chanind commited on
Commit
54f1264
·
verified ·
1 Parent(s): 8123876

Upload folder using huggingface_hub

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +12 -0
  2. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/cfg.json +1 -0
  3. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/feature_absorption/sae/layer_12_sae.parquet +3 -0
  4. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/k_sparse_probing/sae/layer_12_sae_metadata.parquet +3 -0
  5. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/k_sparse_probing/sae/layer_12_sae_metrics.parquet +3 -0
  6. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/k_sparse_probing/sae/layer_12_sae_raw_results.parquet +3 -0
  7. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/runner_cfg.json +1 -0
  8. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json +45 -0
  9. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json +45 -0
  10. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json +45 -0
  11. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json +45 -0
  12. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json +45 -0
  13. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json +45 -0
  14. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json +45 -0
  15. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json +45 -0
  16. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json +45 -0
  17. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json +45 -0
  18. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json +45 -0
  19. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json +45 -0
  20. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json +45 -0
  21. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json +45 -0
  22. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json +45 -0
  23. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json +45 -0
  24. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json +45 -0
  25. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json +45 -0
  26. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json +45 -0
  27. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json +45 -0
  28. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json +45 -0
  29. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json +45 -0
  30. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json +45 -0
  31. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json +45 -0
  32. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json +45 -0
  33. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json +45 -0
  34. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json +45 -0
  35. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json +45 -0
  36. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json +45 -0
  37. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json +45 -0
  38. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json +45 -0
  39. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json +45 -0
  40. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json +45 -0
  41. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json +45 -0
  42. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json +45 -0
  43. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json +45 -0
  44. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json +45 -0
  45. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json +45 -0
  46. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json +45 -0
  47. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json +45 -0
  48. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json +45 -0
  49. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json +45 -0
  50. baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json +45 -0
.gitattributes CHANGED
@@ -48,3 +48,15 @@ baseline/gemma-2-2b-btk/k-150/seed-0/model.layers.12/w-32768/t-300M/l0-143.2/sae
48
  baseline/gemma-2-2b-btk/k-200/seed-0/model.layers.12/w-32768/t-300M/l0-193.0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
49
  baseline/gemma-2-2b-btk/k-250/seed-0/model.layers.12/w-32768/t-300M/l0-243.2/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
50
  baseline/gemma-2-2b-btk/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-45.0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
48
  baseline/gemma-2-2b-btk/k-200/seed-0/model.layers.12/w-32768/t-300M/l0-193.0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
49
  baseline/gemma-2-2b-btk/k-250/seed-0/model.layers.12/w-32768/t-300M/l0-243.2/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
50
  baseline/gemma-2-2b-btk/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-45.0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
51
+ baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
52
+ baseline/gemma-2-2b-btk-mat/k-100/seed-1/model.layers.12/w-32768/t-300M/l0-95.0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
53
+ baseline/gemma-2-2b-btk-mat/k-150/seed-0/model.layers.12/w-32768/t-300M/l0-144.9/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
54
+ baseline/gemma-2-2b-btk-mat/k-150/seed-1/model.layers.12/w-32768/t-300M/l0-145.0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
55
+ baseline/gemma-2-2b-btk-mat/k-200/seed-0/model.layers.12/w-32768/t-300M/l0-194.8/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
56
+ baseline/gemma-2-2b-btk-mat/k-200/seed-2/model.layers.12/w-32768/t-300M/l0-194.9/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
57
+ baseline/gemma-2-2b-btk-mat/k-250/seed-0/model.layers.12/w-32768/t-300M/l0-245.1/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
58
+ baseline/gemma-2-2b-btk-mat/k-250/seed-1/model.layers.12/w-32768/t-300M/l0-245.1/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
59
+ baseline/gemma-2-2b-btk-mat/k-250/seed-2/model.layers.12/w-32768/t-300M/l0-245.0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
60
+ baseline/gemma-2-2b-btk-mat/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-45.4/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
61
+ baseline/gemma-2-2b-btk-mat/k-50/seed-1/model.layers.12/w-32768/t-300M/l0-45.6/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
62
+ baseline/gemma-2-2b-btk-mat/k-50/seed-2/model.layers.12/w-32768/t-300M/l0-45.5/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/cfg.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"apply_b_dec_to_input": true, "normalize_activations": "none", "metadata": {"sae_lens_version": "6.37.3", "sae_lens_training_version": "6.37.3", "model_name": "gemma-2-2b", "hook_name": "blocks.12.hook_resid_post", "hook_head_index": null, "context_size": 1024, "prepend_bos": true, "seqpos_slice": [null], "model_from_pretrained_kwargs": {"center_writing_weights": false}, "dataset_path": "chanind/pile-uncopyrighted-gemma-1024-abbrv-2B", "model_class_name": "AutoModelForCausalLM", "exclude_special_tokens": true, "sequence_separator_token": "bos", "disable_concat_sequences": false}, "dtype": "float32", "d_in": 2304, "d_sae": 32768, "reshape_activations": "none", "device": "cuda", "architecture": "jumprelu"}
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/feature_absorption/sae/layer_12_sae.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d6930226ce8246aa6dc12075a8033111721584de26c83c3daec20489de8325a4
3
+ size 11128085
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/k_sparse_probing/sae/layer_12_sae_metadata.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8ab6883a8f26ef18523d1f94fa35774ae9f04aae8800268544d9f418c90907fa
3
+ size 22838
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/k_sparse_probing/sae/layer_12_sae_metrics.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:669a9d7794befddf126cc2bfc4a5c8e7b8379dcd7a6f7966ac11183d4bee71da
3
+ size 142823
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/k_sparse_probing/sae/layer_12_sae_raw_results.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:64b0bd284d28ad9a9d2d02716c0efd8d15da2895f6ae8c23a9a786d503005122
3
+ size 118798284
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/runner_cfg.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"sae": {"d_in": 2304, "d_sae": 32768, "dtype": "float32", "device": "cpu", "apply_b_dec_to_input": true, "normalize_activations": "none", "reshape_activations": "none", "metadata": {"sae_lens_version": "6.37.0", "sae_lens_training_version": "6.37.0"}, "decoder_init_norm": 0.1, "k": 100, "use_sparse_activations": false, "aux_loss_coefficient": 1.0, "rescale_acts_by_decoder_norm": true, "topk_threshold_lr": 0.01, "matryoshka_widths": [2048, 8192, 32768], "matryoshka_loss_multipliers": null, "skip_final_matryoshka_width": true, "include_outer_loss": true, "detach_matryoshka_losses": false, "normalize_reconstruction_losses_by_d_in": false, "normalize_losses_by_num_matryoshka_steps": false, "matryoshka_loss_probabilities": 1.0, "initial_matryoshka_loss_probabilities": null, "transition_matryoshka_loss_probabilities_duration": 0, "transition_matryoshka_loss_probabilities_start_step": 0, "pin_matryoshka_encoder": false, "pin_matryoshka_encoder_at_step": 0, "pin_loss_coefficient": 1.0, "use_matryoshka_aux_loss": true, "architecture": "xmatryoshka_batchtopk"}, "model_name": "google/gemma-2-2b", "model_class_name": "AutoModelForCausalLM", "hook_name": "model.layers.12", "hook_eval": "NOT_IN_USE", "hook_head_index": null, "dataset_path": "chanind/pile-uncopyrighted-gemma-1024-abbrv-2B", "dataset_trust_remote_code": true, "streaming": false, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "from_pretrained_path": null, "n_batches_in_buffer": 64, "training_tokens": 300000000, "store_batch_size_prompts": 12, "seqpos_slice": [null], "disable_concat_sequences": false, "sequence_separator_token": "bos", "activations_mixing_fraction": 0.5, "device": "cuda", "act_store_device": "cuda", "seed": 0, "dtype": "float32", "prepend_bos": true, "autocast": true, "autocast_lm": true, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "train_batch_size_tokens": 4096, "adam_beta1": 0.9, "adam_beta2": 0.999, "lr": 0.0003, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 2.9999999999999997e-05, "lr_decay_steps": 14648, "n_restart_cycles": 1, "dead_feature_window": 1000, "feature_sampling_window": 2000, "dead_feature_threshold": 1e-08, "n_eval_batches": 10, "eval_batch_size_prompts": 6, "logger": {"log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "log_weights_to_wandb": true, "wandb_project": "sae-rethink", "wandb_id": null, "run_name": "btk-mat-k-100-layer-12-2026-02-21T15:01:43", "wandb_entity": "chanind", "wandb_log_frequency": 10, "eval_every_n_wandb_logs": 100}, "n_checkpoints": 2, "checkpoint_path": "/home/dev/project-storage/shared/checkpoints/2c1f1718c03876f1aba79153eb09f1986d219c379ff40dc473aad6e15d2b8faf", "save_final_checkpoint": false, "output_path": "/home/dev/project-storage/saes/refinement-300t/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8", "resume_from_checkpoint": null, "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "6.37.0", "sae_lens_training_version": "6.37.0", "exclude_special_tokens": true, "b_dec_init_method": "zeros"}
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.9694229846591836,
4
+ "test_acc": 0.9694510739856802,
5
+ "test_auc": 0.9694656488549618,
6
+ "val_auc": 0.972972972972973,
7
+ "k": 1,
8
+ "dataset": "100_news_fake",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 3641
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9985680184405502,
18
+ "test_acc": 0.9985680190930788,
19
+ "test_auc": 0.9999143317512048,
20
+ "val_auc": 1.0,
21
+ "k": 16,
22
+ "dataset": "100_news_fake",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 3641,
28
+ 11117,
29
+ 3714,
30
+ 277,
31
+ 1463,
32
+ 1857,
33
+ 1314,
34
+ 1227,
35
+ 1425,
36
+ 52,
37
+ 1079,
38
+ 1145,
39
+ 43,
40
+ 2156,
41
+ 1566,
42
+ 1882
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.9338934548324378,
4
+ "test_acc": 0.9338950097213221,
5
+ "test_auc": 0.966521172288193,
6
+ "val_auc": 0.9466721785861927,
7
+ "k": 1,
8
+ "dataset": "105_click_bait",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1647
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9643438301288979,
18
+ "test_acc": 0.9643551523007129,
19
+ "test_auc": 0.9810706101355483,
20
+ "val_auc": 0.9818106655642828,
21
+ "k": 16,
22
+ "dataset": "105_click_bait",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1647,
28
+ 3268,
29
+ 1223,
30
+ 1681,
31
+ 27059,
32
+ 1425,
33
+ 1256,
34
+ 992,
35
+ 2362,
36
+ 1062,
37
+ 277,
38
+ 1692,
39
+ 1697,
40
+ 7287,
41
+ 43,
42
+ 939
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5592935556133121,
4
+ "test_acc": 0.5733333333333334,
5
+ "test_auc": 0.5959511472458301,
6
+ "val_auc": 0.5966386554621848,
7
+ "k": 1,
8
+ "dataset": "106_hate_hate",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 2007
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5985774473358663,
18
+ "test_acc": 0.6088050314465409,
19
+ "test_auc": 0.6361947477517343,
20
+ "val_auc": 0.6326530612244898,
21
+ "k": 16,
22
+ "dataset": "106_hate_hate",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 2007,
28
+ 3512,
29
+ 570,
30
+ 1286,
31
+ 3768,
32
+ 776,
33
+ 3714,
34
+ 23329,
35
+ 288,
36
+ 1972,
37
+ 300,
38
+ 43,
39
+ 925,
40
+ 1037,
41
+ 71,
42
+ 23075
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.6238427790876061,
4
+ "test_acc": 0.6528301886792452,
5
+ "test_auc": 0.6555989181186768,
6
+ "val_auc": 0.6790716286514605,
7
+ "k": 1,
8
+ "dataset": "107_hate_offensive",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 15341
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7187736013665411,
18
+ "test_acc": 0.72,
19
+ "test_auc": 0.7753235568418057,
20
+ "val_auc": 0.8167266906762705,
21
+ "k": 16,
22
+ "dataset": "107_hate_offensive",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 15341,
28
+ 611,
29
+ 2000,
30
+ 1149,
31
+ 1196,
32
+ 5682,
33
+ 7469,
34
+ 1972,
35
+ 304,
36
+ 1247,
37
+ 24482,
38
+ 154,
39
+ 11775,
40
+ 6718,
41
+ 19782,
42
+ 1702
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.7847110890235477,
4
+ "test_acc": 0.7849056603773585,
5
+ "test_auc": 0.8274008672062573,
6
+ "val_auc": 0.8729491796718688,
7
+ "k": 1,
8
+ "dataset": "110_aimade_humangpt3",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1425
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7906867899893731,
18
+ "test_acc": 0.7906918238993711,
19
+ "test_auc": 0.8996813037257263,
20
+ "val_auc": 0.9099639855942377,
21
+ "k": 16,
22
+ "dataset": "110_aimade_humangpt3",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1425,
28
+ 1586,
29
+ 1062,
30
+ 443,
31
+ 1857,
32
+ 1119,
33
+ 1494,
34
+ 13902,
35
+ 1647,
36
+ 611,
37
+ 1725,
38
+ 1383,
39
+ 506,
40
+ 6676,
41
+ 1871,
42
+ 1702
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.8423610823905348,
4
+ "test_acc": 0.8425157232704402,
5
+ "test_auc": 0.8761276769828836,
6
+ "val_auc": 0.9139655862344938,
7
+ "k": 1,
8
+ "dataset": "113_movie_sent",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 971
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8802117251639489,
18
+ "test_acc": 0.880251572327044,
19
+ "test_auc": 0.9439720861657109,
20
+ "val_auc": 0.9623849539815926,
21
+ "k": 16,
22
+ "dataset": "113_movie_sent",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 971,
28
+ 2007,
29
+ 1950,
30
+ 16406,
31
+ 491,
32
+ 875,
33
+ 30291,
34
+ 1362,
35
+ 11318,
36
+ 6980,
37
+ 1871,
38
+ 970,
39
+ 28832,
40
+ 3108,
41
+ 20164,
42
+ 749
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5149097030915213,
4
+ "test_acc": 0.5466666666666666,
5
+ "test_auc": 0.5502364762513885,
6
+ "val_auc": 0.5308037535699714,
7
+ "k": 1,
8
+ "dataset": "114_nyc_borough_Manhattan",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 3207
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6158791744168511,
18
+ "test_acc": 0.6194871794871795,
19
+ "test_auc": 0.6686302891574377,
20
+ "val_auc": 0.6209710322317421,
21
+ "k": 16,
22
+ "dataset": "114_nyc_borough_Manhattan",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 3207,
28
+ 1948,
29
+ 823,
30
+ 572,
31
+ 1504,
32
+ 1142,
33
+ 69,
34
+ 1474,
35
+ 567,
36
+ 1596,
37
+ 906,
38
+ 27059,
39
+ 7680,
40
+ 7492,
41
+ 5890,
42
+ 6172
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5250082840967224,
4
+ "test_acc": 0.5251282051282051,
5
+ "test_auc": 0.5291513784629885,
6
+ "val_auc": 0.5157078743370053,
7
+ "k": 1,
8
+ "dataset": "115_nyc_borough_Brooklyn",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 379
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5425936863770348,
18
+ "test_acc": 0.5435897435897435,
19
+ "test_auc": 0.5369273235264416,
20
+ "val_auc": 0.5630354957160343,
21
+ "k": 16,
22
+ "dataset": "115_nyc_borough_Brooklyn",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 379,
28
+ 69,
29
+ 6234,
30
+ 1107,
31
+ 1167,
32
+ 25552,
33
+ 1358,
34
+ 1142,
35
+ 6524,
36
+ 823,
37
+ 37,
38
+ 895,
39
+ 1876,
40
+ 524,
41
+ 20491,
42
+ 1658
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5033684672112249,
4
+ "test_acc": 0.5456410256410257,
5
+ "test_auc": 0.547703403238294,
6
+ "val_auc": 0.5203998368013055,
7
+ "k": 1,
8
+ "dataset": "116_nyc_borough_Bronx",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 21577
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5342322889206697,
18
+ "test_acc": 0.5353846153846153,
19
+ "test_auc": 0.5592326724341063,
20
+ "val_auc": 0.5893512851897184,
21
+ "k": 16,
22
+ "dataset": "116_nyc_borough_Bronx",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 21577,
28
+ 642,
29
+ 191,
30
+ 69,
31
+ 333,
32
+ 1365,
33
+ 1742,
34
+ 926,
35
+ 506,
36
+ 1167,
37
+ 532,
38
+ 5159,
39
+ 1948,
40
+ 25768,
41
+ 1087,
42
+ 1358
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.8460372598303634,
4
+ "test_acc": 0.8484848484848485,
5
+ "test_auc": 0.856326530612245,
6
+ "val_auc": 0.8645833333333333,
7
+ "k": 1,
8
+ "dataset": "117_us_state_FL",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 8580
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8181818181818182,
18
+ "test_acc": 0.8181818181818182,
19
+ "test_auc": 0.8926530612244897,
20
+ "val_auc": 0.9645432692307692,
21
+ "k": 16,
22
+ "dataset": "117_us_state_FL",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 8580,
28
+ 4803,
29
+ 7802,
30
+ 4979,
31
+ 69,
32
+ 551,
33
+ 1544,
34
+ 1847,
35
+ 532,
36
+ 4776,
37
+ 2289,
38
+ 1226,
39
+ 6074,
40
+ 1420,
41
+ 5577,
42
+ 29
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.8780918288711562,
4
+ "test_acc": 0.8787878787878788,
5
+ "test_auc": 0.9095918367346938,
6
+ "val_auc": 0.9419070512820512,
7
+ "k": 1,
8
+ "dataset": "118_us_state_CA",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 7260
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8686064482736934,
18
+ "test_acc": 0.8686868686868687,
19
+ "test_auc": 0.9555102040816326,
20
+ "val_auc": 0.9883814102564104,
21
+ "k": 16,
22
+ "dataset": "118_us_state_CA",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 7260,
28
+ 3038,
29
+ 6207,
30
+ 1894,
31
+ 825,
32
+ 5494,
33
+ 4979,
34
+ 5301,
35
+ 5577,
36
+ 4776,
37
+ 11202,
38
+ 7350,
39
+ 1474,
40
+ 3529,
41
+ 5586,
42
+ 1769
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.8671979042043984,
4
+ "test_acc": 0.8686868686868687,
5
+ "test_auc": 0.8775510204081632,
6
+ "val_auc": 0.8233173076923076,
7
+ "k": 1,
8
+ "dataset": "119_us_state_TX",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 7110
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8988661121014063,
18
+ "test_acc": 0.898989898989899,
19
+ "test_auc": 0.9653061224489795,
20
+ "val_auc": 0.9711538461538461,
21
+ "k": 16,
22
+ "dataset": "119_us_state_TX",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 7110,
28
+ 7680,
29
+ 603,
30
+ 4246,
31
+ 1711,
32
+ 1475,
33
+ 6207,
34
+ 4776,
35
+ 5301,
36
+ 7350,
37
+ 4979,
38
+ 1876,
39
+ 7573,
40
+ 4542,
41
+ 461,
42
+ 4943
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.6453671587093286,
4
+ "test_acc": 0.6614651061985263,
5
+ "test_auc": 0.67957374402696,
6
+ "val_auc": 0.7667999999999999,
7
+ "k": 1,
8
+ "dataset": "120_us_timezone_Chicago",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 7260
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7975725290464437,
18
+ "test_acc": 0.7975726051148678,
19
+ "test_auc": 0.8953239307901473,
20
+ "val_auc": 0.9176,
21
+ "k": 16,
22
+ "dataset": "120_us_timezone_Chicago",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 7260,
28
+ 7350,
29
+ 3038,
30
+ 1894,
31
+ 1876,
32
+ 7110,
33
+ 6207,
34
+ 1475,
35
+ 11202,
36
+ 551,
37
+ 29,
38
+ 265,
39
+ 2265,
40
+ 4776,
41
+ 3529,
42
+ 4943
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.6287319067563097,
4
+ "test_acc": 0.6458604247941049,
5
+ "test_auc": 0.6689188478252048,
6
+ "val_auc": 0.696,
7
+ "k": 1,
8
+ "dataset": "121_us_timezone_New_York",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 7260
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8157643326565608,
18
+ "test_acc": 0.8166449934980494,
19
+ "test_auc": 0.9156912642928327,
20
+ "val_auc": 0.9312,
21
+ "k": 16,
22
+ "dataset": "121_us_timezone_New_York",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 7260,
28
+ 4776,
29
+ 7141,
30
+ 3529,
31
+ 6894,
32
+ 2056,
33
+ 7110,
34
+ 7350,
35
+ 3039,
36
+ 265,
37
+ 4992,
38
+ 1711,
39
+ 171,
40
+ 4979,
41
+ 7573,
42
+ 3538
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.8107212964272528,
4
+ "test_acc": 0.811443433029909,
5
+ "test_auc": 0.8552837071853848,
6
+ "val_auc": 0.8687999999999999,
7
+ "k": 1,
8
+ "dataset": "122_us_timezone_Los_Angeles",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 7260
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8375636179289772,
18
+ "test_acc": 0.8391850888599913,
19
+ "test_auc": 0.9246758136787312,
20
+ "val_auc": 0.954,
21
+ "k": 16,
22
+ "dataset": "122_us_timezone_Los_Angeles",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 7260,
28
+ 3038,
29
+ 825,
30
+ 6207,
31
+ 1894,
32
+ 5494,
33
+ 11202,
34
+ 7350,
35
+ 4776,
36
+ 1876,
37
+ 171,
38
+ 1474,
39
+ 4998,
40
+ 29,
41
+ 461,
42
+ 551
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.9569225331461187,
4
+ "test_acc": 0.9569230769230769,
5
+ "test_auc": 0.9831773656040664,
6
+ "val_auc": 0.9951040391676866,
7
+ "k": 1,
8
+ "dataset": "123_world_country_United_Kingdom",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 2372
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9702457059556757,
18
+ "test_acc": 0.9702564102564103,
19
+ "test_auc": 0.9915760595145925,
20
+ "val_auc": 0.9991840065279478,
21
+ "k": 16,
22
+ "dataset": "123_world_country_United_Kingdom",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 2372,
28
+ 4119,
29
+ 7680,
30
+ 1240,
31
+ 1475,
32
+ 962,
33
+ 1711,
34
+ 629,
35
+ 419,
36
+ 1945,
37
+ 2704,
38
+ 29699,
39
+ 4246,
40
+ 1474,
41
+ 1876,
42
+ 30766
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.8414508544270659,
4
+ "test_acc": 0.8451282051282051,
5
+ "test_auc": 0.8456340256505201,
6
+ "val_auc": 0.8837209302325582,
7
+ "k": 1,
8
+ "dataset": "124_world_country_United_States",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 3515
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9487140643026115,
18
+ "test_acc": 0.9487179487179487,
19
+ "test_auc": 0.9912562695660956,
20
+ "val_auc": 0.9983680130558956,
21
+ "k": 16,
22
+ "dataset": "124_world_country_United_States",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 3515,
28
+ 4105,
29
+ 30620,
30
+ 2704,
31
+ 629,
32
+ 1951,
33
+ 962,
34
+ 8079,
35
+ 7877,
36
+ 346,
37
+ 1483,
38
+ 1959,
39
+ 6418,
40
+ 398,
41
+ 946,
42
+ 7680
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.9220315980315978,
4
+ "test_acc": 0.9220512820512821,
5
+ "test_auc": 0.9685133133604874,
6
+ "val_auc": 0.9787841697266421,
7
+ "k": 1,
8
+ "dataset": "125_world_country_Italy",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1951
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9764080228267914,
18
+ "test_acc": 0.9764102564102564,
19
+ "test_auc": 0.996591712391019,
20
+ "val_auc": 1.0,
21
+ "k": 16,
22
+ "dataset": "125_world_country_Italy",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1951,
28
+ 6929,
29
+ 1945,
30
+ 962,
31
+ 2014,
32
+ 3941,
33
+ 11428,
34
+ 2372,
35
+ 629,
36
+ 544,
37
+ 567,
38
+ 30620,
39
+ 1959,
40
+ 171,
41
+ 398,
42
+ 4119
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5706545728894823,
4
+ "test_acc": 0.5726051148677936,
5
+ "test_auc": 0.5969143113962371,
6
+ "val_auc": 0.5952000000000001,
7
+ "k": 1,
8
+ "dataset": "126_art_type_book",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1681
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6892328706914621,
18
+ "test_acc": 0.6913740788903338,
19
+ "test_auc": 0.7646580918438975,
20
+ "val_auc": 0.8032,
21
+ "k": 16,
22
+ "dataset": "126_art_type_book",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1681,
28
+ 4916,
29
+ 1063,
30
+ 282,
31
+ 18329,
32
+ 396,
33
+ 1697,
34
+ 4309,
35
+ 25819,
36
+ 1288,
37
+ 1599,
38
+ 11542,
39
+ 43,
40
+ 360,
41
+ 6885,
42
+ 1358
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.6157958481505632,
4
+ "test_acc": 0.6441265713047247,
5
+ "test_auc": 0.6489498422471107,
6
+ "val_auc": 0.6192000000000002,
7
+ "k": 1,
8
+ "dataset": "127_art_type_song",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 30230
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7490184826687019,
18
+ "test_acc": 0.7490247074122237,
19
+ "test_auc": 0.8271843025728978,
20
+ "val_auc": 0.8338,
21
+ "k": 16,
22
+ "dataset": "127_art_type_song",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 30230,
28
+ 25819,
29
+ 2000,
30
+ 1954,
31
+ 903,
32
+ 328,
33
+ 1504,
34
+ 7765,
35
+ 1365,
36
+ 823,
37
+ 6737,
38
+ 659,
39
+ 1696,
40
+ 1486,
41
+ 1149,
42
+ 1288
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5264326509246606,
4
+ "test_acc": 0.5834416991764196,
5
+ "test_auc": 0.5880338533642175,
6
+ "val_auc": 0.6018,
7
+ "k": 1,
8
+ "dataset": "128_art_type_movie",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 629
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6351584958660529,
18
+ "test_acc": 0.6384915474642393,
19
+ "test_auc": 0.7103735113433272,
20
+ "val_auc": 0.7572,
21
+ "k": 16,
22
+ "dataset": "128_art_type_movie",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 629,
28
+ 1240,
29
+ 1959,
30
+ 1486,
31
+ 2000,
32
+ 1067,
33
+ 25819,
34
+ 5245,
35
+ 903,
36
+ 3492,
37
+ 1037,
38
+ 30230,
39
+ 1696,
40
+ 1945,
41
+ 8950,
42
+ 733
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.6423098629116595,
4
+ "test_acc": 0.6439232409381663,
5
+ "test_auc": 0.6667030369158029,
6
+ "val_auc": 0.5880000000000001,
7
+ "k": 1,
8
+ "dataset": "129_arith_mc_A",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 351
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7245634886630589,
18
+ "test_acc": 0.7249466950959488,
19
+ "test_auc": 0.8233133296963084,
20
+ "val_auc": 0.7984,
21
+ "k": 16,
22
+ "dataset": "129_arith_mc_A",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 351,
28
+ 1239,
29
+ 4657,
30
+ 1217,
31
+ 27547,
32
+ 543,
33
+ 880,
34
+ 964,
35
+ 10049,
36
+ 1806,
37
+ 821,
38
+ 6629,
39
+ 711,
40
+ 2545,
41
+ 2036,
42
+ 6273
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.9393939393939394,
4
+ "test_acc": 0.9393939393939394,
5
+ "test_auc": 0.9622448979591837,
6
+ "val_auc": 0.9524959742351048,
7
+ "k": 1,
8
+ "dataset": "130_temp_cat_Frequency",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 744
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9898969278371137,
18
+ "test_acc": 0.98989898989899,
19
+ "test_auc": 0.9991836734693877,
20
+ "val_auc": 1.0,
21
+ "k": 16,
22
+ "dataset": "130_temp_cat_Frequency",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 744,
28
+ 5915,
29
+ 1881,
30
+ 1024,
31
+ 28850,
32
+ 585,
33
+ 628,
34
+ 1716,
35
+ 1703,
36
+ 1305,
37
+ 25508,
38
+ 15453,
39
+ 500,
40
+ 131,
41
+ 487,
42
+ 1741
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.8076885358438757,
4
+ "test_acc": 0.8080808080808081,
5
+ "test_auc": 0.8375510204081632,
6
+ "val_auc": 0.8305152979066022,
7
+ "k": 1,
8
+ "dataset": "131_temp_cat_Typical Time",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 24006
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9494846391855681,
18
+ "test_acc": 0.9494949494949495,
19
+ "test_auc": 0.9869387755102041,
20
+ "val_auc": 0.9665861513687601,
21
+ "k": 16,
22
+ "dataset": "131_temp_cat_Typical Time",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 24006,
28
+ 500,
29
+ 443,
30
+ 372,
31
+ 251,
32
+ 1305,
33
+ 2863,
34
+ 843,
35
+ 786,
36
+ 744,
37
+ 521,
38
+ 28850,
39
+ 970,
40
+ 4660,
41
+ 7034,
42
+ 18329
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.9393568033273916,
4
+ "test_acc": 0.9393939393939394,
5
+ "test_auc": 0.956734693877551,
6
+ "val_auc": 0.93317230273752,
7
+ "k": 1,
8
+ "dataset": "132_temp_cat_Event Ordering",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 251
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.98989898989899,
18
+ "test_acc": 0.98989898989899,
19
+ "test_auc": 0.9995918367346939,
20
+ "val_auc": 0.998792270531401,
21
+ "k": 16,
22
+ "dataset": "132_temp_cat_Event Ordering",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 251,
28
+ 1563,
29
+ 9806,
30
+ 19411,
31
+ 1500,
32
+ 1978,
33
+ 13938,
34
+ 1206,
35
+ 843,
36
+ 4400,
37
+ 1872,
38
+ 7053,
39
+ 1703,
40
+ 30332,
41
+ 9857,
42
+ 1970
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.9288804594067752,
4
+ "test_acc": 0.9288888888888889,
5
+ "test_auc": 0.9500237041719343,
6
+ "val_auc": 0.9283685064935064,
7
+ "k": 1,
8
+ "dataset": "133_context_type_Causality",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1388
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9555502871292346,
18
+ "test_acc": 0.9555555555555556,
19
+ "test_auc": 0.9737673830594185,
20
+ "val_auc": 0.9904626623376623,
21
+ "k": 16,
22
+ "dataset": "133_context_type_Causality",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1388,
28
+ 144,
29
+ 2820,
30
+ 32303,
31
+ 30332,
32
+ 32667,
33
+ 7853,
34
+ 1977,
35
+ 3014,
36
+ 2863,
37
+ 32623,
38
+ 48,
39
+ 1223,
40
+ 971,
41
+ 174,
42
+ 131
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.7671416697614374,
4
+ "test_acc": 0.7688888888888888,
5
+ "test_auc": 0.8085493046776233,
6
+ "val_auc": 0.7633928571428571,
7
+ "k": 1,
8
+ "dataset": "134_context_type_Belief_states",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1219
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8132226613965745,
18
+ "test_acc": 0.8133333333333334,
19
+ "test_auc": 0.9151390644753477,
20
+ "val_auc": 0.8681006493506493,
21
+ "k": 16,
22
+ "dataset": "134_context_type_Belief_states",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1219,
28
+ 1748,
29
+ 1040,
30
+ 174,
31
+ 620,
32
+ 131,
33
+ 1608,
34
+ 2265,
35
+ 1168,
36
+ 1230,
37
+ 1824,
38
+ 1256,
39
+ 22999,
40
+ 22168,
41
+ 1425,
42
+ 20645
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.9155555555555556,
4
+ "test_acc": 0.9155555555555556,
5
+ "test_auc": 0.9476927939317319,
6
+ "val_auc": 0.9261363636363636,
7
+ "k": 1,
8
+ "dataset": "135_context_type_Event_duration",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 251
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.96,
18
+ "test_acc": 0.96,
19
+ "test_auc": 0.9888590391908976,
20
+ "val_auc": 0.9849837662337663,
21
+ "k": 16,
22
+ "dataset": "135_context_type_Event_duration",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 251,
28
+ 843,
29
+ 363,
30
+ 13938,
31
+ 28850,
32
+ 927,
33
+ 1741,
34
+ 585,
35
+ 5526,
36
+ 500,
37
+ 103,
38
+ 131,
39
+ 144,
40
+ 4400,
41
+ 30332,
42
+ 1500
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.6570665536271968,
4
+ "test_acc": 0.6571304724750758,
5
+ "test_auc": 0.698900163990855,
6
+ "val_auc": 0.7302,
7
+ "k": 1,
8
+ "dataset": "136_glue_mnli_entailment",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1223
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7496306790811983,
18
+ "test_acc": 0.7503250975292588,
19
+ "test_auc": 0.834907354937237,
20
+ "val_auc": 0.896,
21
+ "k": 16,
22
+ "dataset": "136_glue_mnli_entailment",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1223,
28
+ 2012,
29
+ 2954,
30
+ 21568,
31
+ 7582,
32
+ 6869,
33
+ 1332,
34
+ 29034,
35
+ 943,
36
+ 545,
37
+ 970,
38
+ 749,
39
+ 1386,
40
+ 929,
41
+ 1362,
42
+ 4920
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5216814153849533,
4
+ "test_acc": 0.5431296055483311,
5
+ "test_auc": 0.5534492192021117,
6
+ "val_auc": 0.5902,
7
+ "k": 1,
8
+ "dataset": "137_glue_mnli_neutral",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 970
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.681500433500969,
18
+ "test_acc": 0.682271348071088,
19
+ "test_auc": 0.7526909681773566,
20
+ "val_auc": 0.7992,
21
+ "k": 16,
22
+ "dataset": "137_glue_mnli_neutral",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 970,
28
+ 1223,
29
+ 443,
30
+ 1362,
31
+ 2954,
32
+ 2012,
33
+ 21568,
34
+ 1586,
35
+ 141,
36
+ 28905,
37
+ 1242,
38
+ 7582,
39
+ 749,
40
+ 1324,
41
+ 1119,
42
+ 5763
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.6137042067487427,
4
+ "test_acc": 0.6263545730385782,
5
+ "test_auc": 0.6470134424401117,
6
+ "val_auc": 0.6994,
7
+ "k": 1,
8
+ "dataset": "138_glue_mnli_contradiction",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 970
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7306848163248093,
18
+ "test_acc": 0.7334200260078023,
19
+ "test_auc": 0.8277832975840284,
20
+ "val_auc": 0.7992,
21
+ "k": 16,
22
+ "dataset": "138_glue_mnli_contradiction",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 970,
28
+ 1362,
29
+ 28905,
30
+ 2954,
31
+ 6869,
32
+ 7582,
33
+ 2012,
34
+ 21568,
35
+ 29034,
36
+ 1283,
37
+ 61,
38
+ 90,
39
+ 1332,
40
+ 545,
41
+ 1223,
42
+ 929
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.7664083282712159,
4
+ "test_acc": 0.7749152542372881,
5
+ "test_auc": 0.7808739010049531,
6
+ "val_auc": 0.7685185185185186,
7
+ "k": 1,
8
+ "dataset": "139_news_class_Politics",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 2040
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8425485997423536,
18
+ "test_acc": 0.8427118644067797,
19
+ "test_auc": 0.930813412611738,
20
+ "val_auc": 0.9464573268921096,
21
+ "k": 16,
22
+ "dataset": "139_news_class_Politics",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 2040,
28
+ 10581,
29
+ 15698,
30
+ 3268,
31
+ 1566,
32
+ 1816,
33
+ 992,
34
+ 3641,
35
+ 1463,
36
+ 1290,
37
+ 21427,
38
+ 1276,
39
+ 17259,
40
+ 202,
41
+ 1692,
42
+ 1857
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.7568174041843337,
4
+ "test_acc": 0.7654237288135594,
5
+ "test_auc": 0.7742330476221994,
6
+ "val_auc": 0.6644524959742352,
7
+ "k": 1,
8
+ "dataset": "140_news_class_Technology",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1276
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8320907460474617,
18
+ "test_acc": 0.8325423728813559,
19
+ "test_auc": 0.9144934970380911,
20
+ "val_auc": 0.8466183574879227,
21
+ "k": 16,
22
+ "dataset": "140_news_class_Technology",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1276,
28
+ 1816,
29
+ 1290,
30
+ 22618,
31
+ 5098,
32
+ 12639,
33
+ 4957,
34
+ 1692,
35
+ 2035,
36
+ 20639,
37
+ 4306,
38
+ 129,
39
+ 689,
40
+ 52,
41
+ 842,
42
+ 2040
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.6042275488862886,
4
+ "test_acc": 0.6061016949152542,
5
+ "test_auc": 0.6397272690501667,
6
+ "val_auc": 0.7000805152979066,
7
+ "k": 1,
8
+ "dataset": "141_news_class_Entertainment",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 3641
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7993181597635581,
18
+ "test_acc": 0.7993220338983051,
19
+ "test_auc": 0.8917165833802165,
20
+ "val_auc": 0.9122383252818036,
21
+ "k": 16,
22
+ "dataset": "141_news_class_Entertainment",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 3641,
28
+ 756,
29
+ 1956,
30
+ 1463,
31
+ 15698,
32
+ 21427,
33
+ 778,
34
+ 1290,
35
+ 2040,
36
+ 722,
37
+ 3555,
38
+ 1599,
39
+ 1481,
40
+ 21639,
41
+ 1857,
42
+ 3502
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.6062265483526524,
4
+ "test_acc": 0.6064139941690962,
5
+ "test_auc": 0.6676016591867265,
6
+ "val_auc": 0.6158463385354143,
7
+ "k": 1,
8
+ "dataset": "142_cancer_cat_Thyroid_Cancer",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 171
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6986327744296004,
18
+ "test_acc": 0.6997084548104956,
19
+ "test_auc": 0.7782197742418061,
20
+ "val_auc": 0.7198879551820728,
21
+ "k": 16,
22
+ "dataset": "142_cancer_cat_Thyroid_Cancer",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 171,
28
+ 15063,
29
+ 1358,
30
+ 5,
31
+ 759,
32
+ 1289,
33
+ 140,
34
+ 3989,
35
+ 28492,
36
+ 1702,
37
+ 1314,
38
+ 1145,
39
+ 897,
40
+ 36,
41
+ 1713,
42
+ 2783
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.7134598150284418,
4
+ "test_acc": 0.7288629737609329,
5
+ "test_auc": 0.7350741194070447,
6
+ "val_auc": 0.7745098039215687,
7
+ "k": 1,
8
+ "dataset": "143_cancer_cat_Lung_Cancer",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 140
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8771542994116001,
18
+ "test_acc": 0.8775510204081632,
19
+ "test_auc": 0.9626342989256085,
20
+ "val_auc": 0.969187675070028,
21
+ "k": 16,
22
+ "dataset": "143_cancer_cat_Lung_Cancer",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 140,
28
+ 171,
29
+ 396,
30
+ 1702,
31
+ 1145,
32
+ 885,
33
+ 1425,
34
+ 1358,
35
+ 1494,
36
+ 36,
37
+ 553,
38
+ 759,
39
+ 992,
40
+ 13902,
41
+ 1185,
42
+ 1289
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.648351515797792,
4
+ "test_acc": 0.6559766763848397,
5
+ "test_auc": 0.7077383380932952,
6
+ "val_auc": 0.7066826730692277,
7
+ "k": 1,
8
+ "dataset": "144_cancer_cat_Colon_Cancer",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 396
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8051245554317462,
18
+ "test_acc": 0.8075801749271136,
19
+ "test_auc": 0.8610091119271046,
20
+ "val_auc": 0.8407362945178072,
21
+ "k": 16,
22
+ "dataset": "144_cancer_cat_Colon_Cancer",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 396,
28
+ 696,
29
+ 243,
30
+ 3869,
31
+ 140,
32
+ 1945,
33
+ 1902,
34
+ 60,
35
+ 1494,
36
+ 11897,
37
+ 1145,
38
+ 121,
39
+ 26004,
40
+ 1425,
41
+ 244,
42
+ 226
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.3389600705036947,
4
+ "test_acc": 0.5050505050505051,
5
+ "test_auc": 0.5,
6
+ "val_auc": 0.5,
7
+ "k": 1,
8
+ "dataset": "145_disease_class_digestive system diseases",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1353
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5278281150374173,
18
+ "test_acc": 0.5353535353535354,
19
+ "test_auc": 0.5391836734693878,
20
+ "val_auc": 0.6101010101010101,
21
+ "k": 16,
22
+ "dataset": "145_disease_class_digestive system diseases",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1353,
28
+ 23646,
29
+ 83,
30
+ 1297,
31
+ 1242,
32
+ 1004,
33
+ 2790,
34
+ 627,
35
+ 1061,
36
+ 32026,
37
+ 7381,
38
+ 9514,
39
+ 6724,
40
+ 917,
41
+ 2558,
42
+ 13736
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5085399449035812,
4
+ "test_acc": 0.5151515151515151,
5
+ "test_auc": 0.5187755102040816,
6
+ "val_auc": 0.5107070707070708,
7
+ "k": 1,
8
+ "dataset": "146_disease_class_cardiovascular diseases",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 39
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.4942291209860184,
18
+ "test_acc": 0.5050505050505051,
19
+ "test_auc": 0.5355102040816326,
20
+ "val_auc": 0.585050505050505,
21
+ "k": 16,
22
+ "dataset": "146_disease_class_cardiovascular diseases",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 39,
28
+ 443,
29
+ 159,
30
+ 1533,
31
+ 875,
32
+ 304,
33
+ 7359,
34
+ 83,
35
+ 32026,
36
+ 9514,
37
+ 627,
38
+ 1169,
39
+ 2790,
40
+ 29507,
41
+ 733,
42
+ 6265
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5242821676137976,
4
+ "test_acc": 0.5252525252525253,
5
+ "test_auc": 0.5057142857142857,
6
+ "val_auc": 0.5177777777777778,
7
+ "k": 1,
8
+ "dataset": "147_disease_class_nervous system diseases",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 139
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5546462993271503,
18
+ "test_acc": 0.5555555555555556,
19
+ "test_auc": 0.593265306122449,
20
+ "val_auc": 0.6008080808080808,
21
+ "k": 16,
22
+ "dataset": "147_disease_class_nervous system diseases",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 139,
28
+ 30006,
29
+ 2790,
30
+ 1297,
31
+ 12987,
32
+ 4552,
33
+ 23646,
34
+ 7381,
35
+ 3835,
36
+ 1533,
37
+ 1176,
38
+ 778,
39
+ 28513,
40
+ 12503,
41
+ 1242,
42
+ 1473
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.551934283208893,
4
+ "test_acc": 0.5667796610169491,
5
+ "test_auc": 0.5711372921056211,
6
+ "val_auc": 0.517914653784219,
7
+ "k": 1,
8
+ "dataset": "148_twt_emotion_worry",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 971
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5843101007930762,
18
+ "test_acc": 0.584406779661017,
19
+ "test_auc": 0.6255501869808386,
20
+ "val_auc": 0.6177536231884058,
21
+ "k": 16,
22
+ "dataset": "148_twt_emotion_worry",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 971,
28
+ 2039,
29
+ 500,
30
+ 491,
31
+ 1950,
32
+ 1805,
33
+ 1413,
34
+ 1395,
35
+ 650,
36
+ 1490,
37
+ 4493,
38
+ 2030,
39
+ 875,
40
+ 24,
41
+ 1562,
42
+ 749
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5952036293924372,
4
+ "test_acc": 0.5952542372881356,
5
+ "test_auc": 0.6271616786724177,
6
+ "val_auc": 0.6139291465378423,
7
+ "k": 1,
8
+ "dataset": "149_twt_emotion_happiness",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 1395
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6718953810897634,
18
+ "test_acc": 0.672542372881356,
19
+ "test_auc": 0.7339549480976493,
20
+ "val_auc": 0.7751610305958132,
21
+ "k": 16,
22
+ "dataset": "149_twt_emotion_happiness",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 1395,
28
+ 971,
29
+ 1950,
30
+ 875,
31
+ 1805,
32
+ 650,
33
+ 1725,
34
+ 1029,
35
+ 396,
36
+ 491,
37
+ 2013,
38
+ 568,
39
+ 2007,
40
+ 970,
41
+ 7277,
42
+ 681
43
+ ]
44
+ }
45
+ ]
baseline/gemma-2-2b-btk-mat/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-94.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5858024512029095,
4
+ "test_acc": 0.5993220338983051,
5
+ "test_auc": 0.6051826970101452,
6
+ "val_auc": 0.6231884057971014,
7
+ "k": 1,
8
+ "dataset": "150_twt_emotion_sadness",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 971
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6015544570427772,
18
+ "test_acc": 0.6027118644067797,
19
+ "test_auc": 0.6504855618434068,
20
+ "val_auc": 0.7258454106280193,
21
+ "k": 16,
22
+ "dataset": "150_twt_emotion_sadness",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 971,
28
+ 2242,
29
+ 19795,
30
+ 7277,
31
+ 2000,
32
+ 1725,
33
+ 875,
34
+ 6260,
35
+ 2013,
36
+ 1040,
37
+ 1533,
38
+ 491,
39
+ 1475,
40
+ 925,
41
+ 300,
42
+ 2007
43
+ ]
44
+ }
45
+ ]