| { |
| "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": 1745719249324, |
| "eval_result_metrics": { |
| "mean": { |
| "mean_absorption_fraction_score": 0.00849608699429072, |
| "mean_full_absorption_score": 0.005332496852924044, |
| "mean_num_split_features": 1.1923076923076923, |
| "std_dev_absorption_fraction_score": 0.011257446352611816, |
| "std_dev_full_absorption_score": 0.0054153913868388364, |
| "std_dev_num_split_features": 0.4914656259988704 |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "first_letter": "a", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0015349194167306216, |
| "num_full_absorption": 4, |
| "num_probe_true_positives": 2606, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "b", |
| "mean_absorption_fraction": 0.0017258586584180155, |
| "full_absorption_rate": 0.007352941176470588, |
| "num_full_absorption": 12, |
| "num_probe_true_positives": 1632, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "c", |
| "mean_absorption_fraction": 0.0007065002407756726, |
| "full_absorption_rate": 0.0003619254433586681, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 2763, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "d", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.004667444574095682, |
| "num_full_absorption": 8, |
| "num_probe_true_positives": 1714, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "e", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0012391573729863693, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1614, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "f", |
| "mean_absorption_fraction": 0.022999228274488366, |
| "full_absorption_rate": 0.007518796992481203, |
| "num_full_absorption": 9, |
| "num_probe_true_positives": 1197, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "g", |
| "mean_absorption_fraction": 0.017162911537986707, |
| "full_absorption_rate": 0.005235602094240838, |
| "num_full_absorption": 6, |
| "num_probe_true_positives": 1146, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "h", |
| "mean_absorption_fraction": 0.00977314513127678, |
| "full_absorption_rate": 0.008695652173913044, |
| "num_full_absorption": 9, |
| "num_probe_true_positives": 1035, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "i", |
| "mean_absorption_fraction": 0.0013778418224468837, |
| "full_absorption_rate": 0.003486345148169669, |
| "num_full_absorption": 6, |
| "num_probe_true_positives": 1721, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "j", |
| "mean_absorption_fraction": 0.00856764628821801, |
| "full_absorption_rate": 0.00909090909090909, |
| "num_full_absorption": 4, |
| "num_probe_true_positives": 440, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "k", |
| "mean_absorption_fraction": 0.002424164950516951, |
| "full_absorption_rate": 0.001451378809869376, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 689, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "l", |
| "mean_absorption_fraction": 0.0001241204628941615, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 1220, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "m", |
| "mean_absorption_fraction": 0.004420601412823913, |
| "full_absorption_rate": 0.0011435105774728416, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 1749, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "n", |
| "mean_absorption_fraction": 0.0026002795862142694, |
| "full_absorption_rate": 0.0011148272017837235, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 897, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "o", |
| "mean_absorption_fraction": 0.016910753534243702, |
| "full_absorption_rate": 0.014299332697807437, |
| "num_full_absorption": 15, |
| "num_probe_true_positives": 1049, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "p", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0008547008547008547, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 2340, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "q", |
| "mean_absorption_fraction": 0.007804842632591571, |
| "full_absorption_rate": 0.006134969325153374, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 163, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "r", |
| "mean_absorption_fraction": 0.029781651136315337, |
| "full_absorption_rate": 0.014226437462951986, |
| "num_full_absorption": 24, |
| "num_probe_true_positives": 1687, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "s", |
| "mean_absorption_fraction": 0.00312977151299912, |
| "full_absorption_rate": 0.0024946543121881683, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 2806, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "t", |
| "mean_absorption_fraction": 0.0010544198516813646, |
| "full_absorption_rate": 0.0006038647342995169, |
| "num_full_absorption": 1, |
| "num_probe_true_positives": 1656, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "u", |
| "mean_absorption_fraction": 0.043622843678462524, |
| "full_absorption_rate": 0.017834394904458598, |
| "num_full_absorption": 14, |
| "num_probe_true_positives": 785, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "v", |
| "mean_absorption_fraction": 0.0001906515352730492, |
| "full_absorption_rate": 0.0026560424966799467, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 753, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "w", |
| "mean_absorption_fraction": 0.0215796713199841, |
| "full_absorption_rate": 0.01422475106685633, |
| "num_full_absorption": 10, |
| "num_probe_true_positives": 703, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "x", |
| "mean_absorption_fraction": 0.0027988336693133666, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 109, |
| "num_split_features": 3 |
| }, |
| { |
| "first_letter": "y", |
| "mean_absorption_fraction": 0.019691286254293176, |
| "full_absorption_rate": 0.012422360248447204, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 161, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "z", |
| "mean_absorption_fraction": 0.002451238360341718, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 227, |
| "num_split_features": 1 |
| } |
| ], |
| "sae_bench_commit_hash": "Unknown", |
| "sae_lens_id": "blocks.8.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.8.hook_resid_post", |
| "hook_layer": 8, |
| "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 |
| } |
