| { |
| "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": 1745791236588, |
| "eval_result_metrics": { |
| "mean": { |
| "mean_absorption_fraction_score": 0.012628375228045254, |
| "mean_full_absorption_score": 0.005509152897684141, |
| "mean_num_split_features": 1.0384615384615385, |
| "std_dev_absorption_fraction_score": 0.026055714121596404, |
| "std_dev_full_absorption_score": 0.010808806408357102, |
| "std_dev_num_split_features": 0.19611613513818404 |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "first_letter": "a", |
| "mean_absorption_fraction": 0.00038182512409316535, |
| "full_absorption_rate": 0.0007636502481863307, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 2619, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "b", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1614, |
| "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": 2751, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "d", |
| "mean_absorption_fraction": 0.00027670937030250983, |
| "full_absorption_rate": 0.0011926058437686344, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1677, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "e", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0042997542997543, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 1628, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "f", |
| "mean_absorption_fraction": 0.000803473550264643, |
| "full_absorption_rate": 0.0008136696501220504, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 1229, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "g", |
| "mean_absorption_fraction": 0.027705099694449065, |
| "full_absorption_rate": 0.010251630941286114, |
| "num_full_absorption": 11, |
| "num_probe_true_positives": 1073, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "h", |
| "mean_absorption_fraction": 0.004430582307466775, |
| "full_absorption_rate": 0.003727865796831314, |
| "num_full_absorption": 4, |
| "num_probe_true_positives": 1073, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "i", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.005172413793103448, |
| "num_full_absorption": 9, |
| "num_probe_true_positives": 1740, |
| "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": 449, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "k", |
| "mean_absorption_fraction": 0.00011954126046899571, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 670, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "l", |
| "mean_absorption_fraction": 3.878152323728303e-05, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1278, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "m", |
| "mean_absorption_fraction": 0.0006451130149381448, |
| "full_absorption_rate": 0.014168937329700272, |
| "num_full_absorption": 26, |
| "num_probe_true_positives": 1835, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "n", |
| "mean_absorption_fraction": 0.027800108092991965, |
| "full_absorption_rate": 0.01346389228886169, |
| "num_full_absorption": 11, |
| "num_probe_true_positives": 817, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "o", |
| "mean_absorption_fraction": 0.10578539376770135, |
| "full_absorption_rate": 0.0531496062992126, |
| "num_full_absorption": 54, |
| "num_probe_true_positives": 1016, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "p", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0008768084173608067, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 2281, |
| "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": 167, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "r", |
| "mean_absorption_fraction": 0.0023568206854501237, |
| "full_absorption_rate": 0.0012113870381586917, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1651, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "s", |
| "mean_absorption_fraction": 0.007736173244554238, |
| "full_absorption_rate": 0.0020833333333333333, |
| "num_full_absorption": 6, |
| "num_probe_true_positives": 2880, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "t", |
| "mean_absorption_fraction": 0.002051458257589933, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1647, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "u", |
| "mean_absorption_fraction": 0.08022923032695359, |
| "full_absorption_rate": 0.015776699029126214, |
| "num_full_absorption": 13, |
| "num_probe_true_positives": 824, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "v", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.003484320557491289, |
| "num_full_absorption": 3, |
| "num_probe_true_positives": 861, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "w", |
| "mean_absorption_fraction": 0.009599272963032644, |
| "full_absorption_rate": 0.001519756838905775, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 658, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "x", |
| "mean_absorption_fraction": 0.03486746063991207, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 81, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "y", |
| "mean_absorption_fraction": 0.023510712105770097, |
| "full_absorption_rate": 0.006756756756756757, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 148, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "z", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.004524886877828055, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 221, |
| "num_split_features": 1 |
| } |
| ], |
| "sae_bench_commit_hash": "Unknown", |
| "sae_lens_id": "blocks.6.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.6.hook_resid_post", |
| "hook_layer": 6, |
| "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 |
| } |
