0.0009794319294809011, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 1021, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "i", |
| "mean_absorption_fraction": 0.022980356485917362, |
| "full_absorption_rate": 0.008928571428571428, |
| "num_full_absorption": 15, |
| "num_probe_true_positives": 1680, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "j", |
| "mean_absorption_fraction": 0.0015945614363869333, |
| "full_absorption_rate": 0.0021929824561403508, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 456, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "k", |
| "mean_absorption_fraction": 7.567120399590601e-05, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 716, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "l", |
| "mean_absorption_fraction": 0.0026949480040322758, |
| "full_absorption_rate": 0.0008488964346349745, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 1178, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "m", |
| "mean_absorption_fraction": 0.002033027699140422, |
| "full_absorption_rate": 0.004909983633387889, |
| "num_full_absorption": 9, |
| "num_probe_true_positives": 1833, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "n", |
| "mean_absorption_fraction": 0.01759992284881484, |
| "full_absorption_rate": 0.0043859649122807015, |
| "num_full_absorption": 4, |
| "num_probe_true_positives": 912, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "o", |
| "mean_absorption_fraction": 0.04657255897646143, |
| "full_absorption_rate": 0.014005602240896359, |
| "num_full_absorption": 15, |
| "num_probe_true_positives": 1071, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "p", |
| "mean_absorption_fraction": 0.0002547493519222984, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 2335, |
| "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": 196, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "r", |
| "mean_absorption_fraction": 0.00501576483547139, |
| "full_absorption_rate": 0.0011940298507462687, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1675, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "s", |
| "mean_absorption_fraction": 0.06968499034139855, |
| "full_absorption_rate": 0.018651362984218076, |
| "num_full_absorption": 52, |
| "num_probe_true_positives": 2788, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "t", |
| "mean_absorption_fraction": 0.0013445946368539508, |
| "full_absorption_rate": 0.001156737998843262, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1729, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "u", |
| "mean_absorption_fraction": 0.02094825155333275, |
| "full_absorption_rate": 0.008717310087173101, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 803, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "v", |
| "mean_absorption_fraction": 0.0009320381751078891, |
| "full_absorption_rate": 0.0034762456546929316, |
| "num_full_absorption": 3, |
| "num_probe_true_positives": 863, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "w", |
| "mean_absorption_fraction": 0.04030100556238586, |
| "full_absorption_rate": 0.018439716312056736, |
| "num_full_absorption": 13, |
| "num_probe_true_positives": 705, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "x", |
| "mean_absorption_fraction": 0.016657005093132945, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 88, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "y", |
| "mean_absorption_fraction": 0.3377649759379459, |
| "full_absorption_rate": 0.10497237569060773, |
| "num_full_absorption": 19, |
| "num_probe_true_positives": 181, |
| "num_split_features": 4 |
| }, |
| { |
| "first_letter": "z", |
| "mean_absorption_fraction": 0.022054892879394286, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 263, |
| "num_split_features": 1 |
| } |
| ], |
| "sae_bench_commit_hash": "Unknown", |
| "sae_lens_id": "blocks.2.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.2.hook_resid_post", |
| "hook_layer": 2, |
| "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 |
| } |
