| { |
| "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": 1745792603985, |
| "eval_result_metrics": { |
| "mean": { |
| "mean_absorption_fraction_score": 0.023547590975667807, |
| "mean_full_absorption_score": 0.015102862904013376, |
| "mean_num_split_features": 1.1923076923076923, |
| "std_dev_absorption_fraction_score": 0.05709330666538672, |
| "std_dev_full_absorption_score": 0.03585036003841803, |
| "std_dev_num_split_features": 0.4019184762342502 |
| } |
| }, |
| "eval_result_details": [ |
| { |
| "first_letter": "a", |
| "mean_absorption_fraction": 0.00829131909938039, |
| "full_absorption_rate": 0.009262988320579944, |
| "num_full_absorption": 23, |
| "num_probe_true_positives": 2483, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "b", |
| "mean_absorption_fraction": 0.00015296968723221938, |
| "full_absorption_rate": 0.002570694087403599, |
| "num_full_absorption": 4, |
| "num_probe_true_positives": 1556, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "c", |
| "mean_absorption_fraction": 0.2978497724291809, |
| "full_absorption_rate": 0.1870760442699036, |
| "num_full_absorption": 524, |
| "num_probe_true_positives": 2801, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "d", |
| "mean_absorption_fraction": 0.00935465749209983, |
| "full_absorption_rate": 0.008802816901408451, |
| "num_full_absorption": 15, |
| "num_probe_true_positives": 1704, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "e", |
| "mean_absorption_fraction": 0.004631507456330202, |
| "full_absorption_rate": 0.0037220843672456576, |
| "num_full_absorption": 6, |
| "num_probe_true_positives": 1612, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "f", |
| "mean_absorption_fraction": 0.013394373337784406, |
| "full_absorption_rate": 0.008223684210526315, |
| "num_full_absorption": 10, |
| "num_probe_true_positives": 1216, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "g", |
| "mean_absorption_fraction": 0.03266509636691238, |
| "full_absorption_rate": 0.00684931506849315, |
| "num_full_absorption": 8, |
| "num_probe_true_positives": 1168, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "h", |
| "mean_absorption_fraction": 0.0038251595998291146, |
| "full_absorption_rate": 0.003929273084479371, |
| "num_full_absorption": 4, |
| "num_probe_true_positives": 1018, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "i", |
| "mean_absorption_fraction": 0.004747891209880004, |
| "full_absorption_rate": 0.004186602870813397, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 1672, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "j", |
| "mean_absorption_fraction": 0.002386634844868735, |
| "full_absorption_rate": 0.00477326968973747, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 419, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "k", |
| "mean_absorption_fraction": 0.0, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 652, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "l", |
| "mean_absorption_fraction": 0.012960995970106304, |
| "full_absorption_rate": 0.004058441558441558, |
| "num_full_absorption": 5, |
| "num_probe_true_positives": 1232, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "m", |
| "mean_absorption_fraction": 0.007512726826765947, |
| "full_absorption_rate": 0.025081788440567066, |
| "num_full_absorption": 46, |
| "num_probe_true_positives": 1834, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "n", |
| "mean_absorption_fraction": 0.02256460890928388, |
| "full_absorption_rate": 0.008274231678486997, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 846, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "o", |
| "mean_absorption_fraction": 0.0024460203870805054, |
| "full_absorption_rate": 0.004512635379061372, |
| "num_full_absorption": 5, |
| "num_probe_true_positives": 1108, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "p", |
| "mean_absorption_fraction": 0.043701865747193855, |
| "full_absorption_rate": 0.02186026575225032, |
| "num_full_absorption": 51, |
| "num_probe_true_positives": 2333, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "q", |
| "mean_absorption_fraction": 0.02134641404818411, |
| "full_absorption_rate": 0.009569377990430622, |
| "num_full_absorption": 2, |
| "num_probe_true_positives": 209, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "r", |
| "mean_absorption_fraction": 0.01003972588494661, |
| "full_absorption_rate": 0.004281345565749235, |
| "num_full_absorption": 7, |
| "num_probe_true_positives": 1635, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "s", |
| "mean_absorption_fraction": 0.03080760972610719, |
| "full_absorption_rate": 0.019559902200488997, |
| "num_full_absorption": 56, |
| "num_probe_true_positives": 2863, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "t", |
| "mean_absorption_fraction": 0.0006094791143597142, |
| "full_absorption_rate": 0.0018326206475259622, |
| "num_full_absorption": 3, |
| "num_probe_true_positives": 1637, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "u", |
| "mean_absorption_fraction": 0.029777669596100006, |
| "full_absorption_rate": 0.022696929238985315, |
| "num_full_absorption": 17, |
| "num_probe_true_positives": 749, |
| "num_split_features": 2 |
| }, |
| { |
| "first_letter": "v", |
| "mean_absorption_fraction": 0.0029767931655123887, |
| "full_absorption_rate": 0.0038119440914866584, |
| "num_full_absorption": 3, |
| "num_probe_true_positives": 787, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "w", |
| "mean_absorption_fraction": 0.01349058182363775, |
| "full_absorption_rate": 0.007246376811594203, |
| "num_full_absorption": 5, |
| "num_probe_true_positives": 690, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "x", |
| "mean_absorption_fraction": 0.007020790870657701, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 101, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "y", |
| "mean_absorption_fraction": 0.016181185131639735, |
| "full_absorption_rate": 0.0, |
| "num_full_absorption": 0, |
| "num_probe_true_positives": 165, |
| "num_split_features": 1 |
| }, |
| { |
| "first_letter": "z", |
| "mean_absorption_fraction": 0.013501516642289125, |
| "full_absorption_rate": 0.020491803278688523, |
| "num_full_absorption": 5, |
| "num_probe_true_positives": 244, |
| "num_split_features": 1 |
| } |
| ], |
| "sae_bench_commit_hash": "Unknown", |
| "sae_lens_id": "blocks.9.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.9.hook_resid_post", |
| "hook_layer": 9, |
| "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 |
| } |
