Add files using upload-large-folder tool
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/binarized_probing_binarized_linear_results.json +0 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/cfg.json +1 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/binarized_probing_binarized_linear_results.json +0 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json +45 -0
- denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json +45 -0
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9604424609164776,
|
| 4 |
+
"test_acc": 0.9604597701149425,
|
| 5 |
+
"test_auc": 0.9801886601006548,
|
| 6 |
+
"val_auc": 0.9573863636363638,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "68_credit-card",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
1739
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.976089163965159,
|
| 18 |
+
"test_acc": 0.9760919540229885,
|
| 19 |
+
"test_auc": 0.9933860733805941,
|
| 20 |
+
"val_auc": 0.9772727272727272,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "68_credit-card",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
1739,
|
| 28 |
+
3001,
|
| 29 |
+
18497,
|
| 30 |
+
17343,
|
| 31 |
+
820,
|
| 32 |
+
29728,
|
| 33 |
+
16664,
|
| 34 |
+
31872,
|
| 35 |
+
15686,
|
| 36 |
+
752,
|
| 37 |
+
11200,
|
| 38 |
+
6400,
|
| 39 |
+
14099,
|
| 40 |
+
10355,
|
| 41 |
+
32032,
|
| 42 |
+
20636
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9926435910776503,
|
| 4 |
+
"test_acc": 0.9926436781609196,
|
| 5 |
+
"test_auc": 0.9985414186373722,
|
| 6 |
+
"val_auc": 0.9971590909090909,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "69_blood-pressure",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
32545
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9972413793103448,
|
| 18 |
+
"test_acc": 0.9972413793103448,
|
| 19 |
+
"test_auc": 0.9999332012013638,
|
| 20 |
+
"val_auc": 0.9991883116883117,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "69_blood-pressure",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
32545,
|
| 28 |
+
9296,
|
| 29 |
+
3001,
|
| 30 |
+
7740,
|
| 31 |
+
31143,
|
| 32 |
+
23941,
|
| 33 |
+
23956,
|
| 34 |
+
7771,
|
| 35 |
+
8258,
|
| 36 |
+
32032,
|
| 37 |
+
4021,
|
| 38 |
+
31296,
|
| 39 |
+
10275,
|
| 40 |
+
21973,
|
| 41 |
+
22415,
|
| 42 |
+
15914
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9576755436677145,
|
| 4 |
+
"test_acc": 0.9577011494252874,
|
| 5 |
+
"test_auc": 0.9843183478543212,
|
| 6 |
+
"val_auc": 0.9882305194805194,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "74_magnetic-field",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
175
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9622899066339847,
|
| 18 |
+
"test_acc": 0.9622988505747126,
|
| 19 |
+
"test_auc": 0.9884894677742302,
|
| 20 |
+
"val_auc": 0.9967532467532467,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "74_magnetic-field",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
175,
|
| 28 |
+
8258,
|
| 29 |
+
3001,
|
| 30 |
+
281,
|
| 31 |
+
10275,
|
| 32 |
+
30123,
|
| 33 |
+
3324,
|
| 34 |
+
14602,
|
| 35 |
+
7740,
|
| 36 |
+
31393,
|
| 37 |
+
204,
|
| 38 |
+
29809,
|
| 39 |
+
17343,
|
| 40 |
+
29499,
|
| 41 |
+
32173,
|
| 42 |
+
1553
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9857446795610127,
|
| 4 |
+
"test_acc": 0.9857471264367816,
|
| 5 |
+
"test_auc": 0.9929159451269007,
|
| 6 |
+
"val_auc": 0.9910714285714286,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "75_cell-lines",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
22134
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9816042957503781,
|
| 18 |
+
"test_acc": 0.9816091954022989,
|
| 19 |
+
"test_auc": 0.9957012013637101,
|
| 20 |
+
"val_auc": 0.9987824675324676,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "75_cell-lines",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
22134,
|
| 28 |
+
9615,
|
| 29 |
+
23391,
|
| 30 |
+
7771,
|
| 31 |
+
7094,
|
| 32 |
+
3001,
|
| 33 |
+
17297,
|
| 34 |
+
9246,
|
| 35 |
+
1025,
|
| 36 |
+
16664,
|
| 37 |
+
27741,
|
| 38 |
+
23581,
|
| 39 |
+
32032,
|
| 40 |
+
23220,
|
| 41 |
+
29499,
|
| 42 |
+
16033
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8580924325531809,
|
| 4 |
+
"test_acc": 0.8593103448275862,
|
| 5 |
+
"test_auc": 0.9151143696899182,
|
| 6 |
+
"val_auc": 0.9370941558441559,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "77_second-derivative",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
32724
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9839075969448875,
|
| 18 |
+
"test_acc": 0.9839080459770115,
|
| 19 |
+
"test_auc": 0.9939906447859733,
|
| 20 |
+
"val_auc": 0.9955357142857143,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "77_second-derivative",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
32724,
|
| 28 |
+
3001,
|
| 29 |
+
32607,
|
| 30 |
+
27885,
|
| 31 |
+
30257,
|
| 32 |
+
27751,
|
| 33 |
+
32280,
|
| 34 |
+
24452,
|
| 35 |
+
4296,
|
| 36 |
+
26165,
|
| 37 |
+
17812,
|
| 38 |
+
14786,
|
| 39 |
+
29544,
|
| 40 |
+
7614,
|
| 41 |
+
14838,
|
| 42 |
+
32458
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9834469461994598,
|
| 4 |
+
"test_acc": 0.983448275862069,
|
| 5 |
+
"test_auc": 0.9826238568104335,
|
| 6 |
+
"val_auc": 0.9886363636363636,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "78_north-america",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
21738
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9829882468290856,
|
| 18 |
+
"test_acc": 0.9829885057471265,
|
| 19 |
+
"test_auc": 0.998640348503707,
|
| 20 |
+
"val_auc": 1.0,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "78_north-america",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
21738,
|
| 28 |
+
13401,
|
| 29 |
+
16838,
|
| 30 |
+
24263,
|
| 31 |
+
1164,
|
| 32 |
+
29499,
|
| 33 |
+
14002,
|
| 34 |
+
3001,
|
| 35 |
+
19125,
|
| 36 |
+
17235,
|
| 37 |
+
1984,
|
| 38 |
+
32759,
|
| 39 |
+
3491,
|
| 40 |
+
7771,
|
| 41 |
+
17343,
|
| 42 |
+
6489
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9572401861678439,
|
| 4 |
+
"test_acc": 0.9572413793103448,
|
| 5 |
+
"test_auc": 0.9908942245251365,
|
| 6 |
+
"val_auc": 0.9967532467532468,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "79_human-rights",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
4769
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9742528409072686,
|
| 18 |
+
"test_acc": 0.9742528735632184,
|
| 19 |
+
"test_auc": 0.995685135829861,
|
| 20 |
+
"val_auc": 0.9975649350649352,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "79_human-rights",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
4769,
|
| 28 |
+
3001,
|
| 29 |
+
26013,
|
| 30 |
+
18494,
|
| 31 |
+
3071,
|
| 32 |
+
25581,
|
| 33 |
+
16664,
|
| 34 |
+
29544,
|
| 35 |
+
5331,
|
| 36 |
+
14749,
|
| 37 |
+
21973,
|
| 38 |
+
9511,
|
| 39 |
+
5481,
|
| 40 |
+
4029,
|
| 41 |
+
16046,
|
| 42 |
+
26916
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5499822359872555,
|
| 4 |
+
"test_acc": 0.5733333333333334,
|
| 5 |
+
"test_auc": 0.5755668890038772,
|
| 6 |
+
"val_auc": 0.6266506602641057,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "7_hist_fig_ispolitician",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
7771
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.611211948034679,
|
| 18 |
+
"test_acc": 0.6113207547169811,
|
| 19 |
+
"test_auc": 0.6506370887630767,
|
| 20 |
+
"val_auc": 0.6526610644257703,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "7_hist_fig_ispolitician",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
7771,
|
| 28 |
+
34,
|
| 29 |
+
27910,
|
| 30 |
+
9246,
|
| 31 |
+
25759,
|
| 32 |
+
17503,
|
| 33 |
+
28989,
|
| 34 |
+
29228,
|
| 35 |
+
20434,
|
| 36 |
+
16033,
|
| 37 |
+
29544,
|
| 38 |
+
7106,
|
| 39 |
+
12774,
|
| 40 |
+
16210,
|
| 41 |
+
14602,
|
| 42 |
+
20651
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9664314169905743,
|
| 4 |
+
"test_acc": 0.9664367816091954,
|
| 5 |
+
"test_auc": 0.990767391363169,
|
| 6 |
+
"val_auc": 1.0,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "80_side-effects",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
3878
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9553711095424321,
|
| 18 |
+
"test_acc": 0.9554022988505747,
|
| 19 |
+
"test_auc": 0.990767391363169,
|
| 20 |
+
"val_auc": 1.0,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "80_side-effects",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
3878,
|
| 28 |
+
3001,
|
| 29 |
+
17343,
|
| 30 |
+
29499,
|
| 31 |
+
29728,
|
| 32 |
+
27262,
|
| 33 |
+
9246,
|
| 34 |
+
29022,
|
| 35 |
+
22957,
|
| 36 |
+
2619,
|
| 37 |
+
17469,
|
| 38 |
+
8522,
|
| 39 |
+
19209,
|
| 40 |
+
16033,
|
| 41 |
+
9284,
|
| 42 |
+
5616
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.984827522061186,
|
| 4 |
+
"test_acc": 0.9848275862068966,
|
| 5 |
+
"test_auc": 0.9983532827804535,
|
| 6 |
+
"val_auc": 1.0,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "81_public-health",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
11049
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9875861544136975,
|
| 18 |
+
"test_acc": 0.9875862068965517,
|
| 19 |
+
"test_auc": 0.9988711848584879,
|
| 20 |
+
"val_auc": 1.0,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "81_public-health",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
11049,
|
| 28 |
+
3001,
|
| 29 |
+
22293,
|
| 30 |
+
26013,
|
| 31 |
+
7771,
|
| 32 |
+
5336,
|
| 33 |
+
25343,
|
| 34 |
+
9614,
|
| 35 |
+
28463,
|
| 36 |
+
6038,
|
| 37 |
+
8522,
|
| 38 |
+
19451,
|
| 39 |
+
6765,
|
| 40 |
+
17936,
|
| 41 |
+
14099,
|
| 42 |
+
26748
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8603795861136222,
|
| 4 |
+
"test_acc": 0.8606896551724138,
|
| 5 |
+
"test_auc": 0.8902351994155528,
|
| 6 |
+
"val_auc": 0.9346590909090908,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "82_federal-government",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
3001
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9521773371242497,
|
| 18 |
+
"test_acc": 0.952183908045977,
|
| 19 |
+
"test_auc": 0.9873462782076953,
|
| 20 |
+
"val_auc": 0.9898538961038961,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "82_federal-government",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
3001,
|
| 28 |
+
23729,
|
| 29 |
+
21688,
|
| 30 |
+
6075,
|
| 31 |
+
16687,
|
| 32 |
+
5359,
|
| 33 |
+
14602,
|
| 34 |
+
17772,
|
| 35 |
+
25741,
|
| 36 |
+
21105,
|
| 37 |
+
24613,
|
| 38 |
+
30161,
|
| 39 |
+
21814,
|
| 40 |
+
16715,
|
| 41 |
+
16033,
|
| 42 |
+
8522
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9434405043996937,
|
| 4 |
+
"test_acc": 0.9434482758620689,
|
| 5 |
+
"test_auc": 0.9587492051788516,
|
| 6 |
+
"val_auc": 0.9659090909090909,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "83_third-party",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
13222
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9737892139261309,
|
| 18 |
+
"test_acc": 0.9737931034482759,
|
| 19 |
+
"test_auc": 0.9919055076032253,
|
| 20 |
+
"val_auc": 0.9882305194805195,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "83_third-party",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
13222,
|
| 28 |
+
8100,
|
| 29 |
+
3001,
|
| 30 |
+
13751,
|
| 31 |
+
31494,
|
| 32 |
+
27457,
|
| 33 |
+
29728,
|
| 34 |
+
19751,
|
| 35 |
+
19828,
|
| 36 |
+
14055,
|
| 37 |
+
26146,
|
| 38 |
+
8522,
|
| 39 |
+
9414,
|
| 40 |
+
7771,
|
| 41 |
+
28225,
|
| 42 |
+
21891
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8228864399125922,
|
| 4 |
+
"test_acc": 0.8234482758620689,
|
| 5 |
+
"test_auc": 0.8840905554954273,
|
| 6 |
+
"val_auc": 0.8461850649350648,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "84_clinical-trials",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
17100
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9558169584403526,
|
| 18 |
+
"test_acc": 0.9558620689655173,
|
| 19 |
+
"test_auc": 0.9851706667027437,
|
| 20 |
+
"val_auc": 0.9857954545454546,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "84_clinical-trials",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
17100,
|
| 28 |
+
3001,
|
| 29 |
+
348,
|
| 30 |
+
6596,
|
| 31 |
+
15415,
|
| 32 |
+
7771,
|
| 33 |
+
281,
|
| 34 |
+
13526,
|
| 35 |
+
6400,
|
| 36 |
+
12087,
|
| 37 |
+
6175,
|
| 38 |
+
14749,
|
| 39 |
+
531,
|
| 40 |
+
8522,
|
| 41 |
+
17469,
|
| 42 |
+
27919
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9365430050302544,
|
| 4 |
+
"test_acc": 0.9365517241379311,
|
| 5 |
+
"test_auc": 0.9550790762487147,
|
| 6 |
+
"val_auc": 0.9657061688311689,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "85_mental-health",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
19285
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9622969377817433,
|
| 18 |
+
"test_acc": 0.9622988505747126,
|
| 19 |
+
"test_auc": 0.9923680258672005,
|
| 20 |
+
"val_auc": 0.9995941558441559,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "85_mental-health",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
19285,
|
| 28 |
+
3001,
|
| 29 |
+
26013,
|
| 30 |
+
253,
|
| 31 |
+
7540,
|
| 32 |
+
1061,
|
| 33 |
+
6765,
|
| 34 |
+
2341,
|
| 35 |
+
16033,
|
| 36 |
+
18085,
|
| 37 |
+
9614,
|
| 38 |
+
7774,
|
| 39 |
+
30848,
|
| 40 |
+
820,
|
| 41 |
+
18401,
|
| 42 |
+
16046
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6080957035328844,
|
| 4 |
+
"test_acc": 0.6294339622641509,
|
| 5 |
+
"test_auc": 0.6379037941792678,
|
| 6 |
+
"val_auc": 0.7070828331332533,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "87_glue_cola",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
3661
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6080957035328844,
|
| 18 |
+
"test_acc": 0.6294339622641509,
|
| 19 |
+
"test_auc": 0.6918238165783832,
|
| 20 |
+
"val_auc": 0.7070828331332533,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "87_glue_cola",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
3661,
|
| 28 |
+
10683,
|
| 29 |
+
14720,
|
| 30 |
+
20475,
|
| 31 |
+
16956,
|
| 32 |
+
16033,
|
| 33 |
+
9144,
|
| 34 |
+
15722,
|
| 35 |
+
23758,
|
| 36 |
+
2144,
|
| 37 |
+
5884,
|
| 38 |
+
531,
|
| 39 |
+
28820,
|
| 40 |
+
24690,
|
| 41 |
+
30387,
|
| 42 |
+
13605
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6871501787021208,
|
| 4 |
+
"test_acc": 0.6881878209831255,
|
| 5 |
+
"test_auc": 0.748728581825072,
|
| 6 |
+
"val_auc": 0.7700320512820513,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "89_glue_mrpc",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
29180
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7005713398306558,
|
| 18 |
+
"test_acc": 0.7006603081438004,
|
| 19 |
+
"test_auc": 0.776163654449856,
|
| 20 |
+
"val_auc": 0.8597756410256411,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "89_glue_mrpc",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
29180,
|
| 28 |
+
31496,
|
| 29 |
+
11069,
|
| 30 |
+
15087,
|
| 31 |
+
531,
|
| 32 |
+
3494,
|
| 33 |
+
12881,
|
| 34 |
+
15612,
|
| 35 |
+
6333,
|
| 36 |
+
28281,
|
| 37 |
+
11026,
|
| 38 |
+
12642,
|
| 39 |
+
28402,
|
| 40 |
+
20613,
|
| 41 |
+
32194,
|
| 42 |
+
9579
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6446531082675732,
|
| 4 |
+
"test_acc": 0.6520754716981132,
|
| 5 |
+
"test_auc": 0.6717270153381285,
|
| 6 |
+
"val_auc": 0.6450580232092837,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "90_glue_qnli",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
14112
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7593107819700949,
|
| 18 |
+
"test_acc": 0.759748427672956,
|
| 19 |
+
"test_auc": 0.839029648449327,
|
| 20 |
+
"val_auc": 0.797719087635054,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "90_glue_qnli",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
14112,
|
| 28 |
+
6918,
|
| 29 |
+
3661,
|
| 30 |
+
7519,
|
| 31 |
+
11592,
|
| 32 |
+
26203,
|
| 33 |
+
14222,
|
| 34 |
+
10632,
|
| 35 |
+
17355,
|
| 36 |
+
16740,
|
| 37 |
+
531,
|
| 38 |
+
18992,
|
| 39 |
+
15811,
|
| 40 |
+
2681,
|
| 41 |
+
1457,
|
| 42 |
+
31024
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7129552118427894,
|
| 4 |
+
"test_acc": 0.7129559748427673,
|
| 5 |
+
"test_auc": 0.753697575488158,
|
| 6 |
+
"val_auc": 0.7835134053621449,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "91_glue_qqp",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
32623
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7846270992724697,
|
| 18 |
+
"test_acc": 0.7846540880503144,
|
| 19 |
+
"test_auc": 0.8576373945737839,
|
| 20 |
+
"val_auc": 0.8727490996398559,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "91_glue_qqp",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
32623,
|
| 28 |
+
22167,
|
| 29 |
+
31496,
|
| 30 |
+
19245,
|
| 31 |
+
4374,
|
| 32 |
+
7533,
|
| 33 |
+
4990,
|
| 34 |
+
17209,
|
| 35 |
+
1935,
|
| 36 |
+
32172,
|
| 37 |
+
3169,
|
| 38 |
+
6788,
|
| 39 |
+
30748,
|
| 40 |
+
10434,
|
| 41 |
+
20803,
|
| 42 |
+
22169
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7053409665509465,
|
| 4 |
+
"test_acc": 0.7232704402515723,
|
| 5 |
+
"test_auc": 0.7265604193859686,
|
| 6 |
+
"val_auc": 0.6788715486194478,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "92_glue_sst2",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
29212
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8266675396644553,
|
| 18 |
+
"test_acc": 0.8279245283018868,
|
| 19 |
+
"test_auc": 0.9099946938804443,
|
| 20 |
+
"val_auc": 0.9283713485394158,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "92_glue_sst2",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
29212,
|
| 28 |
+
14615,
|
| 29 |
+
9708,
|
| 30 |
+
9404,
|
| 31 |
+
25983,
|
| 32 |
+
16033,
|
| 33 |
+
2014,
|
| 34 |
+
14981,
|
| 35 |
+
25157,
|
| 36 |
+
26579,
|
| 37 |
+
10040,
|
| 38 |
+
8719,
|
| 39 |
+
30606,
|
| 40 |
+
17726,
|
| 41 |
+
17499,
|
| 42 |
+
9783
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7007638062213549,
|
| 4 |
+
"test_acc": 0.7008805031446541,
|
| 5 |
+
"test_auc": 0.7483305722609437,
|
| 6 |
+
"val_auc": 0.7759103641456582,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "94_ai_gen",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
24096
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8652829423798215,
|
| 18 |
+
"test_acc": 0.8654088050314466,
|
| 19 |
+
"test_auc": 0.9484924139704863,
|
| 20 |
+
"val_auc": 0.9783913565426171,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "94_ai_gen",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
24096,
|
| 28 |
+
5809,
|
| 29 |
+
19045,
|
| 30 |
+
71,
|
| 31 |
+
25637,
|
| 32 |
+
15674,
|
| 33 |
+
16033,
|
| 34 |
+
1987,
|
| 35 |
+
11838,
|
| 36 |
+
17800,
|
| 37 |
+
26324,
|
| 38 |
+
16510,
|
| 39 |
+
7771,
|
| 40 |
+
29499,
|
| 41 |
+
8489,
|
| 42 |
+
15290
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.618026418026418,
|
| 4 |
+
"test_acc": 0.6565656565656566,
|
| 5 |
+
"test_auc": 0.6563265306122449,
|
| 6 |
+
"val_auc": 0.626201923076923,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "95_toxic_is",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
1681
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7070707070707071,
|
| 18 |
+
"test_acc": 0.7070707070707071,
|
| 19 |
+
"test_auc": 0.7942857142857143,
|
| 20 |
+
"val_auc": 0.8297275641025641,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "95_toxic_is",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
1681,
|
| 28 |
+
1096,
|
| 29 |
+
27209,
|
| 30 |
+
16664,
|
| 31 |
+
531,
|
| 32 |
+
15487,
|
| 33 |
+
6389,
|
| 34 |
+
14141,
|
| 35 |
+
30612,
|
| 36 |
+
397,
|
| 37 |
+
23705,
|
| 38 |
+
16956,
|
| 39 |
+
13666,
|
| 40 |
+
1133,
|
| 41 |
+
21653,
|
| 42 |
+
13265
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-softmax_topp_zero-p0.8/k-50/seed-0/model.layers.12/w-32768/t-300M/l0-44.8/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8506566512416085,
|
| 4 |
+
"test_acc": 0.8507462686567164,
|
| 5 |
+
"test_auc": 0.9201218403346063,
|
| 6 |
+
"val_auc": 0.96,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "96_spam_is",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
20664
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9680153126089237,
|
| 18 |
+
"test_acc": 0.9680170575692963,
|
| 19 |
+
"test_auc": 0.9939079832696854,
|
| 20 |
+
"val_auc": 0.994,
|
| 21 |
+
"k": 16,
|
| 22 |
+
"dataset": "96_spam_is",
|
| 23 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 24 |
+
"reg_type": "l1",
|
| 25 |
+
"binarize": false,
|
| 26 |
+
"indices": [
|
| 27 |
+
20664,
|
| 28 |
+
9594,
|
| 29 |
+
2014,
|
| 30 |
+
1096,
|
| 31 |
+
4244,
|
| 32 |
+
24434,
|
| 33 |
+
14284,
|
| 34 |
+
6066,
|
| 35 |
+
31584,
|
| 36 |
+
10940,
|
| 37 |
+
17888,
|
| 38 |
+
25793,
|
| 39 |
+
14898,
|
| 40 |
+
31953,
|
| 41 |
+
1708,
|
| 42 |
+
5832
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/binarized_probing_binarized_linear_results.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/cfg.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"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", "d_sae": 32768, "dtype": "float32", "reshape_activations": "none", "normalize_activations": "none", "apply_b_dec_to_input": true, "d_in": 2304, "architecture": "jumprelu"}
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/binarized_probing_binarized_linear_results.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.959367218088842,
|
| 4 |
+
"test_acc": 0.9594272076372315,
|
| 5 |
+
"test_auc": 0.9603050701021457,
|
| 6 |
+
"val_auc": 0.9725439725439726,
|
| 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 |
+
2323
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9894988066825776,
|
| 18 |
+
"test_acc": 0.9894988066825776,
|
| 19 |
+
"test_auc": 0.9989810946579467,
|
| 20 |
+
"val_auc": 0.9957099957099957,
|
| 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 |
+
2323,
|
| 28 |
+
24699,
|
| 29 |
+
14703,
|
| 30 |
+
24732,
|
| 31 |
+
1958,
|
| 32 |
+
10841,
|
| 33 |
+
11876,
|
| 34 |
+
25492,
|
| 35 |
+
7859,
|
| 36 |
+
1565,
|
| 37 |
+
21905,
|
| 38 |
+
14109,
|
| 39 |
+
26523,
|
| 40 |
+
5822,
|
| 41 |
+
21053,
|
| 42 |
+
2849
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9273201077022764,
|
| 4 |
+
"test_acc": 0.9274141283214518,
|
| 5 |
+
"test_auc": 0.9578780333729832,
|
| 6 |
+
"val_auc": 0.9417114510128154,
|
| 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 |
+
24699
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9675953207889516,
|
| 18 |
+
"test_acc": 0.9675955930006481,
|
| 19 |
+
"test_auc": 0.9930327345550829,
|
| 20 |
+
"val_auc": 0.9913187267465895,
|
| 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 |
+
24699,
|
| 28 |
+
24607,
|
| 29 |
+
26064,
|
| 30 |
+
11321,
|
| 31 |
+
16191,
|
| 32 |
+
24732,
|
| 33 |
+
7859,
|
| 34 |
+
15785,
|
| 35 |
+
16791,
|
| 36 |
+
21879,
|
| 37 |
+
18449,
|
| 38 |
+
7569,
|
| 39 |
+
26735,
|
| 40 |
+
13296,
|
| 41 |
+
26530,
|
| 42 |
+
16807
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5455762409816379,
|
| 4 |
+
"test_acc": 0.5650314465408806,
|
| 5 |
+
"test_auc": 0.5760019351134485,
|
| 6 |
+
"val_auc": 0.5798319327731093,
|
| 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 |
+
2754
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5962024914835046,
|
| 18 |
+
"test_acc": 0.599748427672956,
|
| 19 |
+
"test_auc": 0.6406972281601031,
|
| 20 |
+
"val_auc": 0.6686674669867947,
|
| 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 |
+
2754,
|
| 28 |
+
2909,
|
| 29 |
+
11853,
|
| 30 |
+
16807,
|
| 31 |
+
32,
|
| 32 |
+
21584,
|
| 33 |
+
13296,
|
| 34 |
+
27972,
|
| 35 |
+
28026,
|
| 36 |
+
28552,
|
| 37 |
+
29989,
|
| 38 |
+
12531,
|
| 39 |
+
20642,
|
| 40 |
+
31033,
|
| 41 |
+
27330,
|
| 42 |
+
4999
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6795778504540694,
|
| 4 |
+
"test_acc": 0.6908176100628931,
|
| 5 |
+
"test_auc": 0.6998413480379003,
|
| 6 |
+
"val_auc": 0.7402961184473789,
|
| 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 |
+
21725
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7120980613086793,
|
| 18 |
+
"test_acc": 0.7137106918238993,
|
| 19 |
+
"test_auc": 0.7690748162857366,
|
| 20 |
+
"val_auc": 0.8007202881152461,
|
| 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 |
+
21725,
|
| 28 |
+
24844,
|
| 29 |
+
21571,
|
| 30 |
+
5605,
|
| 31 |
+
10176,
|
| 32 |
+
13440,
|
| 33 |
+
14895,
|
| 34 |
+
20132,
|
| 35 |
+
21584,
|
| 36 |
+
4684,
|
| 37 |
+
32446,
|
| 38 |
+
21879,
|
| 39 |
+
30139,
|
| 40 |
+
15364,
|
| 41 |
+
27478,
|
| 42 |
+
19492
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.782253087783424,
|
| 4 |
+
"test_acc": 0.7828930817610062,
|
| 5 |
+
"test_auc": 0.8190025912900655,
|
| 6 |
+
"val_auc": 0.8647458983593438,
|
| 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 |
+
24732
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7905525592084351,
|
| 18 |
+
"test_acc": 0.7906918238993711,
|
| 19 |
+
"test_auc": 0.8812019069626618,
|
| 20 |
+
"val_auc": 0.928171268507403,
|
| 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 |
+
24732,
|
| 28 |
+
4096,
|
| 29 |
+
16791,
|
| 30 |
+
5608,
|
| 31 |
+
2207,
|
| 32 |
+
24347,
|
| 33 |
+
29302,
|
| 34 |
+
4024,
|
| 35 |
+
16191,
|
| 36 |
+
14703,
|
| 37 |
+
6125,
|
| 38 |
+
30907,
|
| 39 |
+
7405,
|
| 40 |
+
21099,
|
| 41 |
+
29518,
|
| 42 |
+
12753
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8113970463113855,
|
| 4 |
+
"test_acc": 0.8128301886792453,
|
| 5 |
+
"test_auc": 0.8313391926799852,
|
| 6 |
+
"val_auc": 0.8367346938775511,
|
| 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 |
+
14822
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.88,
|
| 18 |
+
"test_acc": 0.88,
|
| 19 |
+
"test_auc": 0.9455107595750649,
|
| 20 |
+
"val_auc": 0.976390556222489,
|
| 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 |
+
14822,
|
| 28 |
+
1201,
|
| 29 |
+
21174,
|
| 30 |
+
2754,
|
| 31 |
+
17637,
|
| 32 |
+
32610,
|
| 33 |
+
21386,
|
| 34 |
+
13296,
|
| 35 |
+
11435,
|
| 36 |
+
16287,
|
| 37 |
+
22619,
|
| 38 |
+
31233,
|
| 39 |
+
19847,
|
| 40 |
+
21879,
|
| 41 |
+
18940,
|
| 42 |
+
24751
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5315664314199846,
|
| 4 |
+
"test_acc": 0.5548717948717948,
|
| 5 |
+
"test_auc": 0.5638990810246742,
|
| 6 |
+
"val_auc": 0.5813953488372093,
|
| 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 |
+
13782
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6012577133118125,
|
| 18 |
+
"test_acc": 0.602051282051282,
|
| 19 |
+
"test_auc": 0.6425253307301309,
|
| 20 |
+
"val_auc": 0.704202366381069,
|
| 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 |
+
13782,
|
| 28 |
+
22911,
|
| 29 |
+
15118,
|
| 30 |
+
1728,
|
| 31 |
+
13871,
|
| 32 |
+
18557,
|
| 33 |
+
29372,
|
| 34 |
+
6782,
|
| 35 |
+
26820,
|
| 36 |
+
7385,
|
| 37 |
+
595,
|
| 38 |
+
11973,
|
| 39 |
+
22581,
|
| 40 |
+
27980,
|
| 41 |
+
15866,
|
| 42 |
+
4559
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5161871408684604,
|
| 4 |
+
"test_acc": 0.521025641025641,
|
| 5 |
+
"test_auc": 0.5241820109738446,
|
| 6 |
+
"val_auc": 0.5558955528355773,
|
| 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 |
+
17367
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5261528492378124,
|
| 18 |
+
"test_acc": 0.5261538461538462,
|
| 19 |
+
"test_auc": 0.5394435654896153,
|
| 20 |
+
"val_auc": 0.5716034271725826,
|
| 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 |
+
17367,
|
| 28 |
+
1886,
|
| 29 |
+
26064,
|
| 30 |
+
23231,
|
| 31 |
+
22911,
|
| 32 |
+
13871,
|
| 33 |
+
7341,
|
| 34 |
+
19414,
|
| 35 |
+
595,
|
| 36 |
+
16288,
|
| 37 |
+
21290,
|
| 38 |
+
27980,
|
| 39 |
+
10047,
|
| 40 |
+
26820,
|
| 41 |
+
27712,
|
| 42 |
+
26627
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5088322965624019,
|
| 4 |
+
"test_acc": 0.5384615384615384,
|
| 5 |
+
"test_auc": 0.5416568485542128,
|
| 6 |
+
"val_auc": 0.5085679314565483,
|
| 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 |
+
14126
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.5290027429264157,
|
| 18 |
+
"test_acc": 0.5487179487179488,
|
| 19 |
+
"test_auc": 0.5521720469922914,
|
| 20 |
+
"val_auc": 0.5197878416972664,
|
| 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 |
+
14126,
|
| 28 |
+
12977,
|
| 29 |
+
7341,
|
| 30 |
+
15480,
|
| 31 |
+
8270,
|
| 32 |
+
1728,
|
| 33 |
+
15833,
|
| 34 |
+
595,
|
| 35 |
+
32047,
|
| 36 |
+
15211,
|
| 37 |
+
17367,
|
| 38 |
+
29859,
|
| 39 |
+
27712,
|
| 40 |
+
20745,
|
| 41 |
+
26523,
|
| 42 |
+
15540
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8059046292218132,
|
| 4 |
+
"test_acc": 0.8080808080808081,
|
| 5 |
+
"test_auc": 0.8518367346938776,
|
| 6 |
+
"val_auc": 0.862179487179487,
|
| 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 |
+
7696
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8378878804410719,
|
| 18 |
+
"test_acc": 0.8383838383838383,
|
| 19 |
+
"test_auc": 0.9130612244897959,
|
| 20 |
+
"val_auc": 0.9387019230769231,
|
| 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 |
+
7696,
|
| 28 |
+
23934,
|
| 29 |
+
17791,
|
| 30 |
+
26649,
|
| 31 |
+
23422,
|
| 32 |
+
31575,
|
| 33 |
+
1522,
|
| 34 |
+
24719,
|
| 35 |
+
12238,
|
| 36 |
+
11792,
|
| 37 |
+
3605,
|
| 38 |
+
17337,
|
| 39 |
+
2606,
|
| 40 |
+
10463,
|
| 41 |
+
8475,
|
| 42 |
+
14922
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.888410991636798,
|
| 4 |
+
"test_acc": 0.8888888888888888,
|
| 5 |
+
"test_auc": 0.913469387755102,
|
| 6 |
+
"val_auc": 0.937099358974359,
|
| 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 |
+
1911
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8988661121014063,
|
| 18 |
+
"test_acc": 0.898989898989899,
|
| 19 |
+
"test_auc": 0.9606122448979592,
|
| 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 |
+
1911,
|
| 28 |
+
3791,
|
| 29 |
+
24088,
|
| 30 |
+
10463,
|
| 31 |
+
14922,
|
| 32 |
+
17337,
|
| 33 |
+
1522,
|
| 34 |
+
24719,
|
| 35 |
+
4546,
|
| 36 |
+
919,
|
| 37 |
+
8537,
|
| 38 |
+
18276,
|
| 39 |
+
27822,
|
| 40 |
+
14828,
|
| 41 |
+
13274,
|
| 42 |
+
10797
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/119_us_state_TX_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": 0.8924489795918367,
|
| 6 |
+
"val_auc": 0.8373397435897435,
|
| 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 |
+
27822
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8886617839096123,
|
| 18 |
+
"test_acc": 0.8888888888888888,
|
| 19 |
+
"test_auc": 0.9702040816326529,
|
| 20 |
+
"val_auc": 0.9579326923076923,
|
| 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 |
+
27822,
|
| 28 |
+
17337,
|
| 29 |
+
10463,
|
| 30 |
+
13871,
|
| 31 |
+
24088,
|
| 32 |
+
3791,
|
| 33 |
+
7696,
|
| 34 |
+
24574,
|
| 35 |
+
18276,
|
| 36 |
+
8104,
|
| 37 |
+
26049,
|
| 38 |
+
11797,
|
| 39 |
+
12238,
|
| 40 |
+
14922,
|
| 41 |
+
28623,
|
| 42 |
+
7989
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6461887630200418,
|
| 4 |
+
"test_acc": 0.6610316428261812,
|
| 5 |
+
"test_auc": 0.6810757409275179,
|
| 6 |
+
"val_auc": 0.7667999999999999,
|
| 7 |
+
"k": 1,
|
| 8 |
+
"dataset": "120_us_timezone_Chicago",
|
| 9 |
+
"hook_name": "blocks.12.hook_resid_post",
|
| 10 |
+
"reg_type": "l1",
|
| 11 |
+
"binarize": false,
|
| 12 |
+
"indices": [
|
| 13 |
+
1911
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7888316824387592,
|
| 18 |
+
"test_acc": 0.788903337667967,
|
| 19 |
+
"test_auc": 0.8772548742561413,
|
| 20 |
+
"val_auc": 0.9231999999999999,
|
| 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 |
+
1911,
|
| 28 |
+
27822,
|
| 29 |
+
3791,
|
| 30 |
+
10753,
|
| 31 |
+
10797,
|
| 32 |
+
24719,
|
| 33 |
+
1813,
|
| 34 |
+
31575,
|
| 35 |
+
26649,
|
| 36 |
+
18276,
|
| 37 |
+
18297,
|
| 38 |
+
22581,
|
| 39 |
+
919,
|
| 40 |
+
25440,
|
| 41 |
+
18839,
|
| 42 |
+
25926
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6347158507478612,
|
| 4 |
+
"test_acc": 0.6488946684005201,
|
| 5 |
+
"test_auc": 0.6738746484568174,
|
| 6 |
+
"val_auc": 0.7054,
|
| 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 |
+
1911
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8219938512417924,
|
| 18 |
+
"test_acc": 0.8222800173385348,
|
| 19 |
+
"test_auc": 0.9200112433693431,
|
| 20 |
+
"val_auc": 0.9384000000000001,
|
| 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 |
+
1911,
|
| 28 |
+
27822,
|
| 29 |
+
18276,
|
| 30 |
+
2606,
|
| 31 |
+
31575,
|
| 32 |
+
6682,
|
| 33 |
+
10463,
|
| 34 |
+
1522,
|
| 35 |
+
24088,
|
| 36 |
+
18839,
|
| 37 |
+
7696,
|
| 38 |
+
23231,
|
| 39 |
+
19198,
|
| 40 |
+
10797,
|
| 41 |
+
10965,
|
| 42 |
+
3605
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8237901860112662,
|
| 4 |
+
"test_acc": 0.8248807975726051,
|
| 5 |
+
"test_auc": 0.8595818909603611,
|
| 6 |
+
"val_auc": 0.8536,
|
| 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 |
+
1911
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.8502577495230544,
|
| 18 |
+
"test_acc": 0.8513220632856524,
|
| 19 |
+
"test_auc": 0.9384256426983485,
|
| 20 |
+
"val_auc": 0.9344,
|
| 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 |
+
1911,
|
| 28 |
+
3791,
|
| 29 |
+
919,
|
| 30 |
+
12326,
|
| 31 |
+
10753,
|
| 32 |
+
24719,
|
| 33 |
+
10463,
|
| 34 |
+
1522,
|
| 35 |
+
14828,
|
| 36 |
+
15101,
|
| 37 |
+
26649,
|
| 38 |
+
27822,
|
| 39 |
+
26892,
|
| 40 |
+
25926,
|
| 41 |
+
2606,
|
| 42 |
+
23422
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9456214626086489,
|
| 4 |
+
"test_acc": 0.9456410256410256,
|
| 5 |
+
"test_auc": 0.978277426195846,
|
| 6 |
+
"val_auc": 0.9775601795185638,
|
| 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 |
+
20640
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.978461402518014,
|
| 18 |
+
"test_acc": 0.9784615384615385,
|
| 19 |
+
"test_auc": 0.9930698151950719,
|
| 20 |
+
"val_auc": 0.9959200326397388,
|
| 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 |
+
20640,
|
| 28 |
+
13708,
|
| 29 |
+
13871,
|
| 30 |
+
27980,
|
| 31 |
+
8609,
|
| 32 |
+
23422,
|
| 33 |
+
12238,
|
| 34 |
+
8104,
|
| 35 |
+
16807,
|
| 36 |
+
15850,
|
| 37 |
+
27886,
|
| 38 |
+
1728,
|
| 39 |
+
8389,
|
| 40 |
+
6547,
|
| 41 |
+
29067,
|
| 42 |
+
10214
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.8643472702576219,
|
| 4 |
+
"test_acc": 0.8666666666666667,
|
| 5 |
+
"test_auc": 0.8673544955734339,
|
| 6 |
+
"val_auc": 0.9186046511627908,
|
| 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 |
+
26491
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9517945675403169,
|
| 18 |
+
"test_acc": 0.9517948717948718,
|
| 19 |
+
"test_auc": 0.989855084660181,
|
| 20 |
+
"val_auc": 0.996328029375765,
|
| 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 |
+
26491,
|
| 28 |
+
24789,
|
| 29 |
+
8226,
|
| 30 |
+
27886,
|
| 31 |
+
21911,
|
| 32 |
+
15635,
|
| 33 |
+
23422,
|
| 34 |
+
6787,
|
| 35 |
+
29067,
|
| 36 |
+
24574,
|
| 37 |
+
15850,
|
| 38 |
+
20640,
|
| 39 |
+
10214,
|
| 40 |
+
23419,
|
| 41 |
+
13871,
|
| 42 |
+
15101
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9651043454790824,
|
| 4 |
+
"test_acc": 0.9651282051282051,
|
| 5 |
+
"test_auc": 0.9794408725216279,
|
| 6 |
+
"val_auc": 1.0,
|
| 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 |
+
25252
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9856410256410256,
|
| 18 |
+
"test_acc": 0.9856410256410256,
|
| 19 |
+
"test_auc": 0.9983253105328711,
|
| 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 |
+
25252,
|
| 28 |
+
31884,
|
| 29 |
+
2545,
|
| 30 |
+
24088,
|
| 31 |
+
20640,
|
| 32 |
+
955,
|
| 33 |
+
19117,
|
| 34 |
+
26649,
|
| 35 |
+
6787,
|
| 36 |
+
22911,
|
| 37 |
+
27886,
|
| 38 |
+
13708,
|
| 39 |
+
5274,
|
| 40 |
+
24789,
|
| 41 |
+
16807,
|
| 42 |
+
24574
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6370172464072774,
|
| 4 |
+
"test_acc": 0.6371911573472041,
|
| 5 |
+
"test_auc": 0.6821500989807314,
|
| 6 |
+
"val_auc": 0.7288,
|
| 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 |
+
24844
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7302089043602942,
|
| 18 |
+
"test_acc": 0.7312527091460772,
|
| 19 |
+
"test_auc": 0.8064081192759149,
|
| 20 |
+
"val_auc": 0.9024,
|
| 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 |
+
24844,
|
| 28 |
+
11273,
|
| 29 |
+
2934,
|
| 30 |
+
19529,
|
| 31 |
+
8237,
|
| 32 |
+
24076,
|
| 33 |
+
24006,
|
| 34 |
+
10887,
|
| 35 |
+
15785,
|
| 36 |
+
11321,
|
| 37 |
+
24834,
|
| 38 |
+
21898,
|
| 39 |
+
14817,
|
| 40 |
+
32207,
|
| 41 |
+
17367,
|
| 42 |
+
11183
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6662949972785116,
|
| 4 |
+
"test_acc": 0.6727351538794972,
|
| 5 |
+
"test_auc": 0.6949875315843982,
|
| 6 |
+
"val_auc": 0.6752,
|
| 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 |
+
2934
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7670904838625288,
|
| 18 |
+
"test_acc": 0.7672301690507152,
|
| 19 |
+
"test_auc": 0.8469267121712479,
|
| 20 |
+
"val_auc": 0.8635999999999999,
|
| 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 |
+
2934,
|
| 28 |
+
19529,
|
| 29 |
+
8237,
|
| 30 |
+
24844,
|
| 31 |
+
10887,
|
| 32 |
+
22911,
|
| 33 |
+
16801,
|
| 34 |
+
14942,
|
| 35 |
+
11273,
|
| 36 |
+
26523,
|
| 37 |
+
24006,
|
| 38 |
+
32207,
|
| 39 |
+
11973,
|
| 40 |
+
27972,
|
| 41 |
+
7385,
|
| 42 |
+
19414
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.5569633124069554,
|
| 4 |
+
"test_acc": 0.5726051148677936,
|
| 5 |
+
"test_auc": 0.5820450305960939,
|
| 6 |
+
"val_auc": 0.5723999999999999,
|
| 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 |
+
14942
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.6443319017331821,
|
| 18 |
+
"test_acc": 0.6449934980494149,
|
| 19 |
+
"test_auc": 0.7196891238439096,
|
| 20 |
+
"val_auc": 0.7417999999999999,
|
| 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 |
+
14942,
|
| 28 |
+
8237,
|
| 29 |
+
2934,
|
| 30 |
+
15706,
|
| 31 |
+
6787,
|
| 32 |
+
19020,
|
| 33 |
+
14792,
|
| 34 |
+
19414,
|
| 35 |
+
27972,
|
| 36 |
+
16801,
|
| 37 |
+
7442,
|
| 38 |
+
16007,
|
| 39 |
+
4684,
|
| 40 |
+
2418,
|
| 41 |
+
18333,
|
| 42 |
+
12412
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.6259268191712314,
|
| 4 |
+
"test_acc": 0.6268656716417911,
|
| 5 |
+
"test_auc": 0.6701400254591743,
|
| 6 |
+
"val_auc": 0.6436000000000001,
|
| 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 |
+
20332
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.7776449999545051,
|
| 18 |
+
"test_acc": 0.7782515991471215,
|
| 19 |
+
"test_auc": 0.8465902891434804,
|
| 20 |
+
"val_auc": 0.8392000000000001,
|
| 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 |
+
20332,
|
| 28 |
+
23903,
|
| 29 |
+
29752,
|
| 30 |
+
26642,
|
| 31 |
+
8228,
|
| 32 |
+
11005,
|
| 33 |
+
9643,
|
| 34 |
+
1717,
|
| 35 |
+
30809,
|
| 36 |
+
27021,
|
| 37 |
+
32353,
|
| 38 |
+
8584,
|
| 39 |
+
8855,
|
| 40 |
+
23365,
|
| 41 |
+
15962,
|
| 42 |
+
18915
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9595877138734282,
|
| 4 |
+
"test_acc": 0.9595959595959596,
|
| 5 |
+
"test_auc": 0.9665306122448979,
|
| 6 |
+
"val_auc": 0.932367149758454,
|
| 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 |
+
22397
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9898969278371137,
|
| 18 |
+
"test_acc": 0.98989898989899,
|
| 19 |
+
"test_auc": 0.9995918367346939,
|
| 20 |
+
"val_auc": 0.998792270531401,
|
| 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 |
+
22397,
|
| 28 |
+
3508,
|
| 29 |
+
9696,
|
| 30 |
+
14433,
|
| 31 |
+
23660,
|
| 32 |
+
5542,
|
| 33 |
+
29667,
|
| 34 |
+
31372,
|
| 35 |
+
21726,
|
| 36 |
+
12753,
|
| 37 |
+
730,
|
| 38 |
+
5860,
|
| 39 |
+
14612,
|
| 40 |
+
24699,
|
| 41 |
+
17025,
|
| 42 |
+
1646
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.7272727272727273,
|
| 4 |
+
"test_acc": 0.7272727272727273,
|
| 5 |
+
"test_auc": 0.7812244897959183,
|
| 6 |
+
"val_auc": 0.7944847020933978,
|
| 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 |
+
17025
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.9494330117733285,
|
| 18 |
+
"test_acc": 0.9494949494949495,
|
| 19 |
+
"test_auc": 0.9869387755102041,
|
| 20 |
+
"val_auc": 0.9760466988727858,
|
| 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 |
+
17025,
|
| 28 |
+
6844,
|
| 29 |
+
18564,
|
| 30 |
+
16191,
|
| 31 |
+
26167,
|
| 32 |
+
20,
|
| 33 |
+
12553,
|
| 34 |
+
3508,
|
| 35 |
+
29586,
|
| 36 |
+
22743,
|
| 37 |
+
21854,
|
| 38 |
+
15007,
|
| 39 |
+
22397,
|
| 40 |
+
15898,
|
| 41 |
+
30536,
|
| 42 |
+
18868
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|
denoised/gemma-2-2b-btk-vnorm_topp-p0.9/k-100/seed-0/model.layers.12/w-32768/t-300M/l0-93.5/denoised/sae_probes/sae_probes_gemma-2-2b--vnorm_topp-p0.9/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"test_f1": 0.9187278859141043,
|
| 4 |
+
"test_acc": 0.9191919191919192,
|
| 5 |
+
"test_auc": 0.9199999999999999,
|
| 6 |
+
"val_auc": 0.9259259259259259,
|
| 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 |
+
12776
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"test_f1": 0.98989898989899,
|
| 18 |
+
"test_acc": 0.98989898989899,
|
| 19 |
+
"test_auc": 0.9993877551020408,
|
| 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 |
+
12776,
|
| 28 |
+
16642,
|
| 29 |
+
30536,
|
| 30 |
+
17968,
|
| 31 |
+
13157,
|
| 32 |
+
30619,
|
| 33 |
+
21,
|
| 34 |
+
29586,
|
| 35 |
+
16281,
|
| 36 |
+
12753,
|
| 37 |
+
26780,
|
| 38 |
+
25621,
|
| 39 |
+
27282,
|
| 40 |
+
18564,
|
| 41 |
+
15053,
|
| 42 |
+
27739
|
| 43 |
+
]
|
| 44 |
+
}
|
| 45 |
+
]
|