Add files using upload-large-folder tool
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +6 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/binarized_probing_binarized_linear_results.json +0 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/cfg.json +1 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/binarized_probing_binarized_linear_results.json +0 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json +45 -0
- gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json +45 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
gemma-2-2b-btk-vnorm_topp-p0.9/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.2/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
gemma-2-2b-btk-vnorm_topp-p0.95/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.1/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
gemma-2-2b-btk-vnorm_topp-p0.99/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.3/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 39 |
+
gemma-2-2b-btk/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.1/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 40 |
+
gemma-2-2b-btk-vnorm_threshold_renorm-t0.01/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.2/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
gemma-2-2b-btk-vnorm_threshold_renorm-t0.05/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.4/saebench_autointerp_custom_sae_eval_results.json filter=lfs diff=lfs merge=lfs -text
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/binarized_probing_binarized_linear_results.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/cfg.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"d_sae": 32768, "device": "cuda", "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}, "d_in": 2304, "normalize_activations": "none", "apply_b_dec_to_input": true, "reshape_activations": "none", "architecture": "jumprelu"}
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/binarized_probing_binarized_linear_results.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.93121837219097,
|
| 4 |
+
"test_acc": 0.9312649164677804,
|
| 5 |
+
"test_auc": 0.955455244719555,
|
| 6 |
+
"val_auc": 0.9755469755469756,
|
| 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 |
+
15241
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9914079383726543,
|
| 18 |
+
"test_acc": 0.9914081145584726,
|
| 19 |
+
"test_auc": 0.9992308084895412,
|
| 20 |
+
"val_auc": 0.9987129987129987,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "100_news_fake",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
15241,
|
| 28 |
+
2096,
|
| 29 |
+
13467,
|
| 30 |
+
13729,
|
| 31 |
+
15362,
|
| 32 |
+
22093,
|
| 33 |
+
24729,
|
| 34 |
+
3247,
|
| 35 |
+
29588,
|
| 36 |
+
1944,
|
| 37 |
+
17721,
|
| 38 |
+
21044,
|
| 39 |
+
15133,
|
| 40 |
+
31708,
|
| 41 |
+
7470,
|
| 42 |
+
26936
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9169039976007368,
|
| 4 |
+
"test_acc": 0.9170447180816591,
|
| 5 |
+
"test_auc": 0.9544221218658226,
|
| 6 |
+
"val_auc": 0.9429516329061596,
|
| 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 |
+
13467
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9675940957759462,
|
| 18 |
+
"test_acc": 0.9675955930006481,
|
| 19 |
+
"test_auc": 0.9923035825890606,
|
| 20 |
+
"val_auc": 0.9875981810665565,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "105_click_bait",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
13467,
|
| 28 |
+
8731,
|
| 29 |
+
5640,
|
| 30 |
+
7737,
|
| 31 |
+
31172,
|
| 32 |
+
24729,
|
| 33 |
+
12247,
|
| 34 |
+
23439,
|
| 35 |
+
5608,
|
| 36 |
+
6230,
|
| 37 |
+
24343,
|
| 38 |
+
29054,
|
| 39 |
+
31511,
|
| 40 |
+
7470,
|
| 41 |
+
15884,
|
| 42 |
+
685
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5533157644411559,
|
| 4 |
+
"test_acc": 0.5733333333333334,
|
| 5 |
+
"test_auc": 0.5832134224572396,
|
| 6 |
+
"val_auc": 0.5874349739895958,
|
| 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 |
+
27644
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5868814782148504,
|
| 18 |
+
"test_acc": 0.5869182389937107,
|
| 19 |
+
"test_auc": 0.6290941421047676,
|
| 20 |
+
"val_auc": 0.6138455382152861,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "106_hate_hate",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
27644,
|
| 28 |
+
23418,
|
| 29 |
+
21889,
|
| 30 |
+
15996,
|
| 31 |
+
17673,
|
| 32 |
+
15157,
|
| 33 |
+
14214,
|
| 34 |
+
24021,
|
| 35 |
+
21451,
|
| 36 |
+
32669,
|
| 37 |
+
12004,
|
| 38 |
+
25217,
|
| 39 |
+
16204,
|
| 40 |
+
12891,
|
| 41 |
+
20530,
|
| 42 |
+
4985
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6877817082284876,
|
| 4 |
+
"test_acc": 0.6991194968553459,
|
| 5 |
+
"test_auc": 0.7040638395040602,
|
| 6 |
+
"val_auc": 0.7611044417767108,
|
| 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 |
+
9503
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7240197573778583,
|
| 18 |
+
"test_acc": 0.7255345911949685,
|
| 19 |
+
"test_auc": 0.7796097166795437,
|
| 20 |
+
"val_auc": 0.8235294117647058,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "107_hate_offensive",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
9503,
|
| 28 |
+
15241,
|
| 29 |
+
26759,
|
| 30 |
+
32706,
|
| 31 |
+
26611,
|
| 32 |
+
3639,
|
| 33 |
+
21661,
|
| 34 |
+
25715,
|
| 35 |
+
15996,
|
| 36 |
+
1944,
|
| 37 |
+
3733,
|
| 38 |
+
298,
|
| 39 |
+
21782,
|
| 40 |
+
16674,
|
| 41 |
+
2055,
|
| 42 |
+
5595
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7773904209042783,
|
| 4 |
+
"test_acc": 0.7781132075471698,
|
| 5 |
+
"test_auc": 0.8204201302429575,
|
| 6 |
+
"val_auc": 0.865546218487395,
|
| 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 |
+
24729
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7790196395230139,
|
| 18 |
+
"test_acc": 0.7791194968553459,
|
| 19 |
+
"test_auc": 0.8672613182871765,
|
| 20 |
+
"val_auc": 0.9151660664265707,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "110_aimade_humangpt3",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
24729,
|
| 28 |
+
18492,
|
| 29 |
+
31511,
|
| 30 |
+
12104,
|
| 31 |
+
21997,
|
| 32 |
+
13729,
|
| 33 |
+
11196,
|
| 34 |
+
13301,
|
| 35 |
+
16154,
|
| 36 |
+
1944,
|
| 37 |
+
31172,
|
| 38 |
+
25007,
|
| 39 |
+
16696,
|
| 40 |
+
27334,
|
| 41 |
+
6158,
|
| 42 |
+
30851
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7617440877364794,
|
| 4 |
+
"test_acc": 0.7617610062893082,
|
| 5 |
+
"test_auc": 0.815240967698491,
|
| 6 |
+
"val_auc": 0.7971188475390156,
|
| 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 |
+
20530
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8830115231008825,
|
| 18 |
+
"test_acc": 0.8830188679245283,
|
| 19 |
+
"test_auc": 0.9475104274362836,
|
| 20 |
+
"val_auc": 0.9647859143657462,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "113_movie_sent",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
20530,
|
| 28 |
+
12591,
|
| 29 |
+
23889,
|
| 30 |
+
11910,
|
| 31 |
+
15869,
|
| 32 |
+
17920,
|
| 33 |
+
1944,
|
| 34 |
+
27644,
|
| 35 |
+
8265,
|
| 36 |
+
16003,
|
| 37 |
+
6046,
|
| 38 |
+
29208,
|
| 39 |
+
20265,
|
| 40 |
+
13582,
|
| 41 |
+
29251,
|
| 42 |
+
19069
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5190074888840321,
|
| 4 |
+
"test_acc": 0.5497435897435897,
|
| 5 |
+
"test_auc": 0.5556455549197159,
|
| 6 |
+
"val_auc": 0.5552835577315383,
|
| 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 |
+
2354
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5910541228879087,
|
| 18 |
+
"test_acc": 0.5917948717948718,
|
| 19 |
+
"test_auc": 0.6438171138115596,
|
| 20 |
+
"val_auc": 0.6136270909832721,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "114_nyc_borough_Manhattan",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
2354,
|
| 28 |
+
30706,
|
| 29 |
+
14742,
|
| 30 |
+
7841,
|
| 31 |
+
13986,
|
| 32 |
+
30057,
|
| 33 |
+
9884,
|
| 34 |
+
12087,
|
| 35 |
+
22553,
|
| 36 |
+
31569,
|
| 37 |
+
13255,
|
| 38 |
+
31808,
|
| 39 |
+
15521,
|
| 40 |
+
8081,
|
| 41 |
+
17891,
|
| 42 |
+
24057
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5082113178258992,
|
| 4 |
+
"test_acc": 0.5230769230769231,
|
| 5 |
+
"test_auc": 0.5324565758912041,
|
| 6 |
+
"val_auc": 0.6005711954304365,
|
| 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 |
+
14742
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5432264763125402,
|
| 18 |
+
"test_acc": 0.5435897435897435,
|
| 19 |
+
"test_auc": 0.5561420708923823,
|
| 20 |
+
"val_auc": 0.6752345981232151,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "115_nyc_borough_Brooklyn",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
14742,
|
| 28 |
+
7401,
|
| 29 |
+
25113,
|
| 30 |
+
32593,
|
| 31 |
+
7347,
|
| 32 |
+
14232,
|
| 33 |
+
9300,
|
| 34 |
+
27814,
|
| 35 |
+
913,
|
| 36 |
+
27879,
|
| 37 |
+
25889,
|
| 38 |
+
1220,
|
| 39 |
+
30945,
|
| 40 |
+
18849,
|
| 41 |
+
1566,
|
| 42 |
+
25582
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5341901640457114,
|
| 4 |
+
"test_acc": 0.5558974358974359,
|
| 5 |
+
"test_auc": 0.5563882249974753,
|
| 6 |
+
"val_auc": 0.5091799265605874,
|
| 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 |
+
23987
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5536497755891315,
|
| 18 |
+
"test_acc": 0.556923076923077,
|
| 19 |
+
"test_auc": 0.5698930386777527,
|
| 20 |
+
"val_auc": 0.5393716850265198,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "116_nyc_borough_Bronx",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
23987,
|
| 28 |
+
30057,
|
| 29 |
+
28553,
|
| 30 |
+
25902,
|
| 31 |
+
14232,
|
| 32 |
+
14742,
|
| 33 |
+
17891,
|
| 34 |
+
31289,
|
| 35 |
+
26936,
|
| 36 |
+
27279,
|
| 37 |
+
31991,
|
| 38 |
+
27879,
|
| 39 |
+
13986,
|
| 40 |
+
26307,
|
| 41 |
+
16288,
|
| 42 |
+
26950
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8671979042043984,
|
| 4 |
+
"test_acc": 0.8686868686868687,
|
| 5 |
+
"test_auc": 0.8771428571428571,
|
| 6 |
+
"val_auc": 0.8587740384615383,
|
| 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 |
+
5563
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.83818577936225,
|
| 18 |
+
"test_acc": 0.8383838383838383,
|
| 19 |
+
"test_auc": 0.9130612244897959,
|
| 20 |
+
"val_auc": 0.9443108974358974,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "117_us_state_FL",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
5563,
|
| 28 |
+
13862,
|
| 29 |
+
4118,
|
| 30 |
+
31545,
|
| 31 |
+
20145,
|
| 32 |
+
18966,
|
| 33 |
+
5019,
|
| 34 |
+
5730,
|
| 35 |
+
16630,
|
| 36 |
+
913,
|
| 37 |
+
24525,
|
| 38 |
+
12601,
|
| 39 |
+
14401,
|
| 40 |
+
2405,
|
| 41 |
+
3334,
|
| 42 |
+
26632
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8681220810253067,
|
| 4 |
+
"test_acc": 0.8686868686868687,
|
| 5 |
+
"test_auc": 0.8953061224489796,
|
| 6 |
+
"val_auc": 0.9403044871794872,
|
| 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 |
+
5230
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8787631416202846,
|
| 18 |
+
"test_acc": 0.8787878787878788,
|
| 19 |
+
"test_auc": 0.9485714285714285,
|
| 20 |
+
"val_auc": 0.9855769230769231,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "118_us_state_CA",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
5230,
|
| 28 |
+
27653,
|
| 29 |
+
30766,
|
| 30 |
+
16726,
|
| 31 |
+
28240,
|
| 32 |
+
6446,
|
| 33 |
+
31405,
|
| 34 |
+
31545,
|
| 35 |
+
1220,
|
| 36 |
+
2405,
|
| 37 |
+
547,
|
| 38 |
+
15387,
|
| 39 |
+
30019,
|
| 40 |
+
29386,
|
| 41 |
+
18823,
|
| 42 |
+
22469
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8677158770616714,
|
| 4 |
+
"test_acc": 0.8686868686868687,
|
| 5 |
+
"test_auc": 0.8842857142857143,
|
| 6 |
+
"val_auc": 0.8120993589743589,
|
| 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 |
+
13064
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8888208408469712,
|
| 18 |
+
"test_acc": 0.8888888888888888,
|
| 19 |
+
"test_auc": 0.9630612244897959,
|
| 20 |
+
"val_auc": 0.9435096153846153,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "119_us_state_TX",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
13064,
|
| 28 |
+
31545,
|
| 29 |
+
1220,
|
| 30 |
+
14742,
|
| 31 |
+
30766,
|
| 32 |
+
22788,
|
| 33 |
+
2143,
|
| 34 |
+
18738,
|
| 35 |
+
28240,
|
| 36 |
+
547,
|
| 37 |
+
5730,
|
| 38 |
+
4080,
|
| 39 |
+
30019,
|
| 40 |
+
7816,
|
| 41 |
+
5787,
|
| 42 |
+
32558
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6284219674688128,
|
| 4 |
+
"test_acc": 0.6380580840918942,
|
| 5 |
+
"test_auc": 0.6705486102864804,
|
| 6 |
+
"val_auc": 0.7806,
|
| 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 |
+
5230
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.798240286256853,
|
| 18 |
+
"test_acc": 0.7984395318595578,
|
| 19 |
+
"test_auc": 0.8933420614747752,
|
| 20 |
+
"val_auc": 0.9312,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "120_us_timezone_Chicago",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
5230,
|
| 28 |
+
31490,
|
| 29 |
+
31545,
|
| 30 |
+
18966,
|
| 31 |
+
27653,
|
| 32 |
+
21803,
|
| 33 |
+
13064,
|
| 34 |
+
1220,
|
| 35 |
+
4080,
|
| 36 |
+
2405,
|
| 37 |
+
6446,
|
| 38 |
+
5162,
|
| 39 |
+
547,
|
| 40 |
+
24779,
|
| 41 |
+
28240,
|
| 42 |
+
15140
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6462295730872706,
|
| 4 |
+
"test_acc": 0.6532293021239706,
|
| 5 |
+
"test_auc": 0.6868864434727582,
|
| 6 |
+
"val_auc": 0.741,
|
| 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 |
+
5230
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8131330469593859,
|
| 18 |
+
"test_acc": 0.8131772865192891,
|
| 19 |
+
"test_auc": 0.9138897698867094,
|
| 20 |
+
"val_auc": 0.9463999999999999,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "121_us_timezone_New_York",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
5230,
|
| 28 |
+
1220,
|
| 29 |
+
547,
|
| 30 |
+
5162,
|
| 31 |
+
31490,
|
| 32 |
+
25479,
|
| 33 |
+
31545,
|
| 34 |
+
30766,
|
| 35 |
+
11342,
|
| 36 |
+
30019,
|
| 37 |
+
13064,
|
| 38 |
+
16726,
|
| 39 |
+
28817,
|
| 40 |
+
18966,
|
| 41 |
+
26106,
|
| 42 |
+
5787
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8229420277015342,
|
| 4 |
+
"test_acc": 0.824013870827915,
|
| 5 |
+
"test_auc": 0.8654478333215588,
|
| 6 |
+
"val_auc": 0.8602,
|
| 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 |
+
5230
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8573605894842493,
|
| 18 |
+
"test_acc": 0.857824013870828,
|
| 19 |
+
"test_auc": 0.9381193811336864,
|
| 20 |
+
"val_auc": 0.9259999999999999,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "122_us_timezone_Los_Angeles",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
5230,
|
| 28 |
+
27653,
|
| 29 |
+
6446,
|
| 30 |
+
1220,
|
| 31 |
+
2405,
|
| 32 |
+
31545,
|
| 33 |
+
18966,
|
| 34 |
+
31490,
|
| 35 |
+
547,
|
| 36 |
+
29456,
|
| 37 |
+
18823,
|
| 38 |
+
22469,
|
| 39 |
+
31405,
|
| 40 |
+
26106,
|
| 41 |
+
21803,
|
| 42 |
+
15387
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8501694101411182,
|
| 4 |
+
"test_acc": 0.8502564102564103,
|
| 5 |
+
"test_auc": 0.9382005924529573,
|
| 6 |
+
"val_auc": 0.952264381884945,
|
| 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 |
+
4761
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9692305750247127,
|
| 18 |
+
"test_acc": 0.9692307692307692,
|
| 19 |
+
"test_auc": 0.992358703335914,
|
| 20 |
+
"val_auc": 0.9942880456956343,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "123_world_country_United_Kingdom",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
4761,
|
| 28 |
+
29647,
|
| 29 |
+
17749,
|
| 30 |
+
24616,
|
| 31 |
+
14742,
|
| 32 |
+
17891,
|
| 33 |
+
16922,
|
| 34 |
+
22788,
|
| 35 |
+
20104,
|
| 36 |
+
3183,
|
| 37 |
+
14232,
|
| 38 |
+
3399,
|
| 39 |
+
5456,
|
| 40 |
+
1440,
|
| 41 |
+
18823,
|
| 42 |
+
13986
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9007755255132825,
|
| 4 |
+
"test_acc": 0.9015384615384615,
|
| 5 |
+
"test_auc": 0.9044795839364459,
|
| 6 |
+
"val_auc": 0.9302325581395349,
|
| 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 |
+
18659
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9610233449818711,
|
| 18 |
+
"test_acc": 0.961025641025641,
|
| 19 |
+
"test_auc": 0.9936252398424614,
|
| 20 |
+
"val_auc": 1.0,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "124_world_country_United_States",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
18659,
|
| 28 |
+
29522,
|
| 29 |
+
29868,
|
| 30 |
+
4761,
|
| 31 |
+
17233,
|
| 32 |
+
13901,
|
| 33 |
+
917,
|
| 34 |
+
16116,
|
| 35 |
+
5456,
|
| 36 |
+
5757,
|
| 37 |
+
3399,
|
| 38 |
+
11807,
|
| 39 |
+
2143,
|
| 40 |
+
14232,
|
| 41 |
+
23070,
|
| 42 |
+
27115
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9589579532770056,
|
| 4 |
+
"test_acc": 0.958974358974359,
|
| 5 |
+
"test_auc": 0.981570000336621,
|
| 6 |
+
"val_auc": 0.9995920032639738,
|
| 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 |
+
9122
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9784615384615385,
|
| 18 |
+
"test_acc": 0.9784615384615385,
|
| 19 |
+
"test_auc": 0.9976941461608375,
|
| 20 |
+
"val_auc": 1.0,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "125_world_country_Italy",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
9122,
|
| 28 |
+
1440,
|
| 29 |
+
31139,
|
| 30 |
+
14639,
|
| 31 |
+
30766,
|
| 32 |
+
16597,
|
| 33 |
+
29868,
|
| 34 |
+
28273,
|
| 35 |
+
7147,
|
| 36 |
+
22304,
|
| 37 |
+
14638,
|
| 38 |
+
17749,
|
| 39 |
+
14717,
|
| 40 |
+
17233,
|
| 41 |
+
11807,
|
| 42 |
+
1706
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6194097198734697,
|
| 4 |
+
"test_acc": 0.6194191590810576,
|
| 5 |
+
"test_auc": 0.6620954153207441,
|
| 6 |
+
"val_auc": 0.6876,
|
| 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 |
+
12247
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7056066890035801,
|
| 18 |
+
"test_acc": 0.7069787602947551,
|
| 19 |
+
"test_auc": 0.7886081219815386,
|
| 20 |
+
"val_auc": 0.8029999999999999,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "126_art_type_book",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
12247,
|
| 28 |
+
25878,
|
| 29 |
+
23439,
|
| 30 |
+
7995,
|
| 31 |
+
15827,
|
| 32 |
+
24487,
|
| 33 |
+
11302,
|
| 34 |
+
6230,
|
| 35 |
+
24343,
|
| 36 |
+
30018,
|
| 37 |
+
25582,
|
| 38 |
+
8943,
|
| 39 |
+
2029,
|
| 40 |
+
1944,
|
| 41 |
+
13255,
|
| 42 |
+
20252
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6542916134157861,
|
| 4 |
+
"test_acc": 0.6666666666666666,
|
| 5 |
+
"test_auc": 0.6795425541989024,
|
| 6 |
+
"val_auc": 0.6714,
|
| 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 |
+
7995
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7533041828469449,
|
| 18 |
+
"test_acc": 0.753359341135674,
|
| 19 |
+
"test_auc": 0.8327319583754835,
|
| 20 |
+
"val_auc": 0.8496,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "127_art_type_song",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
7995,
|
| 28 |
+
12247,
|
| 29 |
+
15827,
|
| 30 |
+
7841,
|
| 31 |
+
24343,
|
| 32 |
+
26478,
|
| 33 |
+
30057,
|
| 34 |
+
14589,
|
| 35 |
+
1257,
|
| 36 |
+
25878,
|
| 37 |
+
22192,
|
| 38 |
+
298,
|
| 39 |
+
13255,
|
| 40 |
+
1944,
|
| 41 |
+
26759,
|
| 42 |
+
26936
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5207261228832126,
|
| 4 |
+
"test_acc": 0.5574338968357174,
|
| 5 |
+
"test_auc": 0.5650048626069285,
|
| 6 |
+
"val_auc": 0.5660000000000001,
|
| 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 |
+
14639
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6397715526290674,
|
| 18 |
+
"test_acc": 0.6406588643259644,
|
| 19 |
+
"test_auc": 0.7022074131081453,
|
| 20 |
+
"val_auc": 0.7304,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "128_art_type_movie",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
14639,
|
| 28 |
+
12460,
|
| 29 |
+
11807,
|
| 30 |
+
15827,
|
| 31 |
+
22192,
|
| 32 |
+
7995,
|
| 33 |
+
1440,
|
| 34 |
+
18634,
|
| 35 |
+
31086,
|
| 36 |
+
13860,
|
| 37 |
+
14589,
|
| 38 |
+
25901,
|
| 39 |
+
15387,
|
| 40 |
+
21236,
|
| 41 |
+
24750,
|
| 42 |
+
26478
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6095264034364154,
|
| 4 |
+
"test_acc": 0.6204690831556503,
|
| 5 |
+
"test_auc": 0.6398254228041461,
|
| 6 |
+
"val_auc": 0.5784,
|
| 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 |
+
1185
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7840098463055911,
|
| 18 |
+
"test_acc": 0.7846481876332623,
|
| 19 |
+
"test_auc": 0.8584106201127477,
|
| 20 |
+
"val_auc": 0.8552,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "129_arith_mc_A",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
1185,
|
| 28 |
+
14542,
|
| 29 |
+
30074,
|
| 30 |
+
17809,
|
| 31 |
+
20449,
|
| 32 |
+
18271,
|
| 33 |
+
3095,
|
| 34 |
+
5725,
|
| 35 |
+
4049,
|
| 36 |
+
19539,
|
| 37 |
+
12744,
|
| 38 |
+
23811,
|
| 39 |
+
12127,
|
| 40 |
+
11220,
|
| 41 |
+
5806,
|
| 42 |
+
16618
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9696969696969697,
|
| 4 |
+
"test_acc": 0.9696969696969697,
|
| 5 |
+
"test_auc": 0.9779591836734693,
|
| 6 |
+
"val_auc": 0.9601449275362319,
|
| 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 |
+
8845
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 1.0,
|
| 18 |
+
"test_acc": 1.0,
|
| 19 |
+
"test_auc": 1.0,
|
| 20 |
+
"val_auc": 1.0,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "130_temp_cat_Frequency",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
8845,
|
| 28 |
+
19764,
|
| 29 |
+
10969,
|
| 30 |
+
32236,
|
| 31 |
+
2739,
|
| 32 |
+
22584,
|
| 33 |
+
2468,
|
| 34 |
+
10348,
|
| 35 |
+
31889,
|
| 36 |
+
21218,
|
| 37 |
+
25240,
|
| 38 |
+
26911,
|
| 39 |
+
27607,
|
| 40 |
+
30061,
|
| 41 |
+
992,
|
| 42 |
+
11039
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8482990353199857,
|
| 4 |
+
"test_acc": 0.8484848484848485,
|
| 5 |
+
"test_auc": 0.9122448979591837,
|
| 6 |
+
"val_auc": 0.874194847020934,
|
| 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 |
+
6330
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9696598070639972,
|
| 18 |
+
"test_acc": 0.9696969696969697,
|
| 19 |
+
"test_auc": 0.9967346938775511,
|
| 20 |
+
"val_auc": 0.9842995169082126,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "131_temp_cat_Typical Time",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
6330,
|
| 28 |
+
14285,
|
| 29 |
+
15453,
|
| 30 |
+
10969,
|
| 31 |
+
6714,
|
| 32 |
+
16835,
|
| 33 |
+
29275,
|
| 34 |
+
24776,
|
| 35 |
+
22739,
|
| 36 |
+
23688,
|
| 37 |
+
1300,
|
| 38 |
+
19764,
|
| 39 |
+
5209,
|
| 40 |
+
13597,
|
| 41 |
+
7673,
|
| 42 |
+
27902
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9189439402205359,
|
| 4 |
+
"test_acc": 0.9191919191919192,
|
| 5 |
+
"test_auc": 0.9273469387755101,
|
| 6 |
+
"val_auc": 0.9074074074074074,
|
| 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 |
+
14619
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9797856011091305,
|
| 18 |
+
"test_acc": 0.9797979797979798,
|
| 19 |
+
"test_auc": 0.9987755102040816,
|
| 20 |
+
"val_auc": 1.0,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "132_temp_cat_Event Ordering",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
14619,
|
| 28 |
+
6330,
|
| 29 |
+
22584,
|
| 30 |
+
12148,
|
| 31 |
+
20180,
|
| 32 |
+
15587,
|
| 33 |
+
9869,
|
| 34 |
+
13597,
|
| 35 |
+
11535,
|
| 36 |
+
22212,
|
| 37 |
+
11039,
|
| 38 |
+
18480,
|
| 39 |
+
3372,
|
| 40 |
+
393,
|
| 41 |
+
23209,
|
| 42 |
+
5835
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9466645596291614,
|
| 4 |
+
"test_acc": 0.9466666666666667,
|
| 5 |
+
"test_auc": 0.9630214917825538,
|
| 6 |
+
"val_auc": 0.9492694805194806,
|
| 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 |
+
8498
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9377777777777778,
|
| 18 |
+
"test_acc": 0.9377777777777778,
|
| 19 |
+
"test_auc": 0.9700537294563844,
|
| 20 |
+
"val_auc": 0.9784902597402598,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "133_context_type_Causality",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
8498,
|
| 28 |
+
4560,
|
| 29 |
+
12873,
|
| 30 |
+
16863,
|
| 31 |
+
4994,
|
| 32 |
+
24976,
|
| 33 |
+
27381,
|
| 34 |
+
13597,
|
| 35 |
+
13999,
|
| 36 |
+
18973,
|
| 37 |
+
4451,
|
| 38 |
+
6029,
|
| 39 |
+
2696,
|
| 40 |
+
4654,
|
| 41 |
+
31741,
|
| 42 |
+
30356
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6827403980843597,
|
| 4 |
+
"test_acc": 0.6844444444444444,
|
| 5 |
+
"test_auc": 0.7470764854614412,
|
| 6 |
+
"val_auc": 0.5669642857142857,
|
| 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 |
+
1474
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7986401860797997,
|
| 18 |
+
"test_acc": 0.8,
|
| 19 |
+
"test_auc": 0.9027338811630847,
|
| 20 |
+
"val_auc": 0.854301948051948,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "134_context_type_Belief_states",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
1474,
|
| 28 |
+
18973,
|
| 29 |
+
992,
|
| 30 |
+
2893,
|
| 31 |
+
4771,
|
| 32 |
+
1206,
|
| 33 |
+
15587,
|
| 34 |
+
21218,
|
| 35 |
+
26821,
|
| 36 |
+
16165,
|
| 37 |
+
4654,
|
| 38 |
+
18220,
|
| 39 |
+
6475,
|
| 40 |
+
6375,
|
| 41 |
+
12479,
|
| 42 |
+
2972
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8610700735578785,
|
| 4 |
+
"test_acc": 0.8622222222222222,
|
| 5 |
+
"test_auc": 0.875869152970923,
|
| 6 |
+
"val_auc": 0.8942775974025974,
|
| 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 |
+
11535
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9555450187029134,
|
| 18 |
+
"test_acc": 0.9555555555555556,
|
| 19 |
+
"test_auc": 0.9943900126422249,
|
| 20 |
+
"val_auc": 0.9866071428571428,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "135_context_type_Event_duration",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
11535,
|
| 28 |
+
13229,
|
| 29 |
+
22584,
|
| 30 |
+
15575,
|
| 31 |
+
23209,
|
| 32 |
+
22468,
|
| 33 |
+
7516,
|
| 34 |
+
14500,
|
| 35 |
+
27719,
|
| 36 |
+
24487,
|
| 37 |
+
12148,
|
| 38 |
+
6514,
|
| 39 |
+
16858,
|
| 40 |
+
6330,
|
| 41 |
+
2690,
|
| 42 |
+
19764
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7229756787564391,
|
| 4 |
+
"test_acc": 0.7234503684438665,
|
| 5 |
+
"test_auc": 0.7764553624708959,
|
| 6 |
+
"val_auc": 0.8342,
|
| 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 |
+
21276
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7527822692245206,
|
| 18 |
+
"test_acc": 0.7537928045080191,
|
| 19 |
+
"test_auc": 0.8245786366963734,
|
| 20 |
+
"val_auc": 0.896,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "136_glue_mnli_entailment",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
21276,
|
| 28 |
+
5773,
|
| 29 |
+
17840,
|
| 30 |
+
9664,
|
| 31 |
+
24934,
|
| 32 |
+
895,
|
| 33 |
+
13467,
|
| 34 |
+
11534,
|
| 35 |
+
24704,
|
| 36 |
+
23600,
|
| 37 |
+
17201,
|
| 38 |
+
8731,
|
| 39 |
+
30910,
|
| 40 |
+
685,
|
| 41 |
+
13729,
|
| 42 |
+
9313
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6237943949814779,
|
| 4 |
+
"test_acc": 0.6250541829215431,
|
| 5 |
+
"test_auc": 0.661661763976425,
|
| 6 |
+
"val_auc": 0.7132000000000001,
|
| 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 |
+
21276
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6846480360047881,
|
| 18 |
+
"test_acc": 0.6848721283051582,
|
| 19 |
+
"test_auc": 0.7547213132495892,
|
| 20 |
+
"val_auc": 0.7964000000000001,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "137_glue_mnli_neutral",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
21276,
|
| 28 |
+
5826,
|
| 29 |
+
31511,
|
| 30 |
+
27404,
|
| 31 |
+
5773,
|
| 32 |
+
17840,
|
| 33 |
+
9313,
|
| 34 |
+
27656,
|
| 35 |
+
13467,
|
| 36 |
+
24724,
|
| 37 |
+
1102,
|
| 38 |
+
25065,
|
| 39 |
+
16696,
|
| 40 |
+
24934,
|
| 41 |
+
30851,
|
| 42 |
+
30998
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5889330715645289,
|
| 4 |
+
"test_acc": 0.5916775032509753,
|
| 5 |
+
"test_auc": 0.6190534526012317,
|
| 6 |
+
"val_auc": 0.6464,
|
| 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 |
+
21276
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.742481853932304,
|
| 18 |
+
"test_acc": 0.7438231469440832,
|
| 19 |
+
"test_auc": 0.8279155725174776,
|
| 20 |
+
"val_auc": 0.8304,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "138_glue_mnli_contradiction",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
21276,
|
| 28 |
+
8847,
|
| 29 |
+
30998,
|
| 30 |
+
5826,
|
| 31 |
+
27404,
|
| 32 |
+
25065,
|
| 33 |
+
5773,
|
| 34 |
+
27656,
|
| 35 |
+
2489,
|
| 36 |
+
25543,
|
| 37 |
+
11910,
|
| 38 |
+
9664,
|
| 39 |
+
26629,
|
| 40 |
+
8304,
|
| 41 |
+
24934,
|
| 42 |
+
31876
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6703384661277905,
|
| 4 |
+
"test_acc": 0.6705084745762712,
|
| 5 |
+
"test_auc": 0.7103598783613344,
|
| 6 |
+
"val_auc": 0.7586553945249599,
|
| 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 |
+
19527
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7941827407699165,
|
| 18 |
+
"test_acc": 0.7945762711864407,
|
| 19 |
+
"test_auc": 0.8693919537567153,
|
| 20 |
+
"val_auc": 0.8756038647342995,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "139_news_class_Politics",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
19527,
|
| 28 |
+
2198,
|
| 29 |
+
4712,
|
| 30 |
+
1440,
|
| 31 |
+
4103,
|
| 32 |
+
23180,
|
| 33 |
+
23395,
|
| 34 |
+
7737,
|
| 35 |
+
20158,
|
| 36 |
+
14892,
|
| 37 |
+
8017,
|
| 38 |
+
27505,
|
| 39 |
+
30057,
|
| 40 |
+
2371,
|
| 41 |
+
2813,
|
| 42 |
+
8683
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5575765765777074,
|
| 4 |
+
"test_acc": 0.560677966101695,
|
| 5 |
+
"test_auc": 0.5767485558166301,
|
| 6 |
+
"val_auc": 0.6384863123993558,
|
| 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 |
+
3976
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7737203913946582,
|
| 18 |
+
"test_acc": 0.7742372881355932,
|
| 19 |
+
"test_auc": 0.86295334120234,
|
| 20 |
+
"val_auc": 0.8679549114331724,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "140_news_class_Technology",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
3976,
|
| 28 |
+
17032,
|
| 29 |
+
24551,
|
| 30 |
+
13347,
|
| 31 |
+
8934,
|
| 32 |
+
8731,
|
| 33 |
+
17721,
|
| 34 |
+
21474,
|
| 35 |
+
7023,
|
| 36 |
+
13630,
|
| 37 |
+
22764,
|
| 38 |
+
32031,
|
| 39 |
+
22727,
|
| 40 |
+
12845,
|
| 41 |
+
2813,
|
| 42 |
+
2198
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6374248670055985,
|
| 4 |
+
"test_acc": 0.648135593220339,
|
| 5 |
+
"test_auc": 0.6528683265123019,
|
| 6 |
+
"val_auc": 0.623792270531401,
|
| 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 |
+
19527
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7530852820493633,
|
| 18 |
+
"test_acc": 0.7532203389830509,
|
| 19 |
+
"test_auc": 0.8500273944394804,
|
| 20 |
+
"val_auc": 0.8927133655394525,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "141_news_class_Entertainment",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
19527,
|
| 28 |
+
2198,
|
| 29 |
+
7078,
|
| 30 |
+
20568,
|
| 31 |
+
22745,
|
| 32 |
+
1944,
|
| 33 |
+
2371,
|
| 34 |
+
10825,
|
| 35 |
+
5615,
|
| 36 |
+
10209,
|
| 37 |
+
21044,
|
| 38 |
+
20759,
|
| 39 |
+
2813,
|
| 40 |
+
23180,
|
| 41 |
+
7023,
|
| 42 |
+
1440
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5736120110067442,
|
| 4 |
+
"test_acc": 0.597667638483965,
|
| 5 |
+
"test_auc": 0.596015231878145,
|
| 6 |
+
"val_auc": 0.606642657062825,
|
| 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 |
+
27447
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6588863293445828,
|
| 18 |
+
"test_acc": 0.6588921282798834,
|
| 19 |
+
"test_auc": 0.6975724194206445,
|
| 20 |
+
"val_auc": 0.7907162865146059,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "142_cancer_cat_Thyroid_Cancer",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
27447,
|
| 28 |
+
2082,
|
| 29 |
+
11750,
|
| 30 |
+
23447,
|
| 31 |
+
24929,
|
| 32 |
+
28876,
|
| 33 |
+
8731,
|
| 34 |
+
26490,
|
| 35 |
+
3178,
|
| 36 |
+
26213,
|
| 37 |
+
3675,
|
| 38 |
+
1000,
|
| 39 |
+
6859,
|
| 40 |
+
2585,
|
| 41 |
+
23348,
|
| 42 |
+
23418
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6682811759468037,
|
| 4 |
+
"test_acc": 0.6793002915451894,
|
| 5 |
+
"test_auc": 0.6903984768121855,
|
| 6 |
+
"val_auc": 0.7302921168467387,
|
| 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 |
+
15241
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.879902870898198,
|
| 18 |
+
"test_acc": 0.880466472303207,
|
| 19 |
+
"test_auc": 0.9444954440364477,
|
| 20 |
+
"val_auc": 0.9299719887955182,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "143_cancer_cat_Lung_Cancer",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
15241,
|
| 28 |
+
11750,
|
| 29 |
+
32713,
|
| 30 |
+
19552,
|
| 31 |
+
28876,
|
| 32 |
+
1000,
|
| 33 |
+
4985,
|
| 34 |
+
8683,
|
| 35 |
+
13373,
|
| 36 |
+
26213,
|
| 37 |
+
27447,
|
| 38 |
+
13763,
|
| 39 |
+
2082,
|
| 40 |
+
13729,
|
| 41 |
+
23447,
|
| 42 |
+
2585
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.727891728368474,
|
| 4 |
+
"test_acc": 0.7288629737609329,
|
| 5 |
+
"test_auc": 0.7420270637834897,
|
| 6 |
+
"val_auc": 0.617046818727491,
|
| 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 |
+
15241
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8309037900874635,
|
| 18 |
+
"test_acc": 0.8309037900874635,
|
| 19 |
+
"test_auc": 0.9048687610499115,
|
| 20 |
+
"val_auc": 0.8819527811124449,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "144_cancer_cat_Colon_Cancer",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
15241,
|
| 28 |
+
32713,
|
| 29 |
+
21044,
|
| 30 |
+
3675,
|
| 31 |
+
13373,
|
| 32 |
+
29386,
|
| 33 |
+
19552,
|
| 34 |
+
24444,
|
| 35 |
+
1000,
|
| 36 |
+
4985,
|
| 37 |
+
13629,
|
| 38 |
+
1706,
|
| 39 |
+
3178,
|
| 40 |
+
6275,
|
| 41 |
+
28876,
|
| 42 |
+
15133
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.49121989121989124,
|
| 4 |
+
"test_acc": 0.5151515151515151,
|
| 5 |
+
"test_auc": 0.5191836734693878,
|
| 6 |
+
"val_auc": 0.5252525252525253,
|
| 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 |
+
23418
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5428403372328606,
|
| 18 |
+
"test_acc": 0.5454545454545454,
|
| 19 |
+
"test_auc": 0.5489795918367347,
|
| 20 |
+
"val_auc": 0.5886868686868686,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "145_disease_class_digestive system diseases",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
23418,
|
| 28 |
+
30091,
|
| 29 |
+
11423,
|
| 30 |
+
28653,
|
| 31 |
+
30061,
|
| 32 |
+
22019,
|
| 33 |
+
17605,
|
| 34 |
+
15631,
|
| 35 |
+
4672,
|
| 36 |
+
27334,
|
| 37 |
+
19500,
|
| 38 |
+
16547,
|
| 39 |
+
26418,
|
| 40 |
+
9673,
|
| 41 |
+
2816,
|
| 42 |
+
24841
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.45455653966811144,
|
| 4 |
+
"test_acc": 0.5454545454545454,
|
| 5 |
+
"test_auc": 0.55,
|
| 6 |
+
"val_auc": 0.5692929292929293,
|
| 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 |
+
7084
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.469884025439581,
|
| 18 |
+
"test_acc": 0.47474747474747475,
|
| 19 |
+
"test_auc": 0.5338775510204081,
|
| 20 |
+
"val_auc": 0.6824242424242424,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "146_disease_class_cardiovascular diseases",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
7084,
|
| 28 |
+
23418,
|
| 29 |
+
26364,
|
| 30 |
+
25567,
|
| 31 |
+
18452,
|
| 32 |
+
22917,
|
| 33 |
+
2084,
|
| 34 |
+
1169,
|
| 35 |
+
22016,
|
| 36 |
+
8212,
|
| 37 |
+
30073,
|
| 38 |
+
21710,
|
| 39 |
+
7767,
|
| 40 |
+
25525,
|
| 41 |
+
8879,
|
| 42 |
+
22041
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5757575757575758,
|
| 4 |
+
"test_acc": 0.5757575757575758,
|
| 5 |
+
"test_auc": 0.5846938775510204,
|
| 6 |
+
"val_auc": 0.6163636363636363,
|
| 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 |
+
23887
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5352587095444239,
|
| 18 |
+
"test_acc": 0.5353535353535354,
|
| 19 |
+
"test_auc": 0.5812244897959185,
|
| 20 |
+
"val_auc": 0.6763636363636363,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "147_disease_class_nervous system diseases",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
23887,
|
| 28 |
+
24915,
|
| 29 |
+
11416,
|
| 30 |
+
256,
|
| 31 |
+
6230,
|
| 32 |
+
307,
|
| 33 |
+
30073,
|
| 34 |
+
31511,
|
| 35 |
+
29386,
|
| 36 |
+
11423,
|
| 37 |
+
9673,
|
| 38 |
+
22016,
|
| 39 |
+
5043,
|
| 40 |
+
17907,
|
| 41 |
+
21533,
|
| 42 |
+
28566
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5193154056523528,
|
| 4 |
+
"test_acc": 0.5193220338983051,
|
| 5 |
+
"test_auc": 0.5230039749515542,
|
| 6 |
+
"val_auc": 0.6247987117552335,
|
| 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 |
+
1944
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5606557526634622,
|
| 18 |
+
"test_acc": 0.560677966101695,
|
| 19 |
+
"test_auc": 0.5976997863601431,
|
| 20 |
+
"val_auc": 0.6634460547504026,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "148_twt_emotion_worry",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
1944,
|
| 28 |
+
10660,
|
| 29 |
+
3814,
|
| 30 |
+
4627,
|
| 31 |
+
14127,
|
| 32 |
+
15241,
|
| 33 |
+
26967,
|
| 34 |
+
11026,
|
| 35 |
+
24487,
|
| 36 |
+
24343,
|
| 37 |
+
13174,
|
| 38 |
+
25525,
|
| 39 |
+
16706,
|
| 40 |
+
32510,
|
| 41 |
+
9175,
|
| 42 |
+
11206
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5803108967827775,
|
| 4 |
+
"test_acc": 0.5810169491525424,
|
| 5 |
+
"test_auc": 0.6028192371475953,
|
| 6 |
+
"val_auc": 0.6115136876006441,
|
| 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 |
+
1944
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6535561372642272,
|
| 18 |
+
"test_acc": 0.6535593220338983,
|
| 19 |
+
"test_auc": 0.7255803392497968,
|
| 20 |
+
"val_auc": 0.7570450885668277,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "149_twt_emotion_happiness",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
1944,
|
| 28 |
+
27398,
|
| 29 |
+
4985,
|
| 30 |
+
26967,
|
| 31 |
+
22602,
|
| 32 |
+
29988,
|
| 33 |
+
23889,
|
| 34 |
+
14127,
|
| 35 |
+
5168,
|
| 36 |
+
32510,
|
| 37 |
+
21889,
|
| 38 |
+
29777,
|
| 39 |
+
14097,
|
| 40 |
+
9313,
|
| 41 |
+
4627,
|
| 42 |
+
15884
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.538948868276155,
|
| 4 |
+
"test_acc": 0.5769491525423729,
|
| 5 |
+
"test_auc": 0.5803557967737071,
|
| 6 |
+
"val_auc": 0.6209742351046699,
|
| 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 |
+
29988
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6016850582801904,
|
| 18 |
+
"test_acc": 0.6033898305084746,
|
| 19 |
+
"test_auc": 0.6367423782785997,
|
| 20 |
+
"val_auc": 0.748792270531401,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "150_twt_emotion_sadness",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
29988,
|
| 28 |
+
799,
|
| 29 |
+
4627,
|
| 30 |
+
26967,
|
| 31 |
+
11910,
|
| 32 |
+
14097,
|
| 33 |
+
16594,
|
| 34 |
+
12591,
|
| 35 |
+
11026,
|
| 36 |
+
5034,
|
| 37 |
+
22739,
|
| 38 |
+
9207,
|
| 39 |
+
29930,
|
| 40 |
+
6230,
|
| 41 |
+
14108,
|
| 42 |
+
21316
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.4799620918073501,
|
| 4 |
+
"test_acc": 0.5012406947890818,
|
| 5 |
+
"test_auc": 0.4961331954090931,
|
| 6 |
+
"val_auc": 0.562625050020008,
|
| 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 |
+
7664
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.593907960212401,
|
| 18 |
+
"test_acc": 0.5955334987593052,
|
| 19 |
+
"test_auc": 0.6501773311659524,
|
| 20 |
+
"val_auc": 0.6250500200080031,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "151_it_tick_HR Support",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
7664,
|
| 28 |
+
29853,
|
| 29 |
+
24444,
|
| 30 |
+
13345,
|
| 31 |
+
15772,
|
| 32 |
+
16696,
|
| 33 |
+
739,
|
| 34 |
+
15811,
|
| 35 |
+
2354,
|
| 36 |
+
31338,
|
| 37 |
+
8594,
|
| 38 |
+
1613,
|
| 39 |
+
30057,
|
| 40 |
+
29434,
|
| 41 |
+
3150,
|
| 42 |
+
19666
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5642756535207794,
|
| 4 |
+
"test_acc": 0.56575682382134,
|
| 5 |
+
"test_auc": 0.58856706566179,
|
| 6 |
+
"val_auc": 0.6094437775110044,
|
| 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 |
+
6605
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.632401221607177,
|
| 18 |
+
"test_acc": 0.6327543424317618,
|
| 19 |
+
"test_auc": 0.6810748239003005,
|
| 20 |
+
"val_auc": 0.693077230892357,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "152_it_tick_Hardware",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
6605,
|
| 28 |
+
31753,
|
| 29 |
+
1999,
|
| 30 |
+
8642,
|
| 31 |
+
13614,
|
| 32 |
+
16696,
|
| 33 |
+
7664,
|
| 34 |
+
15811,
|
| 35 |
+
6337,
|
| 36 |
+
31666,
|
| 37 |
+
19666,
|
| 38 |
+
17126,
|
| 39 |
+
17891,
|
| 40 |
+
28112,
|
| 41 |
+
23439,
|
| 42 |
+
29853
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
gemma-2-2b-btk-softmax_topp-p0.8/k-150/seed-0/model.layers.12/w-32768/t-500M/l0-142.6/denoised/sae_probes/sae_probes_gemma-2-2b--softmax_topp-p0.8/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5707143164360337,
|
| 4 |
+
"test_acc": 0.5707196029776674,
|
| 5 |
+
"test_auc": 0.5989236983399833,
|
| 6 |
+
"val_auc": 0.6126450580232093,
|
| 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 |
+
31666
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.630145417576009,
|
| 18 |
+
"test_acc": 0.630272952853598,
|
| 19 |
+
"test_auc": 0.715531254617999,
|
| 20 |
+
"val_auc": 0.7434973989595839,
|
| 21 |
+
"k": 16,
|
| 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 |
+
31666,
|
| 28 |
+
7664,
|
| 29 |
+
17934,
|
| 30 |
+
7992,
|
| 31 |
+
17891,
|
| 32 |
+
28156,
|
| 33 |
+
29853,
|
| 34 |
+
8642,
|
| 35 |
+
22775,
|
| 36 |
+
23811,
|
| 37 |
+
19285,
|
| 38 |
+
18525,
|
| 39 |
+
18775,
|
| 40 |
+
13119,
|
| 41 |
+
28793,
|
| 42 |
+
3675
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|