Upload folder using huggingface_hub
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +286 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/cfg.json +1 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json +101 -0
- matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json +101 -0
.gitattributes
CHANGED
|
@@ -265,3 +265,289 @@ cross-arch-panel/gemma-2-2b/matryoshka/k-50/seed-0/snapshots/step-85449-tokens-3
|
|
| 265 |
cross-arch-panel/gemma-2-2b/matryoshka/k-50/seed-0/snapshots/step-85449-tokens-350000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 266 |
cross-arch-panel/gemma-2-2b/matryoshka/k-50/seed-0/snapshots/step-97656-tokens-400000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 267 |
cross-arch-panel/gemma-2-2b/matryoshka/k-50/seed-0/snapshots/step-97656-tokens-400000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 265 |
cross-arch-panel/gemma-2-2b/matryoshka/k-50/seed-0/snapshots/step-85449-tokens-350000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 266 |
cross-arch-panel/gemma-2-2b/matryoshka/k-50/seed-0/snapshots/step-97656-tokens-400000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 267 |
cross-arch-panel/gemma-2-2b/matryoshka/k-50/seed-0/snapshots/step-97656-tokens-400000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 268 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 269 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 270 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 271 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 272 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 273 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 274 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 275 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 276 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 277 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 278 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 279 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 280 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 281 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 282 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 283 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 284 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 285 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 286 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 287 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 288 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 289 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 290 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 291 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 292 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 293 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 294 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 295 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 296 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 297 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 298 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 299 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 300 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 301 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 302 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 303 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 304 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 305 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 306 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 307 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 308 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 309 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 310 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 311 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-1/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 312 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 313 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 314 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 315 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 316 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 317 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 318 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 319 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 320 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 321 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 322 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 323 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 324 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 325 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 326 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 327 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 328 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 329 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 330 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 331 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 332 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 333 |
+
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-2/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 334 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 335 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 336 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 337 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 338 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 339 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 340 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 341 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 342 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 343 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 344 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 345 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 346 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 347 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 348 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 349 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 350 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 351 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 352 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 353 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 354 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 355 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 356 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 357 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 358 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 359 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 360 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 361 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 362 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 363 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 364 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 365 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 366 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 367 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 368 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 369 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 370 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 371 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 372 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 373 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 374 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 375 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 376 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 377 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-1/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 378 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 379 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 380 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 381 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 382 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 383 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 384 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 385 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 386 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 387 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 388 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 389 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 390 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 391 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 392 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 393 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 394 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 395 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 396 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 397 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 398 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 399 |
+
matryoshka-prefix-panel/gemma-2-2b/n-2/k-100/seed-2/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 400 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 401 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 402 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 403 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 404 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 405 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 406 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 407 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 408 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 409 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 410 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 411 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 412 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 413 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 414 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 415 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 416 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 417 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 418 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 419 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 420 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 421 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 422 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 423 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 424 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 425 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 426 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 427 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 428 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 429 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 430 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 431 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 432 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 433 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 434 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 435 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 436 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 437 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 438 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 439 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 440 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 441 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 442 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 443 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-1/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 444 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 445 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 446 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 447 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 448 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 449 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 450 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 451 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 452 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 453 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 454 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 455 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 456 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 457 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 458 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 459 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 460 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 461 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 462 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 463 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 464 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 465 |
+
matryoshka-prefix-panel/gemma-2-2b/n-3/k-100/seed-2/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 466 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 467 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 468 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 469 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 470 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 471 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 472 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 473 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 474 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 475 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 476 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 477 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 478 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 479 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 480 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 481 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 482 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 483 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 484 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 485 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 486 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 487 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 488 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 489 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 490 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 491 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 492 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 493 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 494 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 495 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 496 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 497 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 498 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 499 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 500 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 501 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 502 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 503 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 504 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 505 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 506 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 507 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 508 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 509 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-1/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 510 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 511 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 512 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 513 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 514 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 515 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 516 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 517 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 518 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 519 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 520 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 521 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 522 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 523 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 524 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 525 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 526 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 527 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 528 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 529 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 530 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 531 |
+
matryoshka-prefix-panel/gemma-2-2b/n-4/k-100/seed-2/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 532 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-1-tokens-0/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 533 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-1-tokens-0/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 534 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 535 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-14648-tokens-60000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 536 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 537 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-21972-tokens-90000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 538 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 539 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-29296-tokens-120000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 540 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 541 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-36621-tokens-150000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 542 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 543 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-43945-tokens-180000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 544 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 545 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-51269-tokens-210000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 546 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 547 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-58593-tokens-240000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 548 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 549 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-65917-tokens-270000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 550 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 551 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-7324-tokens-30000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 552 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 553 |
+
matryoshka-prefix-panel/gemma-2-2b/n-5/k-100/seed-0/snapshots/step-73242-tokens-300000000/saebench_core_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/cfg.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"reshape_activations": "none", "d_sae": 32768, "apply_b_dec_to_input": true, "d_in": 2304, "dtype": "float32", "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}, "device": "cuda", "normalize_activations": "none", "architecture": "jumprelu"}
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9110284125213659,
|
| 4 |
+
"test_acc": 0.911217183770883,
|
| 5 |
+
"test_auc": 0.9605725555385435,
|
| 6 |
+
"val_auc": 0.945087945087945,
|
| 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 |
+
24914
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9114345601516296,
|
| 18 |
+
"test_acc": 0.9116945107398569,
|
| 19 |
+
"test_auc": 0.9623538171584389,
|
| 20 |
+
"val_auc": 0.9455169455169455,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "100_news_fake",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
24914,
|
| 28 |
+
13585
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.922515375813254,
|
| 33 |
+
"test_acc": 0.922673031026253,
|
| 34 |
+
"test_auc": 0.9715034595390684,
|
| 35 |
+
"val_auc": 0.9635349635349635,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "100_news_fake",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
24914,
|
| 43 |
+
13585,
|
| 44 |
+
10281,
|
| 45 |
+
15091,
|
| 46 |
+
25419
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.9646769190085163,
|
| 51 |
+
"test_acc": 0.9646778042959427,
|
| 52 |
+
"test_auc": 0.9957056511880545,
|
| 53 |
+
"val_auc": 0.9806949806949807,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "100_news_fake",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
24914,
|
| 61 |
+
13585,
|
| 62 |
+
10281,
|
| 63 |
+
15091,
|
| 64 |
+
25419,
|
| 65 |
+
1567,
|
| 66 |
+
23,
|
| 67 |
+
12367,
|
| 68 |
+
19662,
|
| 69 |
+
15794
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.9818611205719407,
|
| 74 |
+
"test_acc": 0.9818615751789976,
|
| 75 |
+
"test_auc": 0.9975812390180596,
|
| 76 |
+
"val_auc": 0.9867009867009866,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "100_news_fake",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
24914,
|
| 84 |
+
13585,
|
| 85 |
+
10281,
|
| 86 |
+
15091,
|
| 87 |
+
25419,
|
| 88 |
+
1567,
|
| 89 |
+
23,
|
| 90 |
+
12367,
|
| 91 |
+
19662,
|
| 92 |
+
15794,
|
| 93 |
+
27566,
|
| 94 |
+
31202,
|
| 95 |
+
27330,
|
| 96 |
+
32387,
|
| 97 |
+
26571,
|
| 98 |
+
28827
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.939694064473835,
|
| 4 |
+
"test_acc": 0.9397278029812054,
|
| 5 |
+
"test_auc": 0.9792200090051948,
|
| 6 |
+
"val_auc": 0.972302604381976,
|
| 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 |
+
27788
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9494306464619561,
|
| 18 |
+
"test_acc": 0.949449125081011,
|
| 19 |
+
"test_auc": 0.9855849680449991,
|
| 20 |
+
"val_auc": 0.9772633319553534,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "105_click_bait",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
27788,
|
| 28 |
+
29716
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.9649997122807324,
|
| 33 |
+
"test_acc": 0.9650032404407,
|
| 34 |
+
"test_auc": 0.9908183974785455,
|
| 35 |
+
"val_auc": 0.9822240595287308,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "105_click_bait",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
27788,
|
| 43 |
+
29716,
|
| 44 |
+
6051,
|
| 45 |
+
8166,
|
| 46 |
+
10201
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.9643500313555798,
|
| 51 |
+
"test_acc": 0.9643551523007129,
|
| 52 |
+
"test_auc": 0.9932158625834155,
|
| 53 |
+
"val_auc": 0.9822240595287308,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "105_click_bait",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
27788,
|
| 61 |
+
29716,
|
| 62 |
+
6051,
|
| 63 |
+
8166,
|
| 64 |
+
10201,
|
| 65 |
+
22056,
|
| 66 |
+
11357,
|
| 67 |
+
7523,
|
| 68 |
+
29215,
|
| 69 |
+
16628
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.9734279175098244,
|
| 74 |
+
"test_acc": 0.9734283862605314,
|
| 75 |
+
"test_auc": 0.9950068210990369,
|
| 76 |
+
"val_auc": 0.99338569656883,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "105_click_bait",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
27788,
|
| 84 |
+
29716,
|
| 85 |
+
6051,
|
| 86 |
+
8166,
|
| 87 |
+
10201,
|
| 88 |
+
22056,
|
| 89 |
+
11357,
|
| 90 |
+
7523,
|
| 91 |
+
29215,
|
| 92 |
+
16628,
|
| 93 |
+
20526,
|
| 94 |
+
21509,
|
| 95 |
+
11558,
|
| 96 |
+
11496,
|
| 97 |
+
32472,
|
| 98 |
+
17924
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.553939680542583,
|
| 4 |
+
"test_acc": 0.5559748427672956,
|
| 5 |
+
"test_auc": 0.5757477679362536,
|
| 6 |
+
"val_auc": 0.5488195278111244,
|
| 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 |
+
23181
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5635207259160504,
|
| 18 |
+
"test_acc": 0.5647798742138365,
|
| 19 |
+
"test_auc": 0.5813932411783231,
|
| 20 |
+
"val_auc": 0.5950380152060823,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "106_hate_hate",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
23181,
|
| 28 |
+
15073
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.5731308413154667,
|
| 33 |
+
"test_acc": 0.5745911949685535,
|
| 34 |
+
"test_auc": 0.6043442334935634,
|
| 35 |
+
"val_auc": 0.5922368947579032,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "106_hate_hate",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
23181,
|
| 43 |
+
15073,
|
| 44 |
+
16511,
|
| 45 |
+
27850,
|
| 46 |
+
28778
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.5907402457673732,
|
| 51 |
+
"test_acc": 0.5911949685534591,
|
| 52 |
+
"test_auc": 0.6245031841780426,
|
| 53 |
+
"val_auc": 0.5734293717486996,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "106_hate_hate",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
23181,
|
| 61 |
+
15073,
|
| 62 |
+
16511,
|
| 63 |
+
27850,
|
| 64 |
+
28778,
|
| 65 |
+
17096,
|
| 66 |
+
20344,
|
| 67 |
+
28622,
|
| 68 |
+
24948,
|
| 69 |
+
17229
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.5863356655482709,
|
| 74 |
+
"test_acc": 0.5871698113207547,
|
| 75 |
+
"test_auc": 0.6219273365406328,
|
| 76 |
+
"val_auc": 0.5870348139255702,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "106_hate_hate",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
23181,
|
| 84 |
+
15073,
|
| 85 |
+
16511,
|
| 86 |
+
27850,
|
| 87 |
+
28778,
|
| 88 |
+
17096,
|
| 89 |
+
20344,
|
| 90 |
+
28622,
|
| 91 |
+
24948,
|
| 92 |
+
17229,
|
| 93 |
+
12148,
|
| 94 |
+
23853,
|
| 95 |
+
6491,
|
| 96 |
+
15398,
|
| 97 |
+
13488,
|
| 98 |
+
3591
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6175941492244598,
|
| 4 |
+
"test_acc": 0.620880503144654,
|
| 5 |
+
"test_auc": 0.6578319944832558,
|
| 6 |
+
"val_auc": 0.7202881152460985,
|
| 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 |
+
3614
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6366497640755162,
|
| 18 |
+
"test_acc": 0.6374842767295598,
|
| 19 |
+
"test_auc": 0.6838833706820693,
|
| 20 |
+
"val_auc": 0.7366946778711484,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "107_hate_offensive",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
3614,
|
| 28 |
+
11093
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6782385864339769,
|
| 33 |
+
"test_acc": 0.6782389937106919,
|
| 34 |
+
"test_auc": 0.732057290901929,
|
| 35 |
+
"val_auc": 0.7486994797919166,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "107_hate_offensive",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
3614,
|
| 43 |
+
11093,
|
| 44 |
+
5654,
|
| 45 |
+
32192,
|
| 46 |
+
29204
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.6923217475004986,
|
| 51 |
+
"test_acc": 0.6923270440251572,
|
| 52 |
+
"test_auc": 0.7654239478137067,
|
| 53 |
+
"val_auc": 0.725890356142457,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "107_hate_offensive",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
3614,
|
| 61 |
+
11093,
|
| 62 |
+
5654,
|
| 63 |
+
32192,
|
| 64 |
+
29204,
|
| 65 |
+
30620,
|
| 66 |
+
10020,
|
| 67 |
+
22980,
|
| 68 |
+
6481,
|
| 69 |
+
10422
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7081150291059563,
|
| 74 |
+
"test_acc": 0.7081761006289308,
|
| 75 |
+
"test_auc": 0.7761635236684323,
|
| 76 |
+
"val_auc": 0.7831132452981193,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "107_hate_offensive",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
3614,
|
| 84 |
+
11093,
|
| 85 |
+
5654,
|
| 86 |
+
32192,
|
| 87 |
+
29204,
|
| 88 |
+
30620,
|
| 89 |
+
10020,
|
| 90 |
+
22980,
|
| 91 |
+
6481,
|
| 92 |
+
10422,
|
| 93 |
+
25554,
|
| 94 |
+
9269,
|
| 95 |
+
28854,
|
| 96 |
+
21327,
|
| 97 |
+
1774,
|
| 98 |
+
31429
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7025523023000008,
|
| 4 |
+
"test_acc": 0.7038993710691824,
|
| 5 |
+
"test_auc": 0.7566775843789462,
|
| 6 |
+
"val_auc": 0.7254901960784313,
|
| 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 |
+
17498
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7486302458550385,
|
| 18 |
+
"test_acc": 0.7486792452830189,
|
| 19 |
+
"test_auc": 0.8220662677625897,
|
| 20 |
+
"val_auc": 0.8225290116046418,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "110_aimade_humangpt3",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
17498,
|
| 28 |
+
4508
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7788248109174962,
|
| 33 |
+
"test_acc": 0.7788679245283019,
|
| 34 |
+
"test_auc": 0.8636075385377184,
|
| 35 |
+
"val_auc": 0.9065626250500199,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "110_aimade_humangpt3",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
17498,
|
| 43 |
+
4508,
|
| 44 |
+
13399,
|
| 45 |
+
29972,
|
| 46 |
+
29652
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.8084382106616796,
|
| 51 |
+
"test_acc": 0.8085534591194968,
|
| 52 |
+
"test_auc": 0.8846070889352218,
|
| 53 |
+
"val_auc": 0.9375750300120047,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "110_aimade_humangpt3",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
17498,
|
| 61 |
+
4508,
|
| 62 |
+
13399,
|
| 63 |
+
29972,
|
| 64 |
+
29652,
|
| 65 |
+
24680,
|
| 66 |
+
22445,
|
| 67 |
+
10814,
|
| 68 |
+
20557,
|
| 69 |
+
19071
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.8124465023102964,
|
| 74 |
+
"test_acc": 0.8125786163522013,
|
| 75 |
+
"test_auc": 0.8977483420907935,
|
| 76 |
+
"val_auc": 0.9327731092436975,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "110_aimade_humangpt3",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
17498,
|
| 84 |
+
4508,
|
| 85 |
+
13399,
|
| 86 |
+
29972,
|
| 87 |
+
29652,
|
| 88 |
+
24680,
|
| 89 |
+
22445,
|
| 90 |
+
10814,
|
| 91 |
+
20557,
|
| 92 |
+
19071,
|
| 93 |
+
10728,
|
| 94 |
+
13444,
|
| 95 |
+
32101,
|
| 96 |
+
5131,
|
| 97 |
+
2966,
|
| 98 |
+
17875
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7577094434272407,
|
| 4 |
+
"test_acc": 0.7577358490566037,
|
| 5 |
+
"test_auc": 0.8253785420120117,
|
| 6 |
+
"val_auc": 0.8301320528211285,
|
| 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 |
+
16437
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7722945323304855,
|
| 18 |
+
"test_acc": 0.7723270440251573,
|
| 19 |
+
"test_auc": 0.8481881475060732,
|
| 20 |
+
"val_auc": 0.8599439775910362,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "113_movie_sent",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
16437,
|
| 28 |
+
7577
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.834968532569888,
|
| 33 |
+
"test_acc": 0.8349685534591195,
|
| 34 |
+
"test_auc": 0.9025301785549734,
|
| 35 |
+
"val_auc": 0.9211684673869549,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "113_movie_sent",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
16437,
|
| 43 |
+
7577,
|
| 44 |
+
21805,
|
| 45 |
+
20827,
|
| 46 |
+
13171
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.859359213125981,
|
| 51 |
+
"test_acc": 0.8593710691823899,
|
| 52 |
+
"test_auc": 0.9275904546554616,
|
| 53 |
+
"val_auc": 0.9463785514205683,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "113_movie_sent",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
16437,
|
| 61 |
+
7577,
|
| 62 |
+
21805,
|
| 63 |
+
20827,
|
| 64 |
+
13171,
|
| 65 |
+
10320,
|
| 66 |
+
4199,
|
| 67 |
+
21048,
|
| 68 |
+
4952,
|
| 69 |
+
9323
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.8678897463594573,
|
| 74 |
+
"test_acc": 0.8679245283018868,
|
| 75 |
+
"test_auc": 0.9337884377224595,
|
| 76 |
+
"val_auc": 0.941576630652261,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "113_movie_sent",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
16437,
|
| 84 |
+
7577,
|
| 85 |
+
21805,
|
| 86 |
+
20827,
|
| 87 |
+
13171,
|
| 88 |
+
10320,
|
| 89 |
+
4199,
|
| 90 |
+
21048,
|
| 91 |
+
4952,
|
| 92 |
+
9323,
|
| 93 |
+
4278,
|
| 94 |
+
24428,
|
| 95 |
+
16421,
|
| 96 |
+
17550,
|
| 97 |
+
2171,
|
| 98 |
+
13969
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.567867332241291,
|
| 4 |
+
"test_acc": 0.5702564102564103,
|
| 5 |
+
"test_auc": 0.5662848823509611,
|
| 6 |
+
"val_auc": 0.711546307629539,
|
| 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 |
+
3543
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.555227037031718,
|
| 18 |
+
"test_acc": 0.558974358974359,
|
| 19 |
+
"test_auc": 0.5834020601205102,
|
| 20 |
+
"val_auc": 0.7421460628314973,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "114_nyc_borough_Manhattan",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
3543,
|
| 28 |
+
10681
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.5692176134270717,
|
| 33 |
+
"test_acc": 0.5712820512820512,
|
| 34 |
+
"test_auc": 0.5889184367320833,
|
| 35 |
+
"val_auc": 0.7262341901264789,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "114_nyc_borough_Manhattan",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
3543,
|
| 43 |
+
10681,
|
| 44 |
+
15325,
|
| 45 |
+
21486,
|
| 46 |
+
2463
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.5789974143806087,
|
| 51 |
+
"test_acc": 0.5805128205128205,
|
| 52 |
+
"test_auc": 0.6033047766519676,
|
| 53 |
+
"val_auc": 0.7123623011015912,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "114_nyc_borough_Manhattan",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
3543,
|
| 61 |
+
10681,
|
| 62 |
+
15325,
|
| 63 |
+
21486,
|
| 64 |
+
2463,
|
| 65 |
+
23013,
|
| 66 |
+
9610,
|
| 67 |
+
10412,
|
| 68 |
+
16050,
|
| 69 |
+
29999
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.5872690272690274,
|
| 74 |
+
"test_acc": 0.5897435897435898,
|
| 75 |
+
"test_auc": 0.6152211599959605,
|
| 76 |
+
"val_auc": 0.7258261933904528,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "114_nyc_borough_Manhattan",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
3543,
|
| 84 |
+
10681,
|
| 85 |
+
15325,
|
| 86 |
+
21486,
|
| 87 |
+
2463,
|
| 88 |
+
23013,
|
| 89 |
+
9610,
|
| 90 |
+
10412,
|
| 91 |
+
16050,
|
| 92 |
+
29999,
|
| 93 |
+
20199,
|
| 94 |
+
28004,
|
| 95 |
+
5053,
|
| 96 |
+
10915,
|
| 97 |
+
8067,
|
| 98 |
+
12544
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.4944843304843305,
|
| 4 |
+
"test_acc": 0.49743589743589745,
|
| 5 |
+
"test_auc": 0.5110243376981857,
|
| 6 |
+
"val_auc": 0.5022439820481436,
|
| 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 |
+
7629
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5139042885930222,
|
| 18 |
+
"test_acc": 0.5148717948717949,
|
| 19 |
+
"test_auc": 0.5207968660585047,
|
| 20 |
+
"val_auc": 0.5079559363525092,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "115_nyc_borough_Brooklyn",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
7629,
|
| 28 |
+
15569
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.5309613323101205,
|
| 33 |
+
"test_acc": 0.5312820512820513,
|
| 34 |
+
"test_auc": 0.542771905611472,
|
| 35 |
+
"val_auc": 0.5442676458588331,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "115_nyc_borough_Brooklyn",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
7629,
|
| 43 |
+
15569,
|
| 44 |
+
30195,
|
| 45 |
+
1517,
|
| 46 |
+
5093
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.5300487743588448,
|
| 51 |
+
"test_acc": 0.5302564102564102,
|
| 52 |
+
"test_auc": 0.5365780792405831,
|
| 53 |
+
"val_auc": 0.5760913912688699,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "115_nyc_borough_Brooklyn",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
7629,
|
| 61 |
+
15569,
|
| 62 |
+
30195,
|
| 63 |
+
1517,
|
| 64 |
+
5093,
|
| 65 |
+
346,
|
| 66 |
+
15996,
|
| 67 |
+
8929,
|
| 68 |
+
14132,
|
| 69 |
+
17386
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.513163980730496,
|
| 74 |
+
"test_acc": 0.5138461538461538,
|
| 75 |
+
"test_auc": 0.538496818931565,
|
| 76 |
+
"val_auc": 0.6075071399428805,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "115_nyc_borough_Brooklyn",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
7629,
|
| 84 |
+
15569,
|
| 85 |
+
30195,
|
| 86 |
+
1517,
|
| 87 |
+
5093,
|
| 88 |
+
346,
|
| 89 |
+
15996,
|
| 90 |
+
8929,
|
| 91 |
+
14132,
|
| 92 |
+
17386,
|
| 93 |
+
500,
|
| 94 |
+
29471,
|
| 95 |
+
7376,
|
| 96 |
+
2916,
|
| 97 |
+
9346,
|
| 98 |
+
19448
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5798578558643499,
|
| 4 |
+
"test_acc": 0.5805128205128205,
|
| 5 |
+
"test_auc": 0.6089684249503484,
|
| 6 |
+
"val_auc": 0.6191350469196246,
|
| 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 |
+
17781
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.574548577229446,
|
| 18 |
+
"test_acc": 0.5764102564102564,
|
| 19 |
+
"test_auc": 0.5974265324670953,
|
| 20 |
+
"val_auc": 0.6138310893512852,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "116_nyc_borough_Bronx",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
17781,
|
| 28 |
+
6774
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.5627916027969705,
|
| 33 |
+
"test_acc": 0.5682051282051283,
|
| 34 |
+
"test_auc": 0.5943590736190123,
|
| 35 |
+
"val_auc": 0.620563035495716,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "116_nyc_borough_Bronx",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
17781,
|
| 43 |
+
6774,
|
| 44 |
+
4588,
|
| 45 |
+
8555,
|
| 46 |
+
560
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.5632759979094022,
|
| 51 |
+
"test_acc": 0.5651282051282052,
|
| 52 |
+
"test_auc": 0.5759332817181136,
|
| 53 |
+
"val_auc": 0.6344349245206038,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "116_nyc_borough_Bronx",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
17781,
|
| 61 |
+
6774,
|
| 62 |
+
4588,
|
| 63 |
+
8555,
|
| 64 |
+
560,
|
| 65 |
+
9573,
|
| 66 |
+
15320,
|
| 67 |
+
7112,
|
| 68 |
+
12271,
|
| 69 |
+
23939
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.5451961033254565,
|
| 74 |
+
"test_acc": 0.5456410256410257,
|
| 75 |
+
"test_auc": 0.5726259804086579,
|
| 76 |
+
"val_auc": 0.6589147286821705,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "116_nyc_borough_Bronx",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
17781,
|
| 84 |
+
6774,
|
| 85 |
+
4588,
|
| 86 |
+
8555,
|
| 87 |
+
560,
|
| 88 |
+
9573,
|
| 89 |
+
15320,
|
| 90 |
+
7112,
|
| 91 |
+
12271,
|
| 92 |
+
23939,
|
| 93 |
+
27134,
|
| 94 |
+
24903,
|
| 95 |
+
2880,
|
| 96 |
+
6647,
|
| 97 |
+
3055,
|
| 98 |
+
977
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6960397758270098,
|
| 4 |
+
"test_acc": 0.696969696969697,
|
| 5 |
+
"test_auc": 0.7461224489795917,
|
| 6 |
+
"val_auc": 0.7363782051282051,
|
| 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 |
+
12573
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7070707070707071,
|
| 18 |
+
"test_acc": 0.7070707070707071,
|
| 19 |
+
"test_auc": 0.7453061224489795,
|
| 20 |
+
"val_auc": 0.797275641025641,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "117_us_state_FL",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
12573,
|
| 28 |
+
29959
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7169984155278273,
|
| 33 |
+
"test_acc": 0.7171717171717171,
|
| 34 |
+
"test_auc": 0.7816326530612245,
|
| 35 |
+
"val_auc": 0.8589743589743589,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "117_us_state_FL",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
12573,
|
| 43 |
+
29959,
|
| 44 |
+
4152,
|
| 45 |
+
26210,
|
| 46 |
+
18010
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.706711468285306,
|
| 51 |
+
"test_acc": 0.7070707070707071,
|
| 52 |
+
"test_auc": 0.7942857142857144,
|
| 53 |
+
"val_auc": 0.8938301282051282,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "117_us_state_FL",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
12573,
|
| 61 |
+
29959,
|
| 62 |
+
4152,
|
| 63 |
+
26210,
|
| 64 |
+
18010,
|
| 65 |
+
1965,
|
| 66 |
+
10773,
|
| 67 |
+
13688,
|
| 68 |
+
8352,
|
| 69 |
+
25517
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7574272133095662,
|
| 74 |
+
"test_acc": 0.7575757575757576,
|
| 75 |
+
"test_auc": 0.8375510204081633,
|
| 76 |
+
"val_auc": 0.921073717948718,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "117_us_state_FL",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
12573,
|
| 84 |
+
29959,
|
| 85 |
+
4152,
|
| 86 |
+
26210,
|
| 87 |
+
18010,
|
| 88 |
+
1965,
|
| 89 |
+
10773,
|
| 90 |
+
13688,
|
| 91 |
+
8352,
|
| 92 |
+
25517,
|
| 93 |
+
4493,
|
| 94 |
+
28368,
|
| 95 |
+
22127,
|
| 96 |
+
19991,
|
| 97 |
+
132,
|
| 98 |
+
28538
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5947197011026799,
|
| 4 |
+
"test_acc": 0.5959595959595959,
|
| 5 |
+
"test_auc": 0.6657142857142857,
|
| 6 |
+
"val_auc": 0.680088141025641,
|
| 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 |
+
12844
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.706711468285306,
|
| 18 |
+
"test_acc": 0.7070707070707071,
|
| 19 |
+
"test_auc": 0.7377551020408163,
|
| 20 |
+
"val_auc": 0.7694310897435896,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "118_us_state_CA",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
12844,
|
| 28 |
+
5873
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.747165058866643,
|
| 33 |
+
"test_acc": 0.7474747474747475,
|
| 34 |
+
"test_auc": 0.806530612244898,
|
| 35 |
+
"val_auc": 0.8613782051282051,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "118_us_state_CA",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
12844,
|
| 43 |
+
5873,
|
| 44 |
+
32332,
|
| 45 |
+
13572,
|
| 46 |
+
1772
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7565316827006736,
|
| 51 |
+
"test_acc": 0.7575757575757576,
|
| 52 |
+
"test_auc": 0.8424489795918367,
|
| 53 |
+
"val_auc": 0.8974358974358975,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "118_us_state_CA",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
12844,
|
| 61 |
+
5873,
|
| 62 |
+
32332,
|
| 63 |
+
13572,
|
| 64 |
+
1772,
|
| 65 |
+
27458,
|
| 66 |
+
8657,
|
| 67 |
+
12571,
|
| 68 |
+
6123,
|
| 69 |
+
23645
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7460224095738114,
|
| 74 |
+
"test_acc": 0.7474747474747475,
|
| 75 |
+
"test_auc": 0.8428571428571429,
|
| 76 |
+
"val_auc": 0.9258814102564102,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "118_us_state_CA",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
12844,
|
| 84 |
+
5873,
|
| 85 |
+
32332,
|
| 86 |
+
13572,
|
| 87 |
+
1772,
|
| 88 |
+
27458,
|
| 89 |
+
8657,
|
| 90 |
+
12571,
|
| 91 |
+
6123,
|
| 92 |
+
23645,
|
| 93 |
+
15416,
|
| 94 |
+
24846,
|
| 95 |
+
24562,
|
| 96 |
+
7200,
|
| 97 |
+
28770,
|
| 98 |
+
28240
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5894157522064499,
|
| 4 |
+
"test_acc": 0.5959595959595959,
|
| 5 |
+
"test_auc": 0.6502040816326531,
|
| 6 |
+
"val_auc": 0.6682692307692308,
|
| 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 |
+
12728
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6133183688739244,
|
| 18 |
+
"test_acc": 0.6161616161616161,
|
| 19 |
+
"test_auc": 0.6599999999999999,
|
| 20 |
+
"val_auc": 0.764823717948718,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "119_us_state_TX",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
12728,
|
| 28 |
+
17985
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6431847596675966,
|
| 33 |
+
"test_acc": 0.6464646464646465,
|
| 34 |
+
"test_auc": 0.7424489795918368,
|
| 35 |
+
"val_auc": 0.811698717948718,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "119_us_state_TX",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
12728,
|
| 43 |
+
17985,
|
| 44 |
+
29411,
|
| 45 |
+
22336,
|
| 46 |
+
18336
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.693526170798898,
|
| 51 |
+
"test_acc": 0.696969696969697,
|
| 52 |
+
"test_auc": 0.7477551020408163,
|
| 53 |
+
"val_auc": 0.844150641025641,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "119_us_state_TX",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
12728,
|
| 61 |
+
17985,
|
| 62 |
+
29411,
|
| 63 |
+
22336,
|
| 64 |
+
18336,
|
| 65 |
+
27461,
|
| 66 |
+
19252,
|
| 67 |
+
12671,
|
| 68 |
+
13420,
|
| 69 |
+
29965
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7426897817319473,
|
| 74 |
+
"test_acc": 0.7474747474747475,
|
| 75 |
+
"test_auc": 0.8183673469387756,
|
| 76 |
+
"val_auc": 0.8493589743589743,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "119_us_state_TX",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
12728,
|
| 84 |
+
17985,
|
| 85 |
+
29411,
|
| 86 |
+
22336,
|
| 87 |
+
18336,
|
| 88 |
+
27461,
|
| 89 |
+
19252,
|
| 90 |
+
12671,
|
| 91 |
+
13420,
|
| 92 |
+
29965,
|
| 93 |
+
28679,
|
| 94 |
+
29412,
|
| 95 |
+
13602,
|
| 96 |
+
31363,
|
| 97 |
+
25891,
|
| 98 |
+
26896
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6322269021487078,
|
| 4 |
+
"test_acc": 0.6363242306025141,
|
| 5 |
+
"test_auc": 0.674532265313454,
|
| 6 |
+
"val_auc": 0.706,
|
| 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 |
+
6501
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6571380109003478,
|
| 18 |
+
"test_acc": 0.657997399219766,
|
| 19 |
+
"test_auc": 0.714645390444038,
|
| 20 |
+
"val_auc": 0.7976,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "120_us_timezone_Chicago",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
6501,
|
| 28 |
+
30782
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6748990500765276,
|
| 33 |
+
"test_acc": 0.6749024707412223,
|
| 34 |
+
"test_auc": 0.7308573369749024,
|
| 35 |
+
"val_auc": 0.8164,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "120_us_timezone_Chicago",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
6501,
|
| 43 |
+
30782,
|
| 44 |
+
12886,
|
| 45 |
+
8073,
|
| 46 |
+
16762
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.6904802842662108,
|
| 51 |
+
"test_acc": 0.6905071521456437,
|
| 52 |
+
"test_auc": 0.750909765948524,
|
| 53 |
+
"val_auc": 0.8451999999999998,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "120_us_timezone_Chicago",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
6501,
|
| 61 |
+
30782,
|
| 62 |
+
12886,
|
| 63 |
+
8073,
|
| 64 |
+
16762,
|
| 65 |
+
2881,
|
| 66 |
+
26277,
|
| 67 |
+
16792,
|
| 68 |
+
7546,
|
| 69 |
+
22172
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7124760267240441,
|
| 74 |
+
"test_acc": 0.7126137841352406,
|
| 75 |
+
"test_auc": 0.7789347659109458,
|
| 76 |
+
"val_auc": 0.8688,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "120_us_timezone_Chicago",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
6501,
|
| 84 |
+
30782,
|
| 85 |
+
12886,
|
| 86 |
+
8073,
|
| 87 |
+
16762,
|
| 88 |
+
2881,
|
| 89 |
+
26277,
|
| 90 |
+
16792,
|
| 91 |
+
7546,
|
| 92 |
+
22172,
|
| 93 |
+
16457,
|
| 94 |
+
26630,
|
| 95 |
+
5342,
|
| 96 |
+
19540,
|
| 97 |
+
3641,
|
| 98 |
+
1484
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5796291855876481,
|
| 4 |
+
"test_acc": 0.5873428695275249,
|
| 5 |
+
"test_auc": 0.6088641491339749,
|
| 6 |
+
"val_auc": 0.7556,
|
| 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 |
+
23213
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6110580356082951,
|
| 18 |
+
"test_acc": 0.612917208495882,
|
| 19 |
+
"test_auc": 0.654742507301426,
|
| 20 |
+
"val_auc": 0.6612,
|
| 21 |
+
"k": 2,
|
| 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 |
+
23213,
|
| 28 |
+
12771
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6756790884618604,
|
| 33 |
+
"test_acc": 0.6757693974859125,
|
| 34 |
+
"test_auc": 0.7405359539803482,
|
| 35 |
+
"val_auc": 0.7363999999999999,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "121_us_timezone_New_York",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
23213,
|
| 43 |
+
12771,
|
| 44 |
+
20622,
|
| 45 |
+
5921,
|
| 46 |
+
12122
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.689201506371276,
|
| 51 |
+
"test_acc": 0.6892067620286085,
|
| 52 |
+
"test_auc": 0.7581390419987946,
|
| 53 |
+
"val_auc": 0.7932,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "121_us_timezone_New_York",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
23213,
|
| 61 |
+
12771,
|
| 62 |
+
20622,
|
| 63 |
+
5921,
|
| 64 |
+
12122,
|
| 65 |
+
5927,
|
| 66 |
+
29412,
|
| 67 |
+
1484,
|
| 68 |
+
5021,
|
| 69 |
+
7059
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.6917217016384938,
|
| 74 |
+
"test_acc": 0.6918075422626788,
|
| 75 |
+
"test_auc": 0.7662055582528284,
|
| 76 |
+
"val_auc": 0.7919999999999999,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "121_us_timezone_New_York",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
23213,
|
| 84 |
+
12771,
|
| 85 |
+
20622,
|
| 86 |
+
5921,
|
| 87 |
+
12122,
|
| 88 |
+
5927,
|
| 89 |
+
29412,
|
| 90 |
+
1484,
|
| 91 |
+
5021,
|
| 92 |
+
7059,
|
| 93 |
+
23659,
|
| 94 |
+
16436,
|
| 95 |
+
17344,
|
| 96 |
+
7561,
|
| 97 |
+
8276,
|
| 98 |
+
16972
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6273476621715769,
|
| 4 |
+
"test_acc": 0.6285218899003034,
|
| 5 |
+
"test_auc": 0.6681319622836064,
|
| 6 |
+
"val_auc": 0.7388,
|
| 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 |
+
9602
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6836103511080227,
|
| 18 |
+
"test_acc": 0.6840052015604682,
|
| 19 |
+
"test_auc": 0.7348545201200696,
|
| 20 |
+
"val_auc": 0.8347999999999999,
|
| 21 |
+
"k": 2,
|
| 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 |
+
9602,
|
| 28 |
+
23645
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7038392339776374,
|
| 33 |
+
"test_acc": 0.7039445166883398,
|
| 34 |
+
"test_auc": 0.7755474754276765,
|
| 35 |
+
"val_auc": 0.8704,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "122_us_timezone_Los_Angeles",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
9602,
|
| 43 |
+
23645,
|
| 44 |
+
32332,
|
| 45 |
+
8073,
|
| 46 |
+
12571
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7171743129247582,
|
| 51 |
+
"test_acc": 0.717381881231036,
|
| 52 |
+
"test_auc": 0.7824024735412556,
|
| 53 |
+
"val_auc": 0.8632000000000001,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "122_us_timezone_Los_Angeles",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
9602,
|
| 61 |
+
23645,
|
| 62 |
+
32332,
|
| 63 |
+
8073,
|
| 64 |
+
12571,
|
| 65 |
+
2871,
|
| 66 |
+
9131,
|
| 67 |
+
10529,
|
| 68 |
+
10310,
|
| 69 |
+
20025
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.72972094934154,
|
| 74 |
+
"test_acc": 0.729952319029042,
|
| 75 |
+
"test_auc": 0.8067162597458818,
|
| 76 |
+
"val_auc": 0.8528,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "122_us_timezone_Los_Angeles",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
9602,
|
| 84 |
+
23645,
|
| 85 |
+
32332,
|
| 86 |
+
8073,
|
| 87 |
+
12571,
|
| 88 |
+
2871,
|
| 89 |
+
9131,
|
| 90 |
+
10529,
|
| 91 |
+
10310,
|
| 92 |
+
20025,
|
| 93 |
+
30670,
|
| 94 |
+
32439,
|
| 95 |
+
17242,
|
| 96 |
+
3245,
|
| 97 |
+
24203,
|
| 98 |
+
17350
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7531840479208901,
|
| 4 |
+
"test_acc": 0.7538461538461538,
|
| 5 |
+
"test_auc": 0.8457644662874071,
|
| 6 |
+
"val_auc": 0.8643410852713178,
|
| 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 |
+
8031
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.824598778407884,
|
| 18 |
+
"test_acc": 0.8246153846153846,
|
| 19 |
+
"test_auc": 0.8817450432557983,
|
| 20 |
+
"val_auc": 0.9053447572419421,
|
| 21 |
+
"k": 2,
|
| 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 |
+
8031,
|
| 28 |
+
2475
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.8799613596113232,
|
| 33 |
+
"test_acc": 0.88,
|
| 34 |
+
"test_auc": 0.9493764096004308,
|
| 35 |
+
"val_auc": 0.9579763361893106,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "123_world_country_United_Kingdom",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
8031,
|
| 43 |
+
2475,
|
| 44 |
+
10829,
|
| 45 |
+
108,
|
| 46 |
+
15757
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.8891909033862007,
|
| 51 |
+
"test_acc": 0.8892307692307693,
|
| 52 |
+
"test_auc": 0.9639773454068066,
|
| 53 |
+
"val_auc": 0.974296205630355,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "123_world_country_United_Kingdom",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
8031,
|
| 61 |
+
2475,
|
| 62 |
+
10829,
|
| 63 |
+
108,
|
| 64 |
+
15757,
|
| 65 |
+
3101,
|
| 66 |
+
764,
|
| 67 |
+
5294,
|
| 68 |
+
27674,
|
| 69 |
+
26734
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.899423697705803,
|
| 74 |
+
"test_acc": 0.8994871794871795,
|
| 75 |
+
"test_auc": 0.9706466489379607,
|
| 76 |
+
"val_auc": 0.97062423500612,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "123_world_country_United_Kingdom",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
8031,
|
| 84 |
+
2475,
|
| 85 |
+
10829,
|
| 86 |
+
108,
|
| 87 |
+
15757,
|
| 88 |
+
3101,
|
| 89 |
+
764,
|
| 90 |
+
5294,
|
| 91 |
+
27674,
|
| 92 |
+
26734,
|
| 93 |
+
23087,
|
| 94 |
+
1444,
|
| 95 |
+
23767,
|
| 96 |
+
14158,
|
| 97 |
+
13840,
|
| 98 |
+
12497
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8163291096373262,
|
| 4 |
+
"test_acc": 0.8164102564102564,
|
| 5 |
+
"test_auc": 0.8894789106944492,
|
| 6 |
+
"val_auc": 0.949408404732762,
|
| 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 |
+
30842
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.842831911851456,
|
| 18 |
+
"test_acc": 0.8430769230769231,
|
| 19 |
+
"test_auc": 0.9195749318342478,
|
| 20 |
+
"val_auc": 0.9702162382700938,
|
| 21 |
+
"k": 2,
|
| 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 |
+
30842,
|
| 28 |
+
335
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.9055862036505676,
|
| 33 |
+
"test_acc": 0.9056410256410257,
|
| 34 |
+
"test_auc": 0.969369593025213,
|
| 35 |
+
"val_auc": 0.9840881272949816,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "124_world_country_United_States",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
30842,
|
| 43 |
+
335,
|
| 44 |
+
12376,
|
| 45 |
+
7546,
|
| 46 |
+
4132
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.9025362159599493,
|
| 51 |
+
"test_acc": 0.9025641025641026,
|
| 52 |
+
"test_auc": 0.9725738041539032,
|
| 53 |
+
"val_auc": 0.9885760913912689,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "124_world_country_United_States",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
30842,
|
| 61 |
+
335,
|
| 62 |
+
12376,
|
| 63 |
+
7546,
|
| 64 |
+
4132,
|
| 65 |
+
9553,
|
| 66 |
+
6525,
|
| 67 |
+
848,
|
| 68 |
+
29430,
|
| 69 |
+
23913
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.9281702199865419,
|
| 74 |
+
"test_acc": 0.9282051282051282,
|
| 75 |
+
"test_auc": 0.9841451509745178,
|
| 76 |
+
"val_auc": 0.9914320685434517,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "124_world_country_United_States",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
30842,
|
| 84 |
+
335,
|
| 85 |
+
12376,
|
| 86 |
+
7546,
|
| 87 |
+
4132,
|
| 88 |
+
9553,
|
| 89 |
+
6525,
|
| 90 |
+
848,
|
| 91 |
+
29430,
|
| 92 |
+
23913,
|
| 93 |
+
19804,
|
| 94 |
+
693,
|
| 95 |
+
12511,
|
| 96 |
+
26811,
|
| 97 |
+
14668,
|
| 98 |
+
16110
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.816077076378132,
|
| 4 |
+
"test_acc": 0.8164102564102564,
|
| 5 |
+
"test_auc": 0.8972085703706197,
|
| 6 |
+
"val_auc": 0.8714810281517749,
|
| 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 |
+
17911
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8264366094367707,
|
| 18 |
+
"test_acc": 0.8266666666666667,
|
| 19 |
+
"test_auc": 0.8996869424714714,
|
| 20 |
+
"val_auc": 0.8965728274173806,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "125_world_country_Italy",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
17911,
|
| 28 |
+
18619
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.8830510878010878,
|
| 33 |
+
"test_acc": 0.8830769230769231,
|
| 34 |
+
"test_auc": 0.9502621436025178,
|
| 35 |
+
"val_auc": 0.930640554875561,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "125_world_country_Italy",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
17911,
|
| 43 |
+
18619,
|
| 44 |
+
24336,
|
| 45 |
+
125,
|
| 46 |
+
32403
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.9169181826472836,
|
| 51 |
+
"test_acc": 0.916923076923077,
|
| 52 |
+
"test_auc": 0.9790958359982496,
|
| 53 |
+
"val_auc": 0.971032231742146,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "125_world_country_Italy",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
17911,
|
| 61 |
+
18619,
|
| 62 |
+
24336,
|
| 63 |
+
125,
|
| 64 |
+
32403,
|
| 65 |
+
14075,
|
| 66 |
+
15676,
|
| 67 |
+
27608,
|
| 68 |
+
5484,
|
| 69 |
+
25388
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.9569211736236362,
|
| 74 |
+
"test_acc": 0.9569230769230769,
|
| 75 |
+
"test_auc": 0.9898550846601811,
|
| 76 |
+
"val_auc": 0.9857201142390861,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "125_world_country_Italy",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
17911,
|
| 84 |
+
18619,
|
| 85 |
+
24336,
|
| 86 |
+
125,
|
| 87 |
+
32403,
|
| 88 |
+
14075,
|
| 89 |
+
15676,
|
| 90 |
+
27608,
|
| 91 |
+
5484,
|
| 92 |
+
25388,
|
| 93 |
+
32382,
|
| 94 |
+
5641,
|
| 95 |
+
27560,
|
| 96 |
+
29127,
|
| 97 |
+
5604,
|
| 98 |
+
25483
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6133368505929642,
|
| 4 |
+
"test_acc": 0.6146510619852622,
|
| 5 |
+
"test_auc": 0.649764535587218,
|
| 6 |
+
"val_auc": 0.702,
|
| 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 |
+
32265
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6547633682983756,
|
| 18 |
+
"test_acc": 0.6549631556133507,
|
| 19 |
+
"test_auc": 0.7114640279821609,
|
| 20 |
+
"val_auc": 0.7447999999999999,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "126_art_type_book",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
32265,
|
| 28 |
+
12409
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6773218213332712,
|
| 33 |
+
"test_acc": 0.6783701777199826,
|
| 34 |
+
"test_auc": 0.7428905981081679,
|
| 35 |
+
"val_auc": 0.7711999999999999,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "126_art_type_book",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
32265,
|
| 43 |
+
12409,
|
| 44 |
+
26349,
|
| 45 |
+
24110,
|
| 46 |
+
29426
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.6884659677246376,
|
| 51 |
+
"test_acc": 0.6892067620286085,
|
| 52 |
+
"test_auc": 0.7609979843103891,
|
| 53 |
+
"val_auc": 0.8284,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "126_art_type_book",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
32265,
|
| 61 |
+
12409,
|
| 62 |
+
26349,
|
| 63 |
+
24110,
|
| 64 |
+
29426,
|
| 65 |
+
31059,
|
| 66 |
+
19329,
|
| 67 |
+
14734,
|
| 68 |
+
10729,
|
| 69 |
+
4760
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.6934495755517961,
|
| 74 |
+
"test_acc": 0.693974859124404,
|
| 75 |
+
"test_auc": 0.7699430766848896,
|
| 76 |
+
"val_auc": 0.8536,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "126_art_type_book",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
32265,
|
| 84 |
+
12409,
|
| 85 |
+
26349,
|
| 86 |
+
24110,
|
| 87 |
+
29426,
|
| 88 |
+
31059,
|
| 89 |
+
19329,
|
| 90 |
+
14734,
|
| 91 |
+
10729,
|
| 92 |
+
4760,
|
| 93 |
+
3803,
|
| 94 |
+
31186,
|
| 95 |
+
12622,
|
| 96 |
+
11172,
|
| 97 |
+
881,
|
| 98 |
+
10071
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6618550984693806,
|
| 4 |
+
"test_acc": 0.6662332032943217,
|
| 5 |
+
"test_auc": 0.7127127484476484,
|
| 6 |
+
"val_auc": 0.7918000000000001,
|
| 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 |
+
3995
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6799254805274386,
|
| 18 |
+
"test_acc": 0.682271348071088,
|
| 19 |
+
"test_auc": 0.7473375911832745,
|
| 20 |
+
"val_auc": 0.8144,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "127_art_type_song",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
3995,
|
| 28 |
+
14014
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7252914728362629,
|
| 33 |
+
"test_acc": 0.7256176853055917,
|
| 34 |
+
"test_auc": 0.8027949843750235,
|
| 35 |
+
"val_auc": 0.8024,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "127_art_type_song",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
3995,
|
| 43 |
+
14014,
|
| 44 |
+
4760,
|
| 45 |
+
8770,
|
| 46 |
+
13144
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7522409392993334,
|
| 51 |
+
"test_acc": 0.7524924143909839,
|
| 52 |
+
"test_auc": 0.8220586489017423,
|
| 53 |
+
"val_auc": 0.8224,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "127_art_type_song",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
3995,
|
| 61 |
+
14014,
|
| 62 |
+
4760,
|
| 63 |
+
8770,
|
| 64 |
+
13144,
|
| 65 |
+
27262,
|
| 66 |
+
15841,
|
| 67 |
+
8602,
|
| 68 |
+
2290,
|
| 69 |
+
27265
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7485938998737799,
|
| 74 |
+
"test_acc": 0.7490247074122237,
|
| 75 |
+
"test_auc": 0.8211492587342792,
|
| 76 |
+
"val_auc": 0.8104,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "127_art_type_song",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
3995,
|
| 84 |
+
14014,
|
| 85 |
+
4760,
|
| 86 |
+
8770,
|
| 87 |
+
13144,
|
| 88 |
+
27262,
|
| 89 |
+
15841,
|
| 90 |
+
8602,
|
| 91 |
+
2290,
|
| 92 |
+
27265,
|
| 93 |
+
10925,
|
| 94 |
+
15452,
|
| 95 |
+
31186,
|
| 96 |
+
26732,
|
| 97 |
+
23163,
|
| 98 |
+
2307
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5531413032579486,
|
| 4 |
+
"test_acc": 0.5726051148677936,
|
| 5 |
+
"test_auc": 0.5856390006628778,
|
| 6 |
+
"val_auc": 0.6195999999999999,
|
| 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 |
+
31936
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5863278766522667,
|
| 18 |
+
"test_acc": 0.5869094061551798,
|
| 19 |
+
"test_auc": 0.6275306975548678,
|
| 20 |
+
"val_auc": 0.6716000000000001,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "128_art_type_movie",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
31936,
|
| 28 |
+
1835
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6177528848311415,
|
| 33 |
+
"test_acc": 0.6185522323363676,
|
| 34 |
+
"test_auc": 0.6647394108654838,
|
| 35 |
+
"val_auc": 0.6604,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "128_art_type_movie",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
31936,
|
| 43 |
+
1835,
|
| 44 |
+
6698,
|
| 45 |
+
11520,
|
| 46 |
+
27731
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.627505685000984,
|
| 51 |
+
"test_acc": 0.6285218899003034,
|
| 52 |
+
"test_auc": 0.6894612953022857,
|
| 53 |
+
"val_auc": 0.6892,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "128_art_type_movie",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
31936,
|
| 61 |
+
1835,
|
| 62 |
+
6698,
|
| 63 |
+
11520,
|
| 64 |
+
27731,
|
| 65 |
+
14018,
|
| 66 |
+
10166,
|
| 67 |
+
28175,
|
| 68 |
+
14437,
|
| 69 |
+
23034
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.6500763197701822,
|
| 74 |
+
"test_acc": 0.6501950585175552,
|
| 75 |
+
"test_auc": 0.7194989786270762,
|
| 76 |
+
"val_auc": 0.706,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "128_art_type_movie",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
31936,
|
| 84 |
+
1835,
|
| 85 |
+
6698,
|
| 86 |
+
11520,
|
| 87 |
+
27731,
|
| 88 |
+
14018,
|
| 89 |
+
10166,
|
| 90 |
+
28175,
|
| 91 |
+
14437,
|
| 92 |
+
23034,
|
| 93 |
+
3913,
|
| 94 |
+
1610,
|
| 95 |
+
14796,
|
| 96 |
+
32196,
|
| 97 |
+
17012,
|
| 98 |
+
27141
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6352602573904274,
|
| 4 |
+
"test_acc": 0.6396588486140725,
|
| 5 |
+
"test_auc": 0.689079832696854,
|
| 6 |
+
"val_auc": 0.664,
|
| 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 |
+
12178
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.653737435971159,
|
| 18 |
+
"test_acc": 0.6567164179104478,
|
| 19 |
+
"test_auc": 0.734497181305692,
|
| 20 |
+
"val_auc": 0.7220000000000001,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "129_arith_mc_A",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
12178,
|
| 28 |
+
2923
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7078492789143808,
|
| 33 |
+
"test_acc": 0.7078891257995735,
|
| 34 |
+
"test_auc": 0.8044917257683215,
|
| 35 |
+
"val_auc": 0.7672,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "129_arith_mc_A",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
12178,
|
| 43 |
+
2923,
|
| 44 |
+
3102,
|
| 45 |
+
11739,
|
| 46 |
+
9831
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7526652452025586,
|
| 51 |
+
"test_acc": 0.7526652452025586,
|
| 52 |
+
"test_auc": 0.8356792144026186,
|
| 53 |
+
"val_auc": 0.852,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "129_arith_mc_A",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
12178,
|
| 61 |
+
2923,
|
| 62 |
+
3102,
|
| 63 |
+
11739,
|
| 64 |
+
9831,
|
| 65 |
+
32603,
|
| 66 |
+
19951,
|
| 67 |
+
29826,
|
| 68 |
+
17971,
|
| 69 |
+
2960
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7590004752467314,
|
| 74 |
+
"test_acc": 0.7590618336886994,
|
| 75 |
+
"test_auc": 0.8446444808146936,
|
| 76 |
+
"val_auc": 0.8644,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "129_arith_mc_A",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
12178,
|
| 84 |
+
2923,
|
| 85 |
+
3102,
|
| 86 |
+
11739,
|
| 87 |
+
9831,
|
| 88 |
+
32603,
|
| 89 |
+
19951,
|
| 90 |
+
29826,
|
| 91 |
+
17971,
|
| 92 |
+
2960,
|
| 93 |
+
21712,
|
| 94 |
+
23376,
|
| 95 |
+
8817,
|
| 96 |
+
9966,
|
| 97 |
+
6088,
|
| 98 |
+
24057
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8176238654962059,
|
| 4 |
+
"test_acc": 0.8181818181818182,
|
| 5 |
+
"test_auc": 0.896326530612245,
|
| 6 |
+
"val_auc": 0.8854669887278582,
|
| 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 |
+
15871
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8482990353199857,
|
| 18 |
+
"test_acc": 0.8484848484848485,
|
| 19 |
+
"test_auc": 0.9555102040816327,
|
| 20 |
+
"val_auc": 0.9315619967793881,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "130_temp_cat_Frequency",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
15871,
|
| 28 |
+
30387
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.888752625901323,
|
| 33 |
+
"test_acc": 0.8888888888888888,
|
| 34 |
+
"test_auc": 0.9771428571428571,
|
| 35 |
+
"val_auc": 0.9834943639291466,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "130_temp_cat_Frequency",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
15871,
|
| 43 |
+
30387,
|
| 44 |
+
30604,
|
| 45 |
+
22751,
|
| 46 |
+
17547
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.918843894233558,
|
| 51 |
+
"test_acc": 0.9191919191919192,
|
| 52 |
+
"test_auc": 0.9808163265306122,
|
| 53 |
+
"val_auc": 0.9979871175523349,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "130_temp_cat_Frequency",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
15871,
|
| 61 |
+
30387,
|
| 62 |
+
30604,
|
| 63 |
+
22751,
|
| 64 |
+
17547,
|
| 65 |
+
8363,
|
| 66 |
+
11100,
|
| 67 |
+
26620,
|
| 68 |
+
3375,
|
| 69 |
+
22683
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.9797856011091305,
|
| 74 |
+
"test_acc": 0.9797979797979798,
|
| 75 |
+
"test_auc": 0.9987755102040816,
|
| 76 |
+
"val_auc": 0.999597423510467,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "130_temp_cat_Frequency",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
15871,
|
| 84 |
+
30387,
|
| 85 |
+
30604,
|
| 86 |
+
22751,
|
| 87 |
+
17547,
|
| 88 |
+
8363,
|
| 89 |
+
11100,
|
| 90 |
+
26620,
|
| 91 |
+
3375,
|
| 92 |
+
22683,
|
| 93 |
+
22895,
|
| 94 |
+
13907,
|
| 95 |
+
10159,
|
| 96 |
+
12963,
|
| 97 |
+
14231,
|
| 98 |
+
7677
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7761316872427984,
|
| 4 |
+
"test_acc": 0.7777777777777778,
|
| 5 |
+
"test_auc": 0.8497959183673469,
|
| 6 |
+
"val_auc": 0.8548711755233495,
|
| 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 |
+
30892
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8054574951126675,
|
| 18 |
+
"test_acc": 0.8080808080808081,
|
| 19 |
+
"test_auc": 0.9195918367346939,
|
| 20 |
+
"val_auc": 0.8691626409017713,
|
| 21 |
+
"k": 2,
|
| 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 |
+
30892,
|
| 28 |
+
17881
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.8679316529783819,
|
| 33 |
+
"test_acc": 0.8686868686868687,
|
| 34 |
+
"test_auc": 0.966938775510204,
|
| 35 |
+
"val_auc": 0.963768115942029,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "131_temp_cat_Typical Time",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
30892,
|
| 43 |
+
17881,
|
| 44 |
+
30998,
|
| 45 |
+
3067,
|
| 46 |
+
255
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.9089794211919915,
|
| 51 |
+
"test_acc": 0.9090909090909091,
|
| 52 |
+
"test_auc": 0.9751020408163266,
|
| 53 |
+
"val_auc": 0.9778582930756844,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "131_temp_cat_Typical Time",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
30892,
|
| 61 |
+
17881,
|
| 62 |
+
30998,
|
| 63 |
+
3067,
|
| 64 |
+
255,
|
| 65 |
+
27113,
|
| 66 |
+
7149,
|
| 67 |
+
17711,
|
| 68 |
+
1553,
|
| 69 |
+
203
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.9290757032692516,
|
| 74 |
+
"test_acc": 0.9292929292929293,
|
| 75 |
+
"test_auc": 0.9820408163265306,
|
| 76 |
+
"val_auc": 0.9867149758454107,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "131_temp_cat_Typical Time",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
30892,
|
| 84 |
+
17881,
|
| 85 |
+
30998,
|
| 86 |
+
3067,
|
| 87 |
+
255,
|
| 88 |
+
27113,
|
| 89 |
+
7149,
|
| 90 |
+
17711,
|
| 91 |
+
1553,
|
| 92 |
+
203,
|
| 93 |
+
7479,
|
| 94 |
+
13126,
|
| 95 |
+
30571,
|
| 96 |
+
1241,
|
| 97 |
+
29991,
|
| 98 |
+
29025
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9191754277468563,
|
| 4 |
+
"test_acc": 0.9191919191919192,
|
| 5 |
+
"test_auc": 0.9616326530612245,
|
| 6 |
+
"val_auc": 0.9565217391304348,
|
| 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 |
+
32332
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9595959595959596,
|
| 18 |
+
"test_acc": 0.9595959595959596,
|
| 19 |
+
"test_auc": 0.9832653061224489,
|
| 20 |
+
"val_auc": 0.9887278582930757,
|
| 21 |
+
"k": 2,
|
| 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 |
+
32332,
|
| 28 |
+
26254
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.98989898989899,
|
| 33 |
+
"test_acc": 0.98989898989899,
|
| 34 |
+
"test_auc": 1.0,
|
| 35 |
+
"val_auc": 0.9923510466988728,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "132_temp_cat_Event Ordering",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
32332,
|
| 43 |
+
26254,
|
| 44 |
+
23968,
|
| 45 |
+
24789,
|
| 46 |
+
28491
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.98989898989899,
|
| 51 |
+
"test_acc": 0.98989898989899,
|
| 52 |
+
"test_auc": 1.0,
|
| 53 |
+
"val_auc": 0.99597423510467,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "132_temp_cat_Event Ordering",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
32332,
|
| 61 |
+
26254,
|
| 62 |
+
23968,
|
| 63 |
+
24789,
|
| 64 |
+
28491,
|
| 65 |
+
17427,
|
| 66 |
+
30323,
|
| 67 |
+
13670,
|
| 68 |
+
10218,
|
| 69 |
+
13714
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 1.0,
|
| 74 |
+
"test_acc": 1.0,
|
| 75 |
+
"test_auc": 1.0,
|
| 76 |
+
"val_auc": 1.0,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "132_temp_cat_Event Ordering",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
32332,
|
| 84 |
+
26254,
|
| 85 |
+
23968,
|
| 86 |
+
24789,
|
| 87 |
+
28491,
|
| 88 |
+
17427,
|
| 89 |
+
30323,
|
| 90 |
+
13670,
|
| 91 |
+
10218,
|
| 92 |
+
13714,
|
| 93 |
+
15916,
|
| 94 |
+
30526,
|
| 95 |
+
31566,
|
| 96 |
+
12661,
|
| 97 |
+
10961,
|
| 98 |
+
3588
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7465665527211892,
|
| 4 |
+
"test_acc": 0.7466666666666667,
|
| 5 |
+
"test_auc": 0.8030183312262957,
|
| 6 |
+
"val_auc": 0.8689123376623377,
|
| 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 |
+
7859
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8397720797720798,
|
| 18 |
+
"test_acc": 0.84,
|
| 19 |
+
"test_auc": 0.9096871049304677,
|
| 20 |
+
"val_auc": 0.9366883116883117,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "133_context_type_Causality",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
7859,
|
| 28 |
+
8253
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.9066113186661131,
|
| 33 |
+
"test_acc": 0.9066666666666666,
|
| 34 |
+
"test_auc": 0.9483249051833123,
|
| 35 |
+
"val_auc": 0.9715909090909091,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "133_context_type_Causality",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
7859,
|
| 43 |
+
8253,
|
| 44 |
+
26038,
|
| 45 |
+
25710,
|
| 46 |
+
504
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.9288607817303469,
|
| 51 |
+
"test_acc": 0.9288888888888889,
|
| 52 |
+
"test_auc": 0.9616782553729456,
|
| 53 |
+
"val_auc": 0.9878246753246753,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "133_context_type_Causality",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
7859,
|
| 61 |
+
8253,
|
| 62 |
+
26038,
|
| 63 |
+
25710,
|
| 64 |
+
504,
|
| 65 |
+
15850,
|
| 66 |
+
12423,
|
| 67 |
+
4070,
|
| 68 |
+
29499,
|
| 69 |
+
20344
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.9199905168326221,
|
| 74 |
+
"test_acc": 0.92,
|
| 75 |
+
"test_auc": 0.9631795195954489,
|
| 76 |
+
"val_auc": 0.9862012987012987,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "133_context_type_Causality",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
7859,
|
| 84 |
+
8253,
|
| 85 |
+
26038,
|
| 86 |
+
25710,
|
| 87 |
+
504,
|
| 88 |
+
15850,
|
| 89 |
+
12423,
|
| 90 |
+
4070,
|
| 91 |
+
29499,
|
| 92 |
+
20344,
|
| 93 |
+
28029,
|
| 94 |
+
19243,
|
| 95 |
+
10737,
|
| 96 |
+
3061,
|
| 97 |
+
15186,
|
| 98 |
+
10951
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6918721171090685,
|
| 4 |
+
"test_acc": 0.6933333333333334,
|
| 5 |
+
"test_auc": 0.7646965865992416,
|
| 6 |
+
"val_auc": 0.7623782467532468,
|
| 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 |
+
6013
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7598006959822841,
|
| 18 |
+
"test_acc": 0.76,
|
| 19 |
+
"test_auc": 0.8251817319848292,
|
| 20 |
+
"val_auc": 0.765625,
|
| 21 |
+
"k": 2,
|
| 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 |
+
6013,
|
| 28 |
+
26305
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7733243778932327,
|
| 33 |
+
"test_acc": 0.7733333333333333,
|
| 34 |
+
"test_auc": 0.8697850821744626,
|
| 35 |
+
"val_auc": 0.817775974025974,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "134_context_type_Belief_states",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
6013,
|
| 43 |
+
26305,
|
| 44 |
+
2694,
|
| 45 |
+
23296,
|
| 46 |
+
27
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.799381945682534,
|
| 51 |
+
"test_acc": 0.8,
|
| 52 |
+
"test_auc": 0.8940423514538558,
|
| 53 |
+
"val_auc": 0.8616071428571428,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "134_context_type_Belief_states",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
6013,
|
| 61 |
+
26305,
|
| 62 |
+
2694,
|
| 63 |
+
23296,
|
| 64 |
+
27,
|
| 65 |
+
17825,
|
| 66 |
+
3954,
|
| 67 |
+
30352,
|
| 68 |
+
4066,
|
| 69 |
+
12405
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.8132226613965745,
|
| 74 |
+
"test_acc": 0.8133333333333334,
|
| 75 |
+
"test_auc": 0.9082648546144121,
|
| 76 |
+
"val_auc": 0.8754058441558442,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "134_context_type_Belief_states",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
6013,
|
| 84 |
+
26305,
|
| 85 |
+
2694,
|
| 86 |
+
23296,
|
| 87 |
+
27,
|
| 88 |
+
17825,
|
| 89 |
+
3954,
|
| 90 |
+
30352,
|
| 91 |
+
4066,
|
| 92 |
+
12405,
|
| 93 |
+
10340,
|
| 94 |
+
475,
|
| 95 |
+
20694,
|
| 96 |
+
13905,
|
| 97 |
+
8405,
|
| 98 |
+
463
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7755331984811384,
|
| 4 |
+
"test_acc": 0.7777777777777778,
|
| 5 |
+
"test_auc": 0.8369943109987357,
|
| 6 |
+
"val_auc": 0.8794642857142857,
|
| 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 |
+
2945
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8305019114564065,
|
| 18 |
+
"test_acc": 0.8311111111111111,
|
| 19 |
+
"test_auc": 0.8949905183312262,
|
| 20 |
+
"val_auc": 0.9025974025974026,
|
| 21 |
+
"k": 2,
|
| 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 |
+
2945,
|
| 28 |
+
18300
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.8933333333333333,
|
| 33 |
+
"test_acc": 0.8933333333333333,
|
| 34 |
+
"test_auc": 0.9629424778761062,
|
| 35 |
+
"val_auc": 0.9569805194805195,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "135_context_type_Event_duration",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
2945,
|
| 43 |
+
18300,
|
| 44 |
+
18015,
|
| 45 |
+
1985,
|
| 46 |
+
23698
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.9244414592977442,
|
| 51 |
+
"test_acc": 0.9244444444444444,
|
| 52 |
+
"test_auc": 0.9812737041719342,
|
| 53 |
+
"val_auc": 0.9813311688311689,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "135_context_type_Event_duration",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
2945,
|
| 61 |
+
18300,
|
| 62 |
+
18015,
|
| 63 |
+
1985,
|
| 64 |
+
23698,
|
| 65 |
+
15057,
|
| 66 |
+
14289,
|
| 67 |
+
11798,
|
| 68 |
+
22676,
|
| 69 |
+
27226
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.9333069875582077,
|
| 74 |
+
"test_acc": 0.9333333333333333,
|
| 75 |
+
"test_auc": 0.9890170670037927,
|
| 76 |
+
"val_auc": 0.9910714285714286,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "135_context_type_Event_duration",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
2945,
|
| 84 |
+
18300,
|
| 85 |
+
18015,
|
| 86 |
+
1985,
|
| 87 |
+
23698,
|
| 88 |
+
15057,
|
| 89 |
+
14289,
|
| 90 |
+
11798,
|
| 91 |
+
22676,
|
| 92 |
+
27226,
|
| 93 |
+
9650,
|
| 94 |
+
30064,
|
| 95 |
+
27187,
|
| 96 |
+
13729,
|
| 97 |
+
28936,
|
| 98 |
+
22813
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6585799355156691,
|
| 4 |
+
"test_acc": 0.6610316428261812,
|
| 5 |
+
"test_auc": 0.7079564875593922,
|
| 6 |
+
"val_auc": 0.7544,
|
| 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 |
+
336
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6826994458542652,
|
| 18 |
+
"test_acc": 0.682704811443433,
|
| 19 |
+
"test_auc": 0.7504024615162616,
|
| 20 |
+
"val_auc": 0.7904,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "136_glue_mnli_entailment",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
336,
|
| 28 |
+
4831
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7320446114837476,
|
| 33 |
+
"test_acc": 0.7325530992631123,
|
| 34 |
+
"test_auc": 0.8029283866516554,
|
| 35 |
+
"val_auc": 0.842,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "136_glue_mnli_entailment",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
336,
|
| 43 |
+
4831,
|
| 44 |
+
14216,
|
| 45 |
+
24685,
|
| 46 |
+
3746
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.735457062184834,
|
| 51 |
+
"test_acc": 0.7360208062418726,
|
| 52 |
+
"test_auc": 0.815813167669,
|
| 53 |
+
"val_auc": 0.8595999999999999,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "136_glue_mnli_entailment",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
336,
|
| 61 |
+
4831,
|
| 62 |
+
14216,
|
| 63 |
+
24685,
|
| 64 |
+
3746,
|
| 65 |
+
28387,
|
| 66 |
+
21878,
|
| 67 |
+
20578,
|
| 68 |
+
9671,
|
| 69 |
+
12789
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7414962611825544,
|
| 74 |
+
"test_acc": 0.7420892934547031,
|
| 75 |
+
"test_auc": 0.8264778341783396,
|
| 76 |
+
"val_auc": 0.8696,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "136_glue_mnli_entailment",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
336,
|
| 84 |
+
4831,
|
| 85 |
+
14216,
|
| 86 |
+
24685,
|
| 87 |
+
3746,
|
| 88 |
+
28387,
|
| 89 |
+
21878,
|
| 90 |
+
20578,
|
| 91 |
+
9671,
|
| 92 |
+
12789,
|
| 93 |
+
22571,
|
| 94 |
+
32065,
|
| 95 |
+
30719,
|
| 96 |
+
18849,
|
| 97 |
+
27807,
|
| 98 |
+
10104
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5818582479765977,
|
| 4 |
+
"test_acc": 0.5825747724317295,
|
| 5 |
+
"test_auc": 0.6117644273622724,
|
| 6 |
+
"val_auc": 0.6692,
|
| 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 |
+
19284
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6148528962358779,
|
| 18 |
+
"test_acc": 0.6163849154746424,
|
| 19 |
+
"test_auc": 0.6673931767178081,
|
| 20 |
+
"val_auc": 0.7149999999999999,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "137_glue_mnli_neutral",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
19284,
|
| 28 |
+
26464
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6683768427081636,
|
| 33 |
+
"test_acc": 0.6684005201560468,
|
| 34 |
+
"test_auc": 0.7251191601744226,
|
| 35 |
+
"val_auc": 0.7548,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "137_glue_mnli_neutral",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
19284,
|
| 43 |
+
26464,
|
| 44 |
+
11369,
|
| 45 |
+
12671,
|
| 46 |
+
16101
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.6713846276007874,
|
| 51 |
+
"test_acc": 0.671434763762462,
|
| 52 |
+
"test_auc": 0.7426561107261442,
|
| 53 |
+
"val_auc": 0.7784000000000001,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "137_glue_mnli_neutral",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
19284,
|
| 61 |
+
26464,
|
| 62 |
+
11369,
|
| 63 |
+
12671,
|
| 64 |
+
16101,
|
| 65 |
+
32236,
|
| 66 |
+
22895,
|
| 67 |
+
1906,
|
| 68 |
+
13062,
|
| 69 |
+
6621
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.6792367428542656,
|
| 74 |
+
"test_acc": 0.6792371044646728,
|
| 75 |
+
"test_auc": 0.7522888824421559,
|
| 76 |
+
"val_auc": 0.8072,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "137_glue_mnli_neutral",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
19284,
|
| 84 |
+
26464,
|
| 85 |
+
11369,
|
| 86 |
+
12671,
|
| 87 |
+
16101,
|
| 88 |
+
32236,
|
| 89 |
+
22895,
|
| 90 |
+
1906,
|
| 91 |
+
13062,
|
| 92 |
+
6621,
|
| 93 |
+
17990,
|
| 94 |
+
21419,
|
| 95 |
+
11785,
|
| 96 |
+
336,
|
| 97 |
+
13157,
|
| 98 |
+
26114
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6348946751940168,
|
| 4 |
+
"test_acc": 0.635890767230169,
|
| 5 |
+
"test_auc": 0.6821072599397848,
|
| 6 |
+
"val_auc": 0.7736,
|
| 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 |
+
30719
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6646787793243174,
|
| 18 |
+
"test_acc": 0.6649328131772865,
|
| 19 |
+
"test_auc": 0.7198135073750792,
|
| 20 |
+
"val_auc": 0.7688,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "138_glue_mnli_contradiction",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
30719,
|
| 28 |
+
1005
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7017637512829701,
|
| 33 |
+
"test_acc": 0.7017771998266147,
|
| 34 |
+
"test_auc": 0.7688736037854681,
|
| 35 |
+
"val_auc": 0.8776,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "138_glue_mnli_contradiction",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
30719,
|
| 43 |
+
1005,
|
| 44 |
+
12337,
|
| 45 |
+
32510,
|
| 46 |
+
13694
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7139043909306635,
|
| 51 |
+
"test_acc": 0.7139141742522757,
|
| 52 |
+
"test_auc": 0.800300173911475,
|
| 53 |
+
"val_auc": 0.8924,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "138_glue_mnli_contradiction",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
30719,
|
| 61 |
+
1005,
|
| 62 |
+
12337,
|
| 63 |
+
32510,
|
| 64 |
+
13694,
|
| 65 |
+
172,
|
| 66 |
+
1387,
|
| 67 |
+
8654,
|
| 68 |
+
9263,
|
| 69 |
+
16621
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7311739143319592,
|
| 74 |
+
"test_acc": 0.7312527091460772,
|
| 75 |
+
"test_auc": 0.8239240260882245,
|
| 76 |
+
"val_auc": 0.8912,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "138_glue_mnli_contradiction",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
30719,
|
| 84 |
+
1005,
|
| 85 |
+
12337,
|
| 86 |
+
32510,
|
| 87 |
+
13694,
|
| 88 |
+
172,
|
| 89 |
+
1387,
|
| 90 |
+
8654,
|
| 91 |
+
9263,
|
| 92 |
+
16621,
|
| 93 |
+
8165,
|
| 94 |
+
32710,
|
| 95 |
+
28670,
|
| 96 |
+
26345,
|
| 97 |
+
30389,
|
| 98 |
+
27449
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6883181701987255,
|
| 4 |
+
"test_acc": 0.6894915254237288,
|
| 5 |
+
"test_auc": 0.7362062562280982,
|
| 6 |
+
"val_auc": 0.6880032206119163,
|
| 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 |
+
21799
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7125637404202415,
|
| 18 |
+
"test_acc": 0.7132203389830508,
|
| 19 |
+
"test_auc": 0.7969704323908912,
|
| 20 |
+
"val_auc": 0.7773752012882448,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "139_news_class_Politics",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
21799,
|
| 28 |
+
22338
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7857625149211039,
|
| 33 |
+
"test_acc": 0.7857627118644068,
|
| 34 |
+
"test_auc": 0.8669953999404308,
|
| 35 |
+
"val_auc": 0.8651368760064413,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "139_news_class_Politics",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
21799,
|
| 43 |
+
22338,
|
| 44 |
+
15846,
|
| 45 |
+
8183,
|
| 46 |
+
24224
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.8142362635280603,
|
| 51 |
+
"test_acc": 0.8142372881355933,
|
| 52 |
+
"test_auc": 0.8819005489919214,
|
| 53 |
+
"val_auc": 0.8760064412238325,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "139_news_class_Politics",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
21799,
|
| 61 |
+
22338,
|
| 62 |
+
15846,
|
| 63 |
+
8183,
|
| 64 |
+
24224,
|
| 65 |
+
14913,
|
| 66 |
+
7124,
|
| 67 |
+
17400,
|
| 68 |
+
22166,
|
| 69 |
+
21642
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.8386213701925371,
|
| 74 |
+
"test_acc": 0.8386440677966102,
|
| 75 |
+
"test_auc": 0.9014094347184991,
|
| 76 |
+
"val_auc": 0.8900966183574879,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "139_news_class_Politics",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
21799,
|
| 84 |
+
22338,
|
| 85 |
+
15846,
|
| 86 |
+
8183,
|
| 87 |
+
24224,
|
| 88 |
+
14913,
|
| 89 |
+
7124,
|
| 90 |
+
17400,
|
| 91 |
+
22166,
|
| 92 |
+
21642,
|
| 93 |
+
16203,
|
| 94 |
+
24461,
|
| 95 |
+
12999,
|
| 96 |
+
18306,
|
| 97 |
+
3165,
|
| 98 |
+
28499
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.611602977540528,
|
| 4 |
+
"test_acc": 0.615593220338983,
|
| 5 |
+
"test_auc": 0.6656186914650694,
|
| 6 |
+
"val_auc": 0.714573268921095,
|
| 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 |
+
13087
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.679424439229786,
|
| 18 |
+
"test_acc": 0.68,
|
| 19 |
+
"test_auc": 0.7333252437001982,
|
| 20 |
+
"val_auc": 0.787842190016103,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "140_news_class_Technology",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
13087,
|
| 28 |
+
22569
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7333482639913381,
|
| 33 |
+
"test_acc": 0.7335593220338983,
|
| 34 |
+
"test_auc": 0.7870624703533332,
|
| 35 |
+
"val_auc": 0.8401771336553945,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "140_news_class_Technology",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
13087,
|
| 43 |
+
22569,
|
| 44 |
+
14313,
|
| 45 |
+
4854,
|
| 46 |
+
5786
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7524898128754856,
|
| 51 |
+
"test_acc": 0.752542372881356,
|
| 52 |
+
"test_auc": 0.8244880549212548,
|
| 53 |
+
"val_auc": 0.8671497584541062,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "140_news_class_Technology",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
13087,
|
| 61 |
+
22569,
|
| 62 |
+
14313,
|
| 63 |
+
4854,
|
| 64 |
+
5786,
|
| 65 |
+
4868,
|
| 66 |
+
29074,
|
| 67 |
+
19028,
|
| 68 |
+
30527,
|
| 69 |
+
14086
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7626998664077344,
|
| 74 |
+
"test_acc": 0.7627118644067796,
|
| 75 |
+
"test_auc": 0.8373505716061231,
|
| 76 |
+
"val_auc": 0.8719806763285024,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "140_news_class_Technology",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
13087,
|
| 84 |
+
22569,
|
| 85 |
+
14313,
|
| 86 |
+
4854,
|
| 87 |
+
5786,
|
| 88 |
+
4868,
|
| 89 |
+
29074,
|
| 90 |
+
19028,
|
| 91 |
+
30527,
|
| 92 |
+
14086,
|
| 93 |
+
1468,
|
| 94 |
+
19303,
|
| 95 |
+
1521,
|
| 96 |
+
5185,
|
| 97 |
+
20631,
|
| 98 |
+
11228
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.657751043765075,
|
| 4 |
+
"test_acc": 0.6589830508474577,
|
| 5 |
+
"test_auc": 0.6991823954874555,
|
| 6 |
+
"val_auc": 0.698671497584541,
|
| 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 |
+
10008
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7044177507075712,
|
| 18 |
+
"test_acc": 0.7057627118644068,
|
| 19 |
+
"test_auc": 0.7689655933194338,
|
| 20 |
+
"val_auc": 0.8015297906602253,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "141_news_class_Entertainment",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
10008,
|
| 28 |
+
28353
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.714029484700115,
|
| 33 |
+
"test_acc": 0.7145762711864406,
|
| 34 |
+
"test_auc": 0.7866193790838859,
|
| 35 |
+
"val_auc": 0.8623188405797101,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "141_news_class_Entertainment",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
10008,
|
| 43 |
+
28353,
|
| 44 |
+
11021,
|
| 45 |
+
8165,
|
| 46 |
+
10650
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7508463232157626,
|
| 51 |
+
"test_acc": 0.7511864406779661,
|
| 52 |
+
"test_auc": 0.8253043724467095,
|
| 53 |
+
"val_auc": 0.9082125603864735,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "141_news_class_Entertainment",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
10008,
|
| 61 |
+
28353,
|
| 62 |
+
11021,
|
| 63 |
+
8165,
|
| 64 |
+
10650,
|
| 65 |
+
14517,
|
| 66 |
+
25155,
|
| 67 |
+
1730,
|
| 68 |
+
26914,
|
| 69 |
+
29750
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7666985929675937,
|
| 74 |
+
"test_acc": 0.7667796610169492,
|
| 75 |
+
"test_auc": 0.8447121377590981,
|
| 76 |
+
"val_auc": 0.9219001610305958,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "141_news_class_Entertainment",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
10008,
|
| 84 |
+
28353,
|
| 85 |
+
11021,
|
| 86 |
+
8165,
|
| 87 |
+
10650,
|
| 88 |
+
14517,
|
| 89 |
+
25155,
|
| 90 |
+
1730,
|
| 91 |
+
26914,
|
| 92 |
+
29750,
|
| 93 |
+
17959,
|
| 94 |
+
5547,
|
| 95 |
+
17400,
|
| 96 |
+
14871,
|
| 97 |
+
30307,
|
| 98 |
+
5225
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6588341330112564,
|
| 4 |
+
"test_acc": 0.6588921282798834,
|
| 5 |
+
"test_auc": 0.7073643410852714,
|
| 6 |
+
"val_auc": 0.6604641856742697,
|
| 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 |
+
16027
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6384655265904875,
|
| 18 |
+
"test_acc": 0.6384839650145773,
|
| 19 |
+
"test_auc": 0.7000543995648034,
|
| 20 |
+
"val_auc": 0.7286914765906363,
|
| 21 |
+
"k": 2,
|
| 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 |
+
16027,
|
| 28 |
+
12170
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6471943338919963,
|
| 33 |
+
"test_acc": 0.6472303206997084,
|
| 34 |
+
"test_auc": 0.7287841697266422,
|
| 35 |
+
"val_auc": 0.7831132452981192,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "142_cancer_cat_Thyroid_Cancer",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
16027,
|
| 43 |
+
12170,
|
| 44 |
+
23396,
|
| 45 |
+
28208,
|
| 46 |
+
27551
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.6500862857166653,
|
| 51 |
+
"test_acc": 0.6501457725947521,
|
| 52 |
+
"test_auc": 0.739358085135319,
|
| 53 |
+
"val_auc": 0.8047218887555022,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "142_cancer_cat_Thyroid_Cancer",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
16027,
|
| 61 |
+
12170,
|
| 62 |
+
23396,
|
| 63 |
+
28208,
|
| 64 |
+
27551,
|
| 65 |
+
21052,
|
| 66 |
+
8522,
|
| 67 |
+
23225,
|
| 68 |
+
2731,
|
| 69 |
+
12000
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.6730357924579434,
|
| 74 |
+
"test_acc": 0.673469387755102,
|
| 75 |
+
"test_auc": 0.7488780089759282,
|
| 76 |
+
"val_auc": 0.8247298919567827,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "142_cancer_cat_Thyroid_Cancer",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
16027,
|
| 84 |
+
12170,
|
| 85 |
+
23396,
|
| 86 |
+
28208,
|
| 87 |
+
27551,
|
| 88 |
+
21052,
|
| 89 |
+
8522,
|
| 90 |
+
23225,
|
| 91 |
+
2731,
|
| 92 |
+
12000,
|
| 93 |
+
22487,
|
| 94 |
+
20852,
|
| 95 |
+
17017,
|
| 96 |
+
2503,
|
| 97 |
+
13811,
|
| 98 |
+
8288
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6618018309497838,
|
| 4 |
+
"test_acc": 0.6618075801749271,
|
| 5 |
+
"test_auc": 0.7078063375492997,
|
| 6 |
+
"val_auc": 0.75250100040016,
|
| 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 |
+
24596
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6733861061257803,
|
| 18 |
+
"test_acc": 0.673469387755102,
|
| 19 |
+
"test_auc": 0.7269821841425268,
|
| 20 |
+
"val_auc": 0.8387354941976791,
|
| 21 |
+
"k": 2,
|
| 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 |
+
24596,
|
| 28 |
+
19650
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6967156699739387,
|
| 33 |
+
"test_acc": 0.6967930029154519,
|
| 34 |
+
"test_auc": 0.7856147150822794,
|
| 35 |
+
"val_auc": 0.8991596638655462,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "143_cancer_cat_Lung_Cancer",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
24596,
|
| 43 |
+
19650,
|
| 44 |
+
9044,
|
| 45 |
+
1001,
|
| 46 |
+
7614
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7171530846332638,
|
| 51 |
+
"test_acc": 0.717201166180758,
|
| 52 |
+
"test_auc": 0.8068815449476404,
|
| 53 |
+
"val_auc": 0.9079631852741096,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "143_cancer_cat_Lung_Cancer",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
24596,
|
| 61 |
+
19650,
|
| 62 |
+
9044,
|
| 63 |
+
1001,
|
| 64 |
+
7614,
|
| 65 |
+
32184,
|
| 66 |
+
24996,
|
| 67 |
+
4586,
|
| 68 |
+
11783,
|
| 69 |
+
27109
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7754720362552712,
|
| 74 |
+
"test_acc": 0.7755102040816326,
|
| 75 |
+
"test_auc": 0.8457092343261253,
|
| 76 |
+
"val_auc": 0.9483793517406963,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "143_cancer_cat_Lung_Cancer",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
24596,
|
| 84 |
+
19650,
|
| 85 |
+
9044,
|
| 86 |
+
1001,
|
| 87 |
+
7614,
|
| 88 |
+
32184,
|
| 89 |
+
24996,
|
| 90 |
+
4586,
|
| 91 |
+
11783,
|
| 92 |
+
27109,
|
| 93 |
+
27869,
|
| 94 |
+
23569,
|
| 95 |
+
2610,
|
| 96 |
+
19846,
|
| 97 |
+
8505,
|
| 98 |
+
16234
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6397086126945937,
|
| 4 |
+
"test_acc": 0.641399416909621,
|
| 5 |
+
"test_auc": 0.6883924928600572,
|
| 6 |
+
"val_auc": 0.633453381352541,
|
| 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 |
+
24596
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6086307744439726,
|
| 18 |
+
"test_acc": 0.60932944606414,
|
| 19 |
+
"test_auc": 0.6909424724602203,
|
| 20 |
+
"val_auc": 0.6330532212885155,
|
| 21 |
+
"k": 2,
|
| 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 |
+
24596,
|
| 28 |
+
3665
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6647173322617694,
|
| 33 |
+
"test_acc": 0.6647230320699709,
|
| 34 |
+
"test_auc": 0.7291921664626684,
|
| 35 |
+
"val_auc": 0.749499799919968,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "144_cancer_cat_Colon_Cancer",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
24596,
|
| 43 |
+
3665,
|
| 44 |
+
28148,
|
| 45 |
+
7614,
|
| 46 |
+
31510
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7198546620141681,
|
| 51 |
+
"test_acc": 0.7201166180758017,
|
| 52 |
+
"test_auc": 0.7967156262749898,
|
| 53 |
+
"val_auc": 0.8315326130452181,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "144_cancer_cat_Colon_Cancer",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
24596,
|
| 61 |
+
3665,
|
| 62 |
+
28148,
|
| 63 |
+
7614,
|
| 64 |
+
31510,
|
| 65 |
+
26616,
|
| 66 |
+
4365,
|
| 67 |
+
19846,
|
| 68 |
+
2302,
|
| 69 |
+
6677
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7317647455348778,
|
| 74 |
+
"test_acc": 0.7317784256559767,
|
| 75 |
+
"test_auc": 0.8446552427580579,
|
| 76 |
+
"val_auc": 0.8315326130452181,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "144_cancer_cat_Colon_Cancer",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
24596,
|
| 84 |
+
3665,
|
| 85 |
+
28148,
|
| 86 |
+
7614,
|
| 87 |
+
31510,
|
| 88 |
+
26616,
|
| 89 |
+
4365,
|
| 90 |
+
19846,
|
| 91 |
+
2302,
|
| 92 |
+
6677,
|
| 93 |
+
22469,
|
| 94 |
+
30796,
|
| 95 |
+
5999,
|
| 96 |
+
391,
|
| 97 |
+
13482,
|
| 98 |
+
16554
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.4597276380287314,
|
| 4 |
+
"test_acc": 0.47474747474747475,
|
| 5 |
+
"test_auc": 0.476938775510204,
|
| 6 |
+
"val_auc": 0.5632323232323233,
|
| 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 |
+
19207
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.4942291209860184,
|
| 18 |
+
"test_acc": 0.5050505050505051,
|
| 19 |
+
"test_auc": 0.47510204081632657,
|
| 20 |
+
"val_auc": 0.5652525252525252,
|
| 21 |
+
"k": 2,
|
| 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 |
+
19207,
|
| 28 |
+
3054
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.5249617745279682,
|
| 33 |
+
"test_acc": 0.5252525252525253,
|
| 34 |
+
"test_auc": 0.4773469387755102,
|
| 35 |
+
"val_auc": 0.545050505050505,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "145_disease_class_digestive system diseases",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
19207,
|
| 43 |
+
3054,
|
| 44 |
+
14189,
|
| 45 |
+
18132,
|
| 46 |
+
29206
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.48453298938141237,
|
| 51 |
+
"test_acc": 0.48484848484848486,
|
| 52 |
+
"test_auc": 0.4514285714285714,
|
| 53 |
+
"val_auc": 0.6307070707070707,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "145_disease_class_digestive system diseases",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
19207,
|
| 61 |
+
3054,
|
| 62 |
+
14189,
|
| 63 |
+
18132,
|
| 64 |
+
29206,
|
| 65 |
+
5834,
|
| 66 |
+
18537,
|
| 67 |
+
15996,
|
| 68 |
+
29116,
|
| 69 |
+
14598
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.5252525252525253,
|
| 74 |
+
"test_acc": 0.5252525252525253,
|
| 75 |
+
"test_auc": 0.48653061224489796,
|
| 76 |
+
"val_auc": 0.6012121212121213,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "145_disease_class_digestive system diseases",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
19207,
|
| 84 |
+
3054,
|
| 85 |
+
14189,
|
| 86 |
+
18132,
|
| 87 |
+
29206,
|
| 88 |
+
5834,
|
| 89 |
+
18537,
|
| 90 |
+
15996,
|
| 91 |
+
29116,
|
| 92 |
+
14598,
|
| 93 |
+
10916,
|
| 94 |
+
11056,
|
| 95 |
+
25699,
|
| 96 |
+
7761,
|
| 97 |
+
6154,
|
| 98 |
+
14574
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5278281150374173,
|
| 4 |
+
"test_acc": 0.5353535353535354,
|
| 5 |
+
"test_auc": 0.543061224489796,
|
| 6 |
+
"val_auc": 0.6262626262626262,
|
| 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 |
+
15178
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5225221299987655,
|
| 18 |
+
"test_acc": 0.5252525252525253,
|
| 19 |
+
"test_auc": 0.5926530612244898,
|
| 20 |
+
"val_auc": 0.6723232323232323,
|
| 21 |
+
"k": 2,
|
| 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 |
+
15178,
|
| 28 |
+
25544
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6227953173628877,
|
| 33 |
+
"test_acc": 0.6262626262626263,
|
| 34 |
+
"test_auc": 0.6477551020408163,
|
| 35 |
+
"val_auc": 0.7127272727272728,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "146_disease_class_cardiovascular diseases",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
15178,
|
| 43 |
+
25544,
|
| 44 |
+
18554,
|
| 45 |
+
30623,
|
| 46 |
+
18215
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.5820164327534701,
|
| 51 |
+
"test_acc": 0.5858585858585859,
|
| 52 |
+
"test_auc": 0.6473469387755102,
|
| 53 |
+
"val_auc": 0.7385858585858586,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "146_disease_class_cardiovascular diseases",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
15178,
|
| 61 |
+
25544,
|
| 62 |
+
18554,
|
| 63 |
+
30623,
|
| 64 |
+
18215,
|
| 65 |
+
787,
|
| 66 |
+
1550,
|
| 67 |
+
25704,
|
| 68 |
+
28201,
|
| 69 |
+
19667
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.5853506965412945,
|
| 74 |
+
"test_acc": 0.5858585858585859,
|
| 75 |
+
"test_auc": 0.6073469387755103,
|
| 76 |
+
"val_auc": 0.7587878787878788,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "146_disease_class_cardiovascular diseases",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
15178,
|
| 84 |
+
25544,
|
| 85 |
+
18554,
|
| 86 |
+
30623,
|
| 87 |
+
18215,
|
| 88 |
+
787,
|
| 89 |
+
1550,
|
| 90 |
+
25704,
|
| 91 |
+
28201,
|
| 92 |
+
19667,
|
| 93 |
+
28623,
|
| 94 |
+
4695,
|
| 95 |
+
11845,
|
| 96 |
+
28714,
|
| 97 |
+
28064,
|
| 98 |
+
22575
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5739304447261789,
|
| 4 |
+
"test_acc": 0.5757575757575758,
|
| 5 |
+
"test_auc": 0.5955102040816327,
|
| 6 |
+
"val_auc": 0.5896969696969697,
|
| 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 |
+
11366
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6043662430759206,
|
| 18 |
+
"test_acc": 0.6060606060606061,
|
| 19 |
+
"test_auc": 0.6720408163265307,
|
| 20 |
+
"val_auc": 0.6070707070707071,
|
| 21 |
+
"k": 2,
|
| 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 |
+
11366,
|
| 28 |
+
23286
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.5347841156664687,
|
| 33 |
+
"test_acc": 0.5353535353535354,
|
| 34 |
+
"test_auc": 0.5997959183673469,
|
| 35 |
+
"val_auc": 0.6430303030303031,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "147_disease_class_nervous system diseases",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
11366,
|
| 43 |
+
23286,
|
| 44 |
+
30623,
|
| 45 |
+
19805,
|
| 46 |
+
16570
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.5445254796302317,
|
| 51 |
+
"test_acc": 0.5454545454545454,
|
| 52 |
+
"test_auc": 0.6008163265306122,
|
| 53 |
+
"val_auc": 0.6715151515151515,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "147_disease_class_nervous system diseases",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
11366,
|
| 61 |
+
23286,
|
| 62 |
+
30623,
|
| 63 |
+
19805,
|
| 64 |
+
16570,
|
| 65 |
+
7761,
|
| 66 |
+
9280,
|
| 67 |
+
9891,
|
| 68 |
+
1736,
|
| 69 |
+
15541
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.5454545454545454,
|
| 74 |
+
"test_acc": 0.5454545454545454,
|
| 75 |
+
"test_auc": 0.6318367346938776,
|
| 76 |
+
"val_auc": 0.6674747474747474,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "147_disease_class_nervous system diseases",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
11366,
|
| 84 |
+
23286,
|
| 85 |
+
30623,
|
| 86 |
+
19805,
|
| 87 |
+
16570,
|
| 88 |
+
7761,
|
| 89 |
+
9280,
|
| 90 |
+
9891,
|
| 91 |
+
1736,
|
| 92 |
+
15541,
|
| 93 |
+
31114,
|
| 94 |
+
15958,
|
| 95 |
+
12179,
|
| 96 |
+
2394,
|
| 97 |
+
25388,
|
| 98 |
+
12475
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5281090662834479,
|
| 4 |
+
"test_acc": 0.5315254237288135,
|
| 5 |
+
"test_auc": 0.5406559221630208,
|
| 6 |
+
"val_auc": 0.6716988727858293,
|
| 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 |
+
18835
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5479447202592379,
|
| 18 |
+
"test_acc": 0.5484745762711865,
|
| 19 |
+
"test_auc": 0.5775841781484301,
|
| 20 |
+
"val_auc": 0.6517713365539453,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "148_twt_emotion_worry",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
18835,
|
| 28 |
+
2142
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.5810030828369629,
|
| 33 |
+
"test_acc": 0.5810169491525424,
|
| 34 |
+
"test_auc": 0.5980702547866728,
|
| 35 |
+
"val_auc": 0.6839774557165862,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "148_twt_emotion_worry",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
18835,
|
| 43 |
+
2142,
|
| 44 |
+
20632,
|
| 45 |
+
31360,
|
| 46 |
+
18015
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.5909999498914505,
|
| 51 |
+
"test_acc": 0.5911864406779661,
|
| 52 |
+
"test_auc": 0.6166102230900192,
|
| 53 |
+
"val_auc": 0.7028985507246377,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "148_twt_emotion_worry",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
18835,
|
| 61 |
+
2142,
|
| 62 |
+
20632,
|
| 63 |
+
31360,
|
| 64 |
+
18015,
|
| 65 |
+
16686,
|
| 66 |
+
19136,
|
| 67 |
+
14490,
|
| 68 |
+
25432,
|
| 69 |
+
2933
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.586388967271481,
|
| 74 |
+
"test_acc": 0.5864406779661017,
|
| 75 |
+
"test_auc": 0.614944494085375,
|
| 76 |
+
"val_auc": 0.7210144927536232,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "148_twt_emotion_worry",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
18835,
|
| 84 |
+
2142,
|
| 85 |
+
20632,
|
| 86 |
+
31360,
|
| 87 |
+
18015,
|
| 88 |
+
16686,
|
| 89 |
+
19136,
|
| 90 |
+
14490,
|
| 91 |
+
25432,
|
| 92 |
+
2933,
|
| 93 |
+
17670,
|
| 94 |
+
31272,
|
| 95 |
+
25978,
|
| 96 |
+
20595,
|
| 97 |
+
25079,
|
| 98 |
+
31708
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5707380242912606,
|
| 4 |
+
"test_acc": 0.5749152542372882,
|
| 5 |
+
"test_auc": 0.5954153842759594,
|
| 6 |
+
"val_auc": 0.6372785829307568,
|
| 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 |
+
21122
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.598212826068443,
|
| 18 |
+
"test_acc": 0.5993220338983051,
|
| 19 |
+
"test_auc": 0.6362597581199693,
|
| 20 |
+
"val_auc": 0.6521739130434783,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "149_twt_emotion_happiness",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
21122,
|
| 28 |
+
30616
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6037071296232759,
|
| 33 |
+
"test_acc": 0.6040677966101695,
|
| 34 |
+
"test_auc": 0.6576641919743486,
|
| 35 |
+
"val_auc": 0.713365539452496,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "149_twt_emotion_happiness",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
21122,
|
| 43 |
+
30616,
|
| 44 |
+
11534,
|
| 45 |
+
29268,
|
| 46 |
+
15272
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.6436333895592965,
|
| 51 |
+
"test_acc": 0.6440677966101694,
|
| 52 |
+
"test_auc": 0.6984460550168595,
|
| 53 |
+
"val_auc": 0.7797906602254429,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "149_twt_emotion_happiness",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
21122,
|
| 61 |
+
30616,
|
| 62 |
+
11534,
|
| 63 |
+
29268,
|
| 64 |
+
15272,
|
| 65 |
+
28313,
|
| 66 |
+
16958,
|
| 67 |
+
22292,
|
| 68 |
+
3354,
|
| 69 |
+
2945
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.6547952987702924,
|
| 74 |
+
"test_acc": 0.6549152542372881,
|
| 75 |
+
"test_auc": 0.7113618897383004,
|
| 76 |
+
"val_auc": 0.7930756843800322,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "149_twt_emotion_happiness",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
21122,
|
| 84 |
+
30616,
|
| 85 |
+
11534,
|
| 86 |
+
29268,
|
| 87 |
+
15272,
|
| 88 |
+
28313,
|
| 89 |
+
16958,
|
| 90 |
+
22292,
|
| 91 |
+
3354,
|
| 92 |
+
2945,
|
| 93 |
+
26124,
|
| 94 |
+
29998,
|
| 95 |
+
10025,
|
| 96 |
+
10655,
|
| 97 |
+
21325,
|
| 98 |
+
27405
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5404058364771427,
|
| 4 |
+
"test_acc": 0.5423728813559322,
|
| 5 |
+
"test_auc": 0.5534375425165378,
|
| 6 |
+
"val_auc": 0.5599838969404187,
|
| 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 |
+
15332
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5334199259692187,
|
| 18 |
+
"test_acc": 0.5335593220338983,
|
| 19 |
+
"test_auc": 0.5680448459844164,
|
| 20 |
+
"val_auc": 0.6698872785829307,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "150_twt_emotion_sadness",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
15332,
|
| 28 |
+
8438
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.5619791358317653,
|
| 33 |
+
"test_acc": 0.5620338983050848,
|
| 34 |
+
"test_auc": 0.5983092666747564,
|
| 35 |
+
"val_auc": 0.7286634460547504,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "150_twt_emotion_sadness",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
15332,
|
| 43 |
+
8438,
|
| 44 |
+
10447,
|
| 45 |
+
15468,
|
| 46 |
+
23677
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.572075539569831,
|
| 51 |
+
"test_acc": 0.5722033898305084,
|
| 52 |
+
"test_auc": 0.6151283493839009,
|
| 53 |
+
"val_auc": 0.748389694041868,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "150_twt_emotion_sadness",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
15332,
|
| 61 |
+
8438,
|
| 62 |
+
10447,
|
| 63 |
+
15468,
|
| 64 |
+
23677,
|
| 65 |
+
8227,
|
| 66 |
+
17596,
|
| 67 |
+
3389,
|
| 68 |
+
15856,
|
| 69 |
+
25769
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.5871015634449429,
|
| 74 |
+
"test_acc": 0.5871186440677966,
|
| 75 |
+
"test_auc": 0.6209069214165683,
|
| 76 |
+
"val_auc": 0.7987117552334944,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "150_twt_emotion_sadness",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
15332,
|
| 84 |
+
8438,
|
| 85 |
+
10447,
|
| 86 |
+
15468,
|
| 87 |
+
23677,
|
| 88 |
+
8227,
|
| 89 |
+
17596,
|
| 90 |
+
3389,
|
| 91 |
+
15856,
|
| 92 |
+
25769,
|
| 93 |
+
18353,
|
| 94 |
+
20855,
|
| 95 |
+
23461,
|
| 96 |
+
12643,
|
| 97 |
+
11534,
|
| 98 |
+
31483
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-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 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5834561046263635,
|
| 4 |
+
"test_acc": 0.5880893300248139,
|
| 5 |
+
"test_auc": 0.5885301216688833,
|
| 6 |
+
"val_auc": 0.6134453781512605,
|
| 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 |
+
28983
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5696708368310696,
|
| 18 |
+
"test_acc": 0.5732009925558312,
|
| 19 |
+
"test_auc": 0.5991207329688193,
|
| 20 |
+
"val_auc": 0.6634653861544618,
|
| 21 |
+
"k": 2,
|
| 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 |
+
28983,
|
| 28 |
+
30087
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.568711646410621,
|
| 33 |
+
"test_acc": 0.5707196029776674,
|
| 34 |
+
"test_auc": 0.6353381606817399,
|
| 35 |
+
"val_auc": 0.6798719487795118,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "151_it_tick_HR Support",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
28983,
|
| 43 |
+
30087,
|
| 44 |
+
19295,
|
| 45 |
+
32398,
|
| 46 |
+
4815
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.5813165238936575,
|
| 51 |
+
"test_acc": 0.5831265508684863,
|
| 52 |
+
"test_auc": 0.6513472242746663,
|
| 53 |
+
"val_auc": 0.673469387755102,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "151_it_tick_HR Support",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
28983,
|
| 61 |
+
30087,
|
| 62 |
+
19295,
|
| 63 |
+
32398,
|
| 64 |
+
4815,
|
| 65 |
+
32480,
|
| 66 |
+
9157,
|
| 67 |
+
21986,
|
| 68 |
+
23672,
|
| 69 |
+
14517
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.6149530840030368,
|
| 74 |
+
"test_acc": 0.6153846153846154,
|
| 75 |
+
"test_auc": 0.6511748189744347,
|
| 76 |
+
"val_auc": 0.671468587434974,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "151_it_tick_HR Support",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
28983,
|
| 84 |
+
30087,
|
| 85 |
+
19295,
|
| 86 |
+
32398,
|
| 87 |
+
4815,
|
| 88 |
+
32480,
|
| 89 |
+
9157,
|
| 90 |
+
21986,
|
| 91 |
+
23672,
|
| 92 |
+
14517,
|
| 93 |
+
22184,
|
| 94 |
+
2906,
|
| 95 |
+
10811,
|
| 96 |
+
28464,
|
| 97 |
+
8263,
|
| 98 |
+
1256
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5481255225854489,
|
| 4 |
+
"test_acc": 0.5583126550868487,
|
| 5 |
+
"test_auc": 0.5638392197428699,
|
| 6 |
+
"val_auc": 0.5964385754301721,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "152_it_tick_Hardware",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
10440
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.557167313507873,
|
| 18 |
+
"test_acc": 0.5583126550868487,
|
| 19 |
+
"test_auc": 0.6063001822570318,
|
| 20 |
+
"val_auc": 0.6346538615446179,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "152_it_tick_Hardware",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
10440,
|
| 28 |
+
4809
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.5530471331913929,
|
| 33 |
+
"test_acc": 0.5533498759305211,
|
| 34 |
+
"test_auc": 0.6081227525737649,
|
| 35 |
+
"val_auc": 0.6706682673069229,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "152_it_tick_Hardware",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
10440,
|
| 43 |
+
4809,
|
| 44 |
+
7453,
|
| 45 |
+
28142,
|
| 46 |
+
23121
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.5754470989050432,
|
| 51 |
+
"test_acc": 0.575682382133995,
|
| 52 |
+
"test_auc": 0.6351411260529038,
|
| 53 |
+
"val_auc": 0.6770708283313325,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "152_it_tick_Hardware",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
10440,
|
| 61 |
+
4809,
|
| 62 |
+
7453,
|
| 63 |
+
28142,
|
| 64 |
+
23121,
|
| 65 |
+
22840,
|
| 66 |
+
18103,
|
| 67 |
+
18671,
|
| 68 |
+
2872,
|
| 69 |
+
20616
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.5726483622620706,
|
| 74 |
+
"test_acc": 0.5732009925558312,
|
| 75 |
+
"test_auc": 0.6266440076843505,
|
| 76 |
+
"val_auc": 0.6934773909563825,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "152_it_tick_Hardware",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
10440,
|
| 84 |
+
4809,
|
| 85 |
+
7453,
|
| 86 |
+
28142,
|
| 87 |
+
23121,
|
| 88 |
+
22840,
|
| 89 |
+
18103,
|
| 90 |
+
18671,
|
| 91 |
+
2872,
|
| 92 |
+
20616,
|
| 93 |
+
30486,
|
| 94 |
+
8255,
|
| 95 |
+
12578,
|
| 96 |
+
27450,
|
| 97 |
+
12098,
|
| 98 |
+
26863
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6322788233417818,
|
| 4 |
+
"test_acc": 0.6327543424317618,
|
| 5 |
+
"test_auc": 0.6643638244421457,
|
| 6 |
+
"val_auc": 0.7410964385754302,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "153_it_tick_Administrative rights",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
7651
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.650098215463241,
|
| 18 |
+
"test_acc": 0.6501240694789082,
|
| 19 |
+
"test_auc": 0.7164671691049702,
|
| 20 |
+
"val_auc": 0.748499399759904,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "153_it_tick_Administrative rights",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
7651,
|
| 28 |
+
2786
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.6650082816119337,
|
| 33 |
+
"test_acc": 0.6650124069478908,
|
| 34 |
+
"test_auc": 0.7341756563716073,
|
| 35 |
+
"val_auc": 0.7951180472188875,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "153_it_tick_Administrative rights",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
7651,
|
| 43 |
+
2786,
|
| 44 |
+
13150,
|
| 45 |
+
7840,
|
| 46 |
+
22364
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.6648968548507462,
|
| 51 |
+
"test_acc": 0.6650124069478908,
|
| 52 |
+
"test_auc": 0.7612186591793508,
|
| 53 |
+
"val_auc": 0.8263305322128851,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "153_it_tick_Administrative rights",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
7651,
|
| 61 |
+
2786,
|
| 62 |
+
13150,
|
| 63 |
+
7840,
|
| 64 |
+
22364,
|
| 65 |
+
23784,
|
| 66 |
+
23624,
|
| 67 |
+
25785,
|
| 68 |
+
28418,
|
| 69 |
+
11091
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.7021965762251632,
|
| 74 |
+
"test_acc": 0.7022332506203474,
|
| 75 |
+
"test_auc": 0.7781143786020392,
|
| 76 |
+
"val_auc": 0.7915166066426571,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "153_it_tick_Administrative rights",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
7651,
|
| 84 |
+
2786,
|
| 85 |
+
13150,
|
| 86 |
+
7840,
|
| 87 |
+
22364,
|
| 88 |
+
23784,
|
| 89 |
+
23624,
|
| 90 |
+
25785,
|
| 91 |
+
28418,
|
| 92 |
+
11091,
|
| 93 |
+
26732,
|
| 94 |
+
6276,
|
| 95 |
+
6284,
|
| 96 |
+
16438,
|
| 97 |
+
19130,
|
| 98 |
+
10239
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6537595172719612,
|
| 4 |
+
"test_acc": 0.6572958500669344,
|
| 5 |
+
"test_auc": 0.7112586199481011,
|
| 6 |
+
"val_auc": 0.7294949494949495,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "154_athlete_sport_football",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
29327
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.69051455582175,
|
| 18 |
+
"test_acc": 0.6921017402945113,
|
| 19 |
+
"test_auc": 0.7399893908331063,
|
| 20 |
+
"val_auc": 0.7773737373737374,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "154_athlete_sport_football",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
29327,
|
| 28 |
+
9338
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7616571405886635,
|
| 33 |
+
"test_acc": 0.7617135207496654,
|
| 34 |
+
"test_auc": 0.863708047196456,
|
| 35 |
+
"val_auc": 0.9248484848484848,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "154_athlete_sport_football",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
29327,
|
| 43 |
+
9338,
|
| 44 |
+
8705,
|
| 45 |
+
8216,
|
| 46 |
+
26926
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.8045368272622117,
|
| 51 |
+
"test_acc": 0.8045515394912985,
|
| 52 |
+
"test_auc": 0.9013777580249746,
|
| 53 |
+
"val_auc": 0.9478787878787879,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "154_athlete_sport_football",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
29327,
|
| 61 |
+
9338,
|
| 62 |
+
8705,
|
| 63 |
+
8216,
|
| 64 |
+
26926,
|
| 65 |
+
26295,
|
| 66 |
+
696,
|
| 67 |
+
9330,
|
| 68 |
+
19357,
|
| 69 |
+
181
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.8125836680053548,
|
| 74 |
+
"test_acc": 0.8125836680053548,
|
| 75 |
+
"test_auc": 0.9133202391363564,
|
| 76 |
+
"val_auc": 0.9632323232323232,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "154_athlete_sport_football",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
29327,
|
| 84 |
+
9338,
|
| 85 |
+
8705,
|
| 86 |
+
8216,
|
| 87 |
+
26926,
|
| 88 |
+
26295,
|
| 89 |
+
696,
|
| 90 |
+
9330,
|
| 91 |
+
19357,
|
| 92 |
+
181,
|
| 93 |
+
12804,
|
| 94 |
+
8207,
|
| 95 |
+
3290,
|
| 96 |
+
20403,
|
| 97 |
+
19429,
|
| 98 |
+
26652
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|
matryoshka-prefix-panel/gemma-2-2b/n-1/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6427472058949092,
|
| 4 |
+
"test_acc": 0.6452476572958501,
|
| 5 |
+
"test_auc": 0.7073805393471062,
|
| 6 |
+
"val_auc": 0.7765656565656566,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "155_athlete_sport_basketball",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
3682
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7192179613546497,
|
| 18 |
+
"test_acc": 0.7202141900937081,
|
| 19 |
+
"test_auc": 0.7668277157316741,
|
| 20 |
+
"val_auc": 0.8274747474747475,
|
| 21 |
+
"k": 2,
|
| 22 |
+
"dataset": "155_athlete_sport_basketball",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
3682,
|
| 28 |
+
30938
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"test_f1": 0.7401520207392482,
|
| 33 |
+
"test_acc": 0.7402945113788487,
|
| 34 |
+
"test_auc": 0.819529469111554,
|
| 35 |
+
"val_auc": 0.9006060606060606,
|
| 36 |
+
"k": 5,
|
| 37 |
+
"dataset": "155_athlete_sport_basketball",
|
| 38 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 39 |
+
"reg_type": "l1",
|
| 40 |
+
"binarize": false,
|
| 41 |
+
"indices": [
|
| 42 |
+
3682,
|
| 43 |
+
30938,
|
| 44 |
+
14284,
|
| 45 |
+
2528,
|
| 46 |
+
22985
|
| 47 |
+
]
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"test_f1": 0.7705951424657312,
|
| 51 |
+
"test_acc": 0.7710843373493976,
|
| 52 |
+
"test_auc": 0.8497225846224427,
|
| 53 |
+
"val_auc": 0.8941414141414141,
|
| 54 |
+
"k": 10,
|
| 55 |
+
"dataset": "155_athlete_sport_basketball",
|
| 56 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 57 |
+
"reg_type": "l1",
|
| 58 |
+
"binarize": false,
|
| 59 |
+
"indices": [
|
| 60 |
+
3682,
|
| 61 |
+
30938,
|
| 62 |
+
14284,
|
| 63 |
+
2528,
|
| 64 |
+
22985,
|
| 65 |
+
28935,
|
| 66 |
+
17217,
|
| 67 |
+
10493,
|
| 68 |
+
28358,
|
| 69 |
+
24292
|
| 70 |
+
]
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"test_f1": 0.8339634013089565,
|
| 74 |
+
"test_acc": 0.8340026773761714,
|
| 75 |
+
"test_auc": 0.9083955785580136,
|
| 76 |
+
"val_auc": 0.9325252525252525,
|
| 77 |
+
"k": 16,
|
| 78 |
+
"dataset": "155_athlete_sport_basketball",
|
| 79 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 80 |
+
"reg_type": "l1",
|
| 81 |
+
"binarize": false,
|
| 82 |
+
"indices": [
|
| 83 |
+
3682,
|
| 84 |
+
30938,
|
| 85 |
+
14284,
|
| 86 |
+
2528,
|
| 87 |
+
22985,
|
| 88 |
+
28935,
|
| 89 |
+
17217,
|
| 90 |
+
10493,
|
| 91 |
+
28358,
|
| 92 |
+
24292,
|
| 93 |
+
6694,
|
| 94 |
+
29170,
|
| 95 |
+
27111,
|
| 96 |
+
16891,
|
| 97 |
+
3463,
|
| 98 |
+
30529
|
| 99 |
+
]
|
| 100 |
+
}
|
| 101 |
+
]
|