chanind commited on
Commit
bd2e542
·
verified ·
1 Parent(s): bcd5e2e

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 +1 -0
  2. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/btk/cfg.json +1 -0
  3. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/btk/sae_weights.safetensors +3 -0
  4. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/cfg.json +1 -0
  5. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/dashboard.zip +3 -0
  6. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/runner_cfg.json +1 -0
  7. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json +45 -0
  8. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json +45 -0
  9. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json +45 -0
  10. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json +45 -0
  11. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json +45 -0
  12. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json +45 -0
  13. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json +45 -0
  14. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json +45 -0
  15. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json +45 -0
  16. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json +45 -0
  17. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json +45 -0
  18. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json +45 -0
  19. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json +45 -0
  20. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json +45 -0
  21. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json +45 -0
  22. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json +45 -0
  23. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json +45 -0
  24. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json +45 -0
  25. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json +45 -0
  26. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json +45 -0
  27. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json +45 -0
  28. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json +45 -0
  29. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json +45 -0
  30. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json +45 -0
  31. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json +45 -0
  32. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json +45 -0
  33. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json +45 -0
  34. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json +45 -0
  35. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json +45 -0
  36. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json +45 -0
  37. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json +45 -0
  38. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json +45 -0
  39. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json +45 -0
  40. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json +45 -0
  41. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json +45 -0
  42. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json +45 -0
  43. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json +45 -0
  44. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json +45 -0
  45. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json +45 -0
  46. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json +45 -0
  47. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json +45 -0
  48. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json +45 -0
  49. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json +45 -0
  50. gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json +45 -0
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/btk/cfg.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"d_in": 2304, "metadata": {"sae_lens_version": "6.39.0", "sae_lens_training_version": "6.39.0", "model_name": "google/gemma-2-2b", "hook_name": "model.layers.12", "hook_head_index": null, "context_size": 1024, "prepend_bos": true, "seqpos_slice": [null], "model_from_pretrained_kwargs": {}, "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}, "d_sae": 32768, "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "none", "reshape_activations": "none", "dtype": "float32", "architecture": "jumprelu"}
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/btk/sae_weights.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b6325c51465e114a4ccfa887dcad6d34570939b7ef13c50219000bf13cc560d2
3
+ size 604251536
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/cfg.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"d_in": 2304, "metadata": {"sae_lens_version": "6.39.0", "sae_lens_training_version": "6.39.0", "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}, "d_sae": 32768, "device": "cuda", "apply_b_dec_to_input": true, "normalize_activations": "none", "reshape_activations": "none", "dtype": "float32", "architecture": "jumprelu"}
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/dashboard.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4e8aee97ea8dd08894edbffc87c0582fb206b2d4e62fb19604449635c69fd503
3
+ size 6945976548
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": "6.37.4", "sae_lens_training_version": "6.37.4"}, "decoder_init_norm": 0.1, "k": 150, "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_frequency_sorted_matryoshka": false, "firing_frequency_ema_decay": 0.99, "use_floating_decoder": false, "floating_decoder_controls_b_dec": true, "use_matryoshka_aux_loss": true, "initial_k": null, "transition_k_duration_steps": null, "transition_k_start_step": 0, "btk_start_step": 61035, "weight_link_enc_min_cos_sim": 0.7, "weight_link_dec_min_cos_sim": 0.7, "weight_link_enc_max_norm_ratio": 1.1, "weight_link_dec_max_norm_ratio": 1.1, "weight_link_direction_coeff": 1.0, "weight_link_norm_coeff": 1.0, "weight_link_coeff": 1.0, "btk_loss_coeff": 1.0, "btk_k_ratio": 1.0, "architecture": "twin_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, "use_chat_formatting": false, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "from_pretrained_path": null, "n_batches_in_buffer": 64, "training_tokens": 500000000, "store_batch_size_prompts": 12, "seqpos_slice": [null], "disable_concat_sequences": false, "sequence_separator_token": "bos", "activations_mixing_fraction": 1.0, "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": 24414, "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": "twin-btk-mat-k-150-layer-12-2026-04-13T14:09:48", "wandb_entity": "chanind", "wandb_log_frequency": 10, "eval_every_n_wandb_logs": 100}, "n_checkpoints": 0, "checkpoint_path": "/vol/data/shared/checkpoints/1bdd9c609f4c734ebf649edf8d9d13c0bd3134b7d42bd968669bd05716199b8e", "save_final_checkpoint": false, "output_path": "/vol/data/saes/twin/gemma-2-2b-twin-btk-mat/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-143.5", "resume_from_checkpoint": null, "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "6.37.4", "sae_lens_training_version": "6.37.4", "exclude_special_tokens": true, "n_batches_for_norm_estimate": 1000, "b_dec_init_method": "zeros", "vnorm_denoise_config": null}
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.9794742107838612,
4
+ "test_acc": 0.9794749403341289,
5
+ "test_auc": 0.9952809554014742,
6
+ "val_auc": 0.9824109824109826,
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
+ 591
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9937947494033413,
18
+ "test_acc": 0.9937947494033413,
19
+ "test_auc": 0.9997165656874968,
20
+ "val_auc": 0.9982839982839983,
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
+ 591,
28
+ 1912,
29
+ 1873,
30
+ 1727,
31
+ 18788,
32
+ 17,
33
+ 12301,
34
+ 875,
35
+ 1809,
36
+ 7145,
37
+ 7768,
38
+ 4169,
39
+ 206,
40
+ 671,
41
+ 2511,
42
+ 1439
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.9371352919954474,
4
+ "test_acc": 0.9371354504212573,
5
+ "test_auc": 0.9632147873362769,
6
+ "val_auc": 0.9470855725506409,
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
+ 961
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.965648298646884,
18
+ "test_acc": 0.9656513285806869,
19
+ "test_auc": 0.9963407995806536,
20
+ "val_auc": 0.9979330301777595,
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
+ 961,
28
+ 765,
29
+ 1912,
30
+ 1873,
31
+ 5667,
32
+ 915,
33
+ 1854,
34
+ 65,
35
+ 1147,
36
+ 1471,
37
+ 6221,
38
+ 2090,
39
+ 1146,
40
+ 5081,
41
+ 29991,
42
+ 1831
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.5801399806457852,
4
+ "test_acc": 0.5849056603773585,
5
+ "test_auc": 0.5955164808680975,
6
+ "val_auc": 0.5542216886754702,
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
+ 473
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6099633060242315,
18
+ "test_acc": 0.6125786163522012,
19
+ "test_auc": 0.6489218400488488,
20
+ "val_auc": 0.6390556222488996,
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
+ 473,
28
+ 4736,
29
+ 2680,
30
+ 197,
31
+ 1863,
32
+ 7180,
33
+ 1115,
34
+ 1954,
35
+ 6258,
36
+ 1446,
37
+ 828,
38
+ 1534,
39
+ 1834,
40
+ 1520,
41
+ 976,
42
+ 721
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.6455622465734614,
4
+ "test_acc": 0.6616352201257861,
5
+ "test_auc": 0.6706750315683734,
6
+ "val_auc": 0.6894757903161265,
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
+ 453
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7225637496552629,
18
+ "test_acc": 0.7240251572327044,
19
+ "test_auc": 0.7802056425113337,
20
+ "val_auc": 0.7995198079231693,
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
+ 453,
28
+ 7715,
29
+ 409,
30
+ 1346,
31
+ 7468,
32
+ 1956,
33
+ 3544,
34
+ 8954,
35
+ 30258,
36
+ 866,
37
+ 319,
38
+ 1186,
39
+ 1986,
40
+ 473,
41
+ 6488,
42
+ 1440
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.7854055454944832,
4
+ "test_acc": 0.7854088050314465,
5
+ "test_auc": 0.8212894629984235,
6
+ "val_auc": 0.852140856342537,
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
+ 1912
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7891522853666044,
18
+ "test_acc": 0.7891823899371069,
19
+ "test_auc": 0.8958062668917379,
20
+ "val_auc": 0.9181672669067628,
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
+ 1912,
28
+ 1957,
29
+ 59,
30
+ 866,
31
+ 641,
32
+ 1146,
33
+ 306,
34
+ 1007,
35
+ 22764,
36
+ 915,
37
+ 1405,
38
+ 1658,
39
+ 831,
40
+ 7028,
41
+ 942,
42
+ 6044
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8507001394550531,
4
+ "test_acc": 0.8508176100628931,
5
+ "test_auc": 0.8986679260262127,
6
+ "val_auc": 0.9447779111644659,
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
+ 1460
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8716530308731567,
18
+ "test_acc": 0.8716981132075472,
19
+ "test_auc": 0.9356869703373741,
20
+ "val_auc": 0.957983193277311,
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
+ 1460,
28
+ 29254,
29
+ 1954,
30
+ 2407,
31
+ 1390,
32
+ 870,
33
+ 1801,
34
+ 1115,
35
+ 4108,
36
+ 2832,
37
+ 6941,
38
+ 575,
39
+ 1147,
40
+ 1540,
41
+ 4711,
42
+ 7014
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.5421815750405149,
4
+ "test_acc": 0.5466666666666666,
5
+ "test_auc": 0.551883815262396,
6
+ "val_auc": 0.6560587515299877,
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
+ 502
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5909808502528386,
18
+ "test_acc": 0.5917948717948718,
19
+ "test_auc": 0.6476819436496449,
20
+ "val_auc": 0.7099143206854345,
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
+ 502,
28
+ 3072,
29
+ 694,
30
+ 7591,
31
+ 928,
32
+ 2016,
33
+ 1694,
34
+ 1939,
35
+ 738,
36
+ 218,
37
+ 5467,
38
+ 5136,
39
+ 1719,
40
+ 1623,
41
+ 19604,
42
+ 260
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.47928499606596,
4
+ "test_acc": 0.5158974358974359,
5
+ "test_auc": 0.5161683273302589,
6
+ "val_auc": 0.5320277437780498,
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
+ 5838
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5248522938533454,
18
+ "test_acc": 0.5251282051282051,
19
+ "test_auc": 0.5273294173090517,
20
+ "val_auc": 0.6001631986944104,
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
+ 5838,
28
+ 216,
29
+ 1626,
30
+ 1728,
31
+ 218,
32
+ 2016,
33
+ 3064,
34
+ 2799,
35
+ 849,
36
+ 126,
37
+ 2514,
38
+ 65,
39
+ 30082,
40
+ 1309,
41
+ 260,
42
+ 355
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.4928310090563661,
4
+ "test_acc": 0.5312820512820513,
5
+ "test_auc": 0.5312931295654223,
6
+ "val_auc": 0.5430436556507547,
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
+ 1493
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5471738419200378,
18
+ "test_acc": 0.5517948717948717,
19
+ "test_auc": 0.5607306358770661,
20
+ "val_auc": 0.5915952672378622,
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
+ 1493,
28
+ 2514,
29
+ 1787,
30
+ 30288,
31
+ 23429,
32
+ 12645,
33
+ 1576,
34
+ 1963,
35
+ 3205,
36
+ 248,
37
+ 755,
38
+ 4375,
39
+ 1498,
40
+ 30358,
41
+ 1152,
42
+ 65
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8233894735728156,
4
+ "test_acc": 0.8282828282828283,
5
+ "test_auc": 0.8300000000000001,
6
+ "val_auc": 0.8183092948717948,
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
+ 26915
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8181447124304266,
18
+ "test_acc": 0.8181818181818182,
19
+ "test_auc": 0.9191836734693878,
20
+ "val_auc": 0.9579326923076923,
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
+ 26915,
28
+ 390,
29
+ 6811,
30
+ 3518,
31
+ 1785,
32
+ 746,
33
+ 7052,
34
+ 1694,
35
+ 1284,
36
+ 1582,
37
+ 6732,
38
+ 2889,
39
+ 3586,
40
+ 3776,
41
+ 7777,
42
+ 6527
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.888410991636798,
4
+ "test_acc": 0.8888888888888888,
5
+ "test_auc": 0.8965306122448978,
6
+ "val_auc": 0.9397035256410257,
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
+ 4032
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8785398998164956,
18
+ "test_acc": 0.8787878787878788,
19
+ "test_auc": 0.9583673469387755,
20
+ "val_auc": 0.9899839743589743,
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
+ 4032,
28
+ 746,
29
+ 7038,
30
+ 1131,
31
+ 126,
32
+ 4223,
33
+ 6732,
34
+ 7650,
35
+ 2757,
36
+ 1326,
37
+ 6525,
38
+ 538,
39
+ 1727,
40
+ 1429,
41
+ 7777,
42
+ 841
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8473644735326978,
4
+ "test_acc": 0.8484848484848485,
5
+ "test_auc": 0.8759183673469387,
6
+ "val_auc": 0.8052884615384616,
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
+ 8171
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8989280055456527,
18
+ "test_acc": 0.898989898989899,
19
+ "test_auc": 0.9775510204081633,
20
+ "val_auc": 0.9543269230769231,
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
+ 8171,
28
+ 7038,
29
+ 746,
30
+ 557,
31
+ 5467,
32
+ 7052,
33
+ 6732,
34
+ 4223,
35
+ 1823,
36
+ 288,
37
+ 6932,
38
+ 7777,
39
+ 1585,
40
+ 178,
41
+ 5673,
42
+ 850
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.6291773460140262,
4
+ "test_acc": 0.6605981794538361,
5
+ "test_auc": 0.6660862101878756,
6
+ "val_auc": 0.7328,
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
+ 4032
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7861743175870144,
18
+ "test_acc": 0.7863025574338969,
19
+ "test_auc": 0.8812839987914882,
20
+ "val_auc": 0.894,
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
+ 4032,
28
+ 8171,
29
+ 7038,
30
+ 746,
31
+ 5673,
32
+ 126,
33
+ 849,
34
+ 1582,
35
+ 5436,
36
+ 7052,
37
+ 18,
38
+ 373,
39
+ 441,
40
+ 288,
41
+ 425,
42
+ 1694
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.6029657527953877,
4
+ "test_acc": 0.6098829648894668,
5
+ "test_auc": 0.6332497095212399,
6
+ "val_auc": 0.6344000000000001,
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
+ 5436
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8203616024969126,
18
+ "test_acc": 0.8214130905938448,
19
+ "test_auc": 0.9213369989523224,
20
+ "val_auc": 0.9683999999999999,
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
+ 5436,
28
+ 4032,
29
+ 7052,
30
+ 7777,
31
+ 8171,
32
+ 2799,
33
+ 6732,
34
+ 5673,
35
+ 4223,
36
+ 7970,
37
+ 1616,
38
+ 1131,
39
+ 6097,
40
+ 557,
41
+ 488,
42
+ 649
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8317040063586488,
4
+ "test_acc": 0.8339835283918509,
5
+ "test_auc": 0.8464190319579247,
6
+ "val_auc": 0.8568,
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
+ 4032
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8269683091823924,
18
+ "test_acc": 0.8300823580407456,
19
+ "test_auc": 0.9195580514098554,
20
+ "val_auc": 0.9481999999999999,
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
+ 4032,
28
+ 746,
29
+ 7038,
30
+ 7777,
31
+ 126,
32
+ 1429,
33
+ 4223,
34
+ 1582,
35
+ 832,
36
+ 7591,
37
+ 5353,
38
+ 2757,
39
+ 7052,
40
+ 23169,
41
+ 5673,
42
+ 3776
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8758700150817185,
4
+ "test_acc": 0.8758974358974358,
5
+ "test_auc": 0.9490271653145723,
6
+ "val_auc": 0.9640962872297021,
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
+ 3561
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9620141214967085,
18
+ "test_acc": 0.9620512820512821,
19
+ "test_auc": 0.9911552832665702,
20
+ "val_auc": 0.9930640554875562,
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
+ 3561,
28
+ 3350,
29
+ 4099,
30
+ 5467,
31
+ 1293,
32
+ 288,
33
+ 26439,
34
+ 124,
35
+ 1585,
36
+ 6932,
37
+ 1582,
38
+ 1785,
39
+ 12764,
40
+ 57,
41
+ 14471,
42
+ 128
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.7585132995164583,
4
+ "test_acc": 0.76,
5
+ "test_auc": 0.8036405560978894,
6
+ "val_auc": 0.864953080375357,
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
+ 1238
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9558948377016445,
18
+ "test_acc": 0.9558974358974359,
19
+ "test_auc": 0.9910837513044064,
20
+ "val_auc": 1.0,
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
+ 1238,
28
+ 7161,
29
+ 22214,
30
+ 1293,
31
+ 10346,
32
+ 2846,
33
+ 1018,
34
+ 6763,
35
+ 841,
36
+ 1827,
37
+ 806,
38
+ 6894,
39
+ 4561,
40
+ 850,
41
+ 5467,
42
+ 12764
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8892027972027973,
4
+ "test_acc": 0.8892307692307693,
5
+ "test_auc": 0.9603670851987748,
6
+ "val_auc": 0.9661362709098327,
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
+ 1238
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9815384226975771,
18
+ "test_acc": 0.9815384615384616,
19
+ "test_auc": 0.9982874406705491,
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
+ 1238,
28
+ 5855,
29
+ 23310,
30
+ 128,
31
+ 7760,
32
+ 726,
33
+ 806,
34
+ 52,
35
+ 441,
36
+ 896,
37
+ 218,
38
+ 1652,
39
+ 22214,
40
+ 1168,
41
+ 3561,
42
+ 18157
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.5810247630045569,
4
+ "test_acc": 0.6025140875596012,
5
+ "test_auc": 0.613088304039947,
6
+ "val_auc": 0.6372,
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
+ 5925
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7011243941203538,
18
+ "test_acc": 0.7013437364542696,
19
+ "test_auc": 0.7737858138140126,
20
+ "val_auc": 0.824,
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
+ 5925,
28
+ 765,
29
+ 1627,
30
+ 1471,
31
+ 965,
32
+ 976,
33
+ 8182,
34
+ 1245,
35
+ 1974,
36
+ 850,
37
+ 9215,
38
+ 13167,
39
+ 6492,
40
+ 1956,
41
+ 750,
42
+ 455
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.6203802143288761,
4
+ "test_acc": 0.6402254009536195,
5
+ "test_auc": 0.6498235332137849,
6
+ "val_auc": 0.7092,
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
+ 1245
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7596734443124059,
18
+ "test_acc": 0.7598612917208496,
19
+ "test_auc": 0.8486049503893843,
20
+ "val_auc": 0.8296,
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
+ 1245,
28
+ 1301,
29
+ 8182,
30
+ 409,
31
+ 31360,
32
+ 1391,
33
+ 9215,
34
+ 755,
35
+ 31605,
36
+ 1627,
37
+ 1956,
38
+ 1253,
39
+ 3544,
40
+ 5925,
41
+ 726,
42
+ 759
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.580628661854354,
4
+ "test_acc": 0.6016471608149111,
5
+ "test_auc": 0.612021837389013,
6
+ "val_auc": 0.6571999999999999,
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
+ 1293
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6699944893321388,
18
+ "test_acc": 0.6701343736454269,
19
+ "test_auc": 0.7443884614170554,
20
+ "val_auc": 0.8204,
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
+ 1293,
28
+ 806,
29
+ 976,
30
+ 338,
31
+ 1301,
32
+ 841,
33
+ 828,
34
+ 128,
35
+ 1245,
36
+ 409,
37
+ 1939,
38
+ 1133,
39
+ 31605,
40
+ 8182,
41
+ 1186,
42
+ 6777
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.5763080389618984,
4
+ "test_acc": 0.5884861407249466,
5
+ "test_auc": 0.6015093653391526,
6
+ "val_auc": 0.6288,
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
+ 1752
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7269895228760173,
18
+ "test_acc": 0.7270788912579957,
19
+ "test_auc": 0.8215675577377706,
20
+ "val_auc": 0.8056,
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
+ 1752,
28
+ 66,
29
+ 1702,
30
+ 21988,
31
+ 3907,
32
+ 1529,
33
+ 298,
34
+ 1763,
35
+ 1236,
36
+ 1782,
37
+ 1900,
38
+ 7729,
39
+ 1416,
40
+ 5967,
41
+ 892,
42
+ 416
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.9393815708101421,
4
+ "test_acc": 0.9393939393939394,
5
+ "test_auc": 0.9648979591836735,
6
+ "val_auc": 0.9665861513687601,
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
+ 803
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 1.0,
18
+ "test_acc": 1.0,
19
+ "test_auc": 1.0,
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
+ 803,
28
+ 881,
29
+ 26786,
30
+ 1699,
31
+ 1143,
32
+ 12841,
33
+ 743,
34
+ 948,
35
+ 412,
36
+ 5740,
37
+ 2034,
38
+ 774,
39
+ 1779,
40
+ 7295,
41
+ 1043,
42
+ 2030
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8382848088730442,
4
+ "test_acc": 0.8383838383838383,
5
+ "test_auc": 0.8828571428571429,
6
+ "val_auc": 0.8667471819645732,
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
+ 18126
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9494846391855681,
18
+ "test_acc": 0.9494949494949495,
19
+ "test_auc": 0.9820408163265306,
20
+ "val_auc": 0.9661835748792271,
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
+ 18126,
28
+ 1546,
29
+ 1779,
30
+ 12841,
31
+ 907,
32
+ 1389,
33
+ 3926,
34
+ 264,
35
+ 1043,
36
+ 2598,
37
+ 4676,
38
+ 1699,
39
+ 1412,
40
+ 732,
41
+ 803,
42
+ 728
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.9290757032692516,
4
+ "test_acc": 0.9292929292929293,
5
+ "test_auc": 0.9485714285714285,
6
+ "val_auc": 0.9690016103059581,
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
+ 1389
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 1.0,
18
+ "test_acc": 1.0,
19
+ "test_auc": 1.0,
20
+ "val_auc": 1.0,
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
+ 1389,
28
+ 3619,
29
+ 601,
30
+ 673,
31
+ 1404,
32
+ 18725,
33
+ 1412,
34
+ 19,
35
+ 59,
36
+ 2278,
37
+ 12841,
38
+ 1108,
39
+ 5858,
40
+ 112,
41
+ 1043,
42
+ 2393
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.9288860795055486,
4
+ "test_acc": 0.9288888888888889,
5
+ "test_auc": 0.9521965865992414,
6
+ "val_auc": 0.9354707792207794,
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
+ 1896
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9377531840140536,
18
+ "test_acc": 0.9377777777777778,
19
+ "test_auc": 0.9766908975979771,
20
+ "val_auc": 0.9910714285714286,
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
+ 1896,
28
+ 537,
29
+ 18725,
30
+ 16688,
31
+ 6065,
32
+ 2036,
33
+ 1089,
34
+ 975,
35
+ 2574,
36
+ 1816,
37
+ 995,
38
+ 3926,
39
+ 1136,
40
+ 1520,
41
+ 14471,
42
+ 110
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8311044388256777,
4
+ "test_acc": 0.8311111111111111,
5
+ "test_auc": 0.8823878002528445,
6
+ "val_auc": 0.8530844155844155,
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
+ 1536
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8710958341838806,
18
+ "test_acc": 0.8711111111111111,
19
+ "test_auc": 0.939198798988622,
20
+ "val_auc": 0.9068587662337662,
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
+ 1536,
28
+ 950,
29
+ 1286,
30
+ 1164,
31
+ 1187,
32
+ 1907,
33
+ 609,
34
+ 3926,
35
+ 948,
36
+ 1343,
37
+ 7689,
38
+ 110,
39
+ 29973,
40
+ 19935,
41
+ 5137,
42
+ 1854
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.9244444444444444,
4
+ "test_acc": 0.9244444444444444,
5
+ "test_auc": 0.9511694058154236,
6
+ "val_auc": 0.9389204545454546,
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
+ 1389
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.9733333333333334,
18
+ "test_acc": 0.9733333333333334,
19
+ "test_auc": 0.9822218710493045,
20
+ "val_auc": 0.9805194805194805,
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
+ 1389,
28
+ 1043,
29
+ 1568,
30
+ 2278,
31
+ 12841,
32
+ 1779,
33
+ 296,
34
+ 288,
35
+ 1699,
36
+ 1045,
37
+ 918,
38
+ 743,
39
+ 1536,
40
+ 1896,
41
+ 7776,
42
+ 1452
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.6931515798171085,
4
+ "test_acc": 0.693974859124404,
5
+ "test_auc": 0.7382805160526154,
6
+ "val_auc": 0.8124,
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
+ 17441
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7576863748010236,
18
+ "test_acc": 0.7589943649761595,
19
+ "test_auc": 0.8276367429702637,
20
+ "val_auc": 0.8808,
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
+ 17441,
28
+ 6,
29
+ 1149,
30
+ 2993,
31
+ 3478,
32
+ 6259,
33
+ 5197,
34
+ 484,
35
+ 23968,
36
+ 870,
37
+ 1673,
38
+ 1573,
39
+ 3798,
40
+ 772,
41
+ 3733,
42
+ 7742
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.5918740790861167,
4
+ "test_acc": 0.5925444299956654,
5
+ "test_auc": 0.6224384883981355,
6
+ "val_auc": 0.6424,
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
+ 831
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6823735884160967,
18
+ "test_acc": 0.682704811443433,
19
+ "test_auc": 0.7453504609330494,
20
+ "val_auc": 0.8089999999999999,
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
+ 831,
28
+ 6,
29
+ 17441,
30
+ 870,
31
+ 1942,
32
+ 9,
33
+ 3478,
34
+ 1149,
35
+ 5631,
36
+ 5475,
37
+ 496,
38
+ 3358,
39
+ 2993,
40
+ 641,
41
+ 1573,
42
+ 772
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.6031938982411725,
4
+ "test_acc": 0.6168183788469874,
5
+ "test_auc": 0.6442134977550839,
6
+ "val_auc": 0.6803999999999999,
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
+ 870
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.7530309590465332,
18
+ "test_acc": 0.7542262678803641,
19
+ "test_auc": 0.8377493871010896,
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
+ 870,
28
+ 1149,
29
+ 5631,
30
+ 6,
31
+ 1989,
32
+ 17441,
33
+ 1986,
34
+ 1204,
35
+ 3478,
36
+ 1460,
37
+ 1865,
38
+ 422,
39
+ 484,
40
+ 575,
41
+ 1563,
42
+ 6259
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8053932859952562,
4
+ "test_acc": 0.8067796610169492,
5
+ "test_auc": 0.834229627913647,
6
+ "val_auc": 0.8015297906602254,
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
+ 386
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8764820911771912,
18
+ "test_acc": 0.8766101694915254,
19
+ "test_auc": 0.9417325787911881,
20
+ "val_auc": 0.9448470209339774,
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
+ 386,
28
+ 1293,
29
+ 7619,
30
+ 1605,
31
+ 31430,
32
+ 6186,
33
+ 591,
34
+ 15791,
35
+ 642,
36
+ 351,
37
+ 1147,
38
+ 656,
39
+ 4454,
40
+ 15549,
41
+ 51,
42
+ 6365
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.8301655639119235,
4
+ "test_acc": 0.8325423728813559,
5
+ "test_auc": 0.8414146929800369,
6
+ "val_auc": 0.8047504025764896,
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
+ 642
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8604195503375026,
18
+ "test_acc": 0.8610169491525423,
19
+ "test_auc": 0.9290373704279784,
20
+ "val_auc": 0.8776167471819646,
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
+ 642,
28
+ 234,
29
+ 2803,
30
+ 1756,
31
+ 351,
32
+ 2019,
33
+ 607,
34
+ 20009,
35
+ 179,
36
+ 1605,
37
+ 492,
38
+ 1685,
39
+ 735,
40
+ 386,
41
+ 6365,
42
+ 2054
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.6359536817229112,
4
+ "test_acc": 0.64,
5
+ "test_auc": 0.6656977492434355,
6
+ "val_auc": 0.6843800322061192,
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
+ 4454
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8054156793246283,
18
+ "test_acc": 0.8054237288135593,
19
+ "test_auc": 0.9003770872172765,
20
+ "val_auc": 0.9376006441223833,
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
+ 4454,
28
+ 921,
29
+ 1795,
30
+ 386,
31
+ 591,
32
+ 1065,
33
+ 6962,
34
+ 351,
35
+ 15549,
36
+ 1164,
37
+ 1535,
38
+ 15791,
39
+ 28150,
40
+ 7619,
41
+ 310,
42
+ 462
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.6062019613040022,
4
+ "test_acc": 0.6122448979591837,
5
+ "test_auc": 0.6179620563035496,
6
+ "val_auc": 0.5438175270108043,
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
+ 21381
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6520499049822358,
18
+ "test_acc": 0.6530612244897959,
19
+ "test_auc": 0.7547939616483068,
20
+ "val_auc": 0.7302921168467388,
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
+ 21381,
28
+ 6438,
29
+ 1051,
30
+ 58,
31
+ 1246,
32
+ 1336,
33
+ 1912,
34
+ 945,
35
+ 6526,
36
+ 364,
37
+ 1398,
38
+ 1959,
39
+ 405,
40
+ 883,
41
+ 1440,
42
+ 1707
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.7118675852178012,
4
+ "test_acc": 0.7230320699708455,
5
+ "test_auc": 0.7396980824153406,
6
+ "val_auc": 0.7294917967186876,
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
+ 822
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8539667068820793,
18
+ "test_acc": 0.8542274052478134,
19
+ "test_auc": 0.926016591867265,
20
+ "val_auc": 0.9059623849539816,
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
+ 822,
28
+ 1246,
29
+ 405,
30
+ 6526,
31
+ 21381,
32
+ 307,
33
+ 611,
34
+ 945,
35
+ 1152,
36
+ 1912,
37
+ 1336,
38
+ 565,
39
+ 872,
40
+ 268,
41
+ 792,
42
+ 1051
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.6616882630922509,
4
+ "test_acc": 0.6676384839650146,
5
+ "test_auc": 0.6858595131238949,
6
+ "val_auc": 0.6846738695478191,
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
+ 822
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.8278830114955611,
18
+ "test_acc": 0.8279883381924198,
19
+ "test_auc": 0.8875628994968041,
20
+ "val_auc": 0.912765106042417,
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
+ 822,
28
+ 1152,
29
+ 1090,
30
+ 1707,
31
+ 11181,
32
+ 565,
33
+ 307,
34
+ 611,
35
+ 6526,
36
+ 1246,
37
+ 405,
38
+ 550,
39
+ 4173,
40
+ 65,
41
+ 1750,
42
+ 1756
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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
+ 777
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5246703106692887,
18
+ "test_acc": 0.5252525252525253,
19
+ "test_auc": 0.5612244897959183,
20
+ "val_auc": 0.7002020202020203,
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
+ 777,
28
+ 4978,
29
+ 2258,
30
+ 828,
31
+ 7169,
32
+ 24524,
33
+ 16234,
34
+ 209,
35
+ 1997,
36
+ 1064,
37
+ 3833,
38
+ 158,
39
+ 935,
40
+ 496,
41
+ 5312,
42
+ 6273
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.509923546960584,
4
+ "test_acc": 0.5252525252525253,
5
+ "test_auc": 0.506938775510204,
6
+ "val_auc": 0.5915151515151514,
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
+ 828
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5858585858585859,
18
+ "test_acc": 0.5858585858585859,
19
+ "test_auc": 0.6140816326530611,
20
+ "val_auc": 0.6012121212121211,
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
+ 828,
28
+ 653,
29
+ 759,
30
+ 4711,
31
+ 1567,
32
+ 3833,
33
+ 713,
34
+ 29134,
35
+ 6832,
36
+ 20539,
37
+ 7,
38
+ 1957,
39
+ 16181,
40
+ 6446,
41
+ 12389,
42
+ 286
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.5636261691770523,
4
+ "test_acc": 0.5757575757575758,
5
+ "test_auc": 0.5740816326530612,
6
+ "val_auc": 0.5723232323232323,
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
+ 713
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5541916712129478,
18
+ "test_acc": 0.5555555555555556,
19
+ "test_auc": 0.5677551020408164,
20
+ "val_auc": 0.6056565656565656,
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
+ 713,
28
+ 4821,
29
+ 24524,
30
+ 7677,
31
+ 562,
32
+ 120,
33
+ 24361,
34
+ 7768,
35
+ 24921,
36
+ 3449,
37
+ 65,
38
+ 733,
39
+ 31821,
40
+ 1957,
41
+ 563,
42
+ 1667
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.49158923230309076,
4
+ "test_acc": 0.528135593220339,
5
+ "test_auc": 0.5337778954451688,
6
+ "val_auc": 0.5571658615136876,
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
+ 1714
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5776271186440678,
18
+ "test_acc": 0.5776271186440678,
19
+ "test_auc": 0.6218353906741239,
20
+ "val_auc": 0.6151368760064412,
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
+ 1714,
28
+ 1390,
29
+ 1188,
30
+ 288,
31
+ 1460,
32
+ 1801,
33
+ 286,
34
+ 1410,
35
+ 1661,
36
+ 566,
37
+ 1779,
38
+ 750,
39
+ 870,
40
+ 30258,
41
+ 1586,
42
+ 1245
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.5911363415229978,
4
+ "test_acc": 0.5925423728813559,
5
+ "test_auc": 0.6170349288296141,
6
+ "val_auc": 0.6751207729468599,
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
+ 750
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6687002654009802,
18
+ "test_acc": 0.6691525423728814,
19
+ "test_auc": 0.7358403841840317,
20
+ "val_auc": 0.7661030595813204,
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
+ 750,
28
+ 1460,
29
+ 1661,
30
+ 1390,
31
+ 3731,
32
+ 7587,
33
+ 721,
34
+ 866,
35
+ 870,
36
+ 1801,
37
+ 1325,
38
+ 4934,
39
+ 1124,
40
+ 1188,
41
+ 7456,
42
+ 1809
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/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": 0.5920515426753724,
4
+ "test_acc": 0.6006779661016949,
5
+ "test_auc": 0.6100061407669707,
6
+ "val_auc": 0.6384863123993558,
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
+ 1460
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.6126216842100116,
18
+ "test_acc": 0.6149152542372881,
19
+ "test_auc": 0.6556463800730274,
20
+ "val_auc": 0.7238325281803543,
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
+ 1460,
28
+ 1536,
29
+ 1834,
30
+ 1188,
31
+ 1661,
32
+ 15253,
33
+ 7456,
34
+ 1718,
35
+ 866,
36
+ 28318,
37
+ 1801,
38
+ 870,
39
+ 1822,
40
+ 750,
41
+ 669,
42
+ 1873
43
+ ]
44
+ }
45
+ ]
gemma-2-2b-twin-btk-mat/layer-12/k-150/w-32768/t-500M/enc-0.7-dec-0.7/seed-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "test_f1": 0.5730537842260894,
4
+ "test_acc": 0.5732009925558312,
5
+ "test_auc": 0.587483375203192,
6
+ "val_auc": 0.5438175270108043,
7
+ "k": 1,
8
+ "dataset": "151_it_tick_HR Support",
9
+ "hook_name": "blocks.12.hook_resid_post",
10
+ "reg_type": "l1",
11
+ "binarize": false,
12
+ "indices": [
13
+ 907
14
+ ]
15
+ },
16
+ {
17
+ "test_f1": 0.5975639435006681,
18
+ "test_acc": 0.598014888337469,
19
+ "test_auc": 0.665558346879464,
20
+ "val_auc": 0.6890756302521007,
21
+ "k": 16,
22
+ "dataset": "151_it_tick_HR Support",
23
+ "hook_name": "blocks.12.hook_resid_post",
24
+ "reg_type": "l1",
25
+ "binarize": false,
26
+ "indices": [
27
+ 907,
28
+ 1506,
29
+ 4308,
30
+ 1025,
31
+ 2,
32
+ 28659,
33
+ 765,
34
+ 1279,
35
+ 434,
36
+ 2046,
37
+ 504,
38
+ 7954,
39
+ 1692,
40
+ 1098,
41
+ 32074,
42
+ 4803
43
+ ]
44
+ }
45
+ ]