| { |
| "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": 1745790771224, |
| "eval_result_metrics": { |
| "mean": { |
| "mean_absorption_fraction_score": 0.004897855752677167, |
| "mean_full_absorption_score": 0.0023701556913451483, |
| "mean_num_split_features": 1.1153846153846154, |
| "std_dev_absorption_fraction_score": 0.009731170558331502, |
| "std_dev_full_absorption_score": 0.005763844961692995, |
| "std_dev_num_split_features": 0.3258125936084211 |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "first_letter": "a", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0007797270955165692, |
| "num_full_absorption": 2, |
| "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": 0.0043267862479381865, |
| "full_absorption_rate": 0.003663003663003663, |
| "num_full_absorption": 6, |
| "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": 1 |
| }, |
| { |
| "first_letter": "f", |
| "mean_absorption_fraction": 0.00032661809730080796, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1271, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "g", |
| "mean_absorption_fraction": 0.0017683465959328027, |
| "full_absorption_rate": 0.0017683465959328027, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1131, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "h", |
| "mean_absorption_fraction": 0.005684681434291037, |
| "full_absorption_rate": 0.0019455252918287938, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1028, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "i", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.002365464222353637, |
| "num_full_absorption": 4, |
| "num_probe_true_positives": 1691, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "j", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 457, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "k", |
| "mean_absorption_fraction": 0.0015621679770505738, |
| "full_absorption_rate": 0.0013812154696132596, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 724, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "l", |
| "mean_absorption_fraction": 5.967078189300411e-05, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1215, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "m", |
| "mean_absorption_fraction": 0.000868313136268892, |
| "full_absorption_rate": 0.004988913525498891, |
| "num_full_absorption": 9, |
| "num_probe_true_positives": 1804, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "n", |
| "mean_absorption_fraction": 0.016172503654382123, |
| "full_absorption_rate": 0.0022753128555176336, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 879, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "o", |
| "mean_absorption_fraction": 0.0012622126763743582, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1033, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "p", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 2393, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "q", |
| "mean_absorption_fraction": 0.012199244515399827, |
| "full_absorption_rate": 0.00546448087431694, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 183, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "r", |
| "mean_absorption_fraction": 0.004042892858515715, |
| "full_absorption_rate": 0.001735106998264893, |
| "num_full_absorption": 3, |
| "num_probe_true_positives": 1729, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "s", |
| "mean_absorption_fraction": 0.0004242735557323247, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 2794, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "t", |
| "mean_absorption_fraction": 0.00017217478167990625, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1652, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "u", |
| "mean_absorption_fraction": 0.014155204950431925, |
| "full_absorption_rate": 0.0012437810945273632, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 804, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "v", |
| "mean_absorption_fraction": 0.002609080188679245, |
| "full_absorption_rate": 0.0025157232704402514, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 795, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "w", |
| "mean_absorption_fraction": 0.006593761196813876, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 650, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "x", |
| "mean_absorption_fraction": 0.007183094749869198, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 92, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "y", |
| "mean_absorption_fraction": 0.046923850178248884, |
| "full_absorption_rate": 0.029585798816568046, |
| "num_full_absorption": 5, |
| "num_probe_true_positives": 169, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "z", |
| "mean_absorption_fraction": 0.0010093719928036483, |
| "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-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 |
| } |
