diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0e9cd5a43c22e7c98944970bb8aedd4f8d99d336 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6075800641892788, + "test_acc": 0.6081355932203389, + "test_auc": 0.6351933973885193, + "val_auc": 0.6308373590982287, + "k": 1, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14530 + ] + }, + { + "test_f1": 0.6113367780174659, + "test_acc": 0.6115254237288136, + "test_auc": 0.6330514831606933, + "val_auc": 0.6892109500805154, + "k": 2, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14530, + 17630 + ] + }, + { + "test_f1": 0.6222218114076764, + "test_acc": 0.6223728813559322, + "test_auc": 0.6505388798799793, + "val_auc": 0.7375201288244766, + "k": 5, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14530, + 17630, + 31015, + 7467, + 27275 + ] + }, + { + "test_f1": 0.6408101058555727, + "test_acc": 0.6413559322033898, + "test_auc": 0.6763760649818167, + "val_auc": 0.7648953301127215, + "k": 10, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14530, + 17630, + 31015, + 7467, + 27275, + 20399, + 2509, + 13504, + 16025, + 5460 + ] + }, + { + "test_f1": 0.6450860385250187, + "test_acc": 0.6454237288135594, + "test_auc": 0.6769717561490405, + "val_auc": 0.7608695652173914, + "k": 16, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14530, + 17630, + 31015, + 7467, + 27275, + 20399, + 2509, + 13504, + 16025, + 5460, + 14822, + 4648, + 19078, + 29250, + 7125, + 5437 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b09eb33cc20c5a3a5eb84968c1baed2370a13acd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5695601396124615, + "test_acc": 0.5701694915254237, + "test_auc": 0.5770822531834545, + "val_auc": 0.5915861513687601, + "k": 1, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2509 + ] + }, + { + "test_f1": 0.5643123312527998, + "test_acc": 0.5654237288135593, + "test_auc": 0.5857850437391756, + "val_auc": 0.6201690821256038, + "k": 2, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2509, + 25315 + ] + }, + { + "test_f1": 0.5647277566952745, + "test_acc": 0.5647457627118644, + "test_auc": 0.6059043290568591, + "val_auc": 0.7033011272141707, + "k": 5, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2509, + 25315, + 7467, + 1833, + 22739 + ] + }, + { + "test_f1": 0.5721994571697209, + "test_acc": 0.5722033898305084, + "test_auc": 0.601387004372079, + "val_auc": 0.7028985507246377, + "k": 10, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2509, + 25315, + 7467, + 1833, + 22739, + 14873, + 14288, + 17402, + 27151, + 9524 + ] + }, + { + "test_f1": 0.5769491525423729, + "test_acc": 0.5769491525423729, + "test_auc": 0.6028210757005806, + "val_auc": 0.716988727858293, + "k": 16, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2509, + 25315, + 7467, + 1833, + 22739, + 14873, + 14288, + 17402, + 27151, + 9524, + 26568, + 9773, + 6399, + 6201, + 25415, + 8950 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..66676a411375d98e2f81c2619b62696068781307 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5754470989050432, + "test_acc": 0.575682382133995, + "test_auc": 0.627013447613418, + "val_auc": 0.5542216886754702, + "k": 1, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26510 + ] + }, + { + "test_f1": 0.6243630110948751, + "test_acc": 0.6277915632754343, + "test_auc": 0.6455593320526083, + "val_auc": 0.6102440976390556, + "k": 2, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26510, + 4981 + ] + }, + { + "test_f1": 0.604697629182029, + "test_acc": 0.607940446650124, + "test_auc": 0.6626767154327373, + "val_auc": 0.6246498599439776, + "k": 5, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26510, + 4981, + 32528, + 18849, + 10549 + ] + }, + { + "test_f1": 0.5959946125980281, + "test_acc": 0.598014888337469, + "test_auc": 0.648243928870499, + "val_auc": 0.6566626650660263, + "k": 10, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26510, + 4981, + 32528, + 18849, + 10549, + 1765, + 28723, + 12462, + 8489, + 15544 + ] + }, + { + "test_f1": 0.5883652482425866, + "test_acc": 0.5905707196029777, + "test_auc": 0.6557066154376632, + "val_auc": 0.6910764305722289, + "k": 16, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26510, + 4981, + 32528, + 18849, + 10549, + 1765, + 28723, + 12462, + 8489, + 15544, + 13474, + 14050, + 17050, + 9796, + 2176, + 11678 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..eb1a13399257c20590a879ad5a43633477cd48f9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5804126299120017, + "test_acc": 0.5806451612903226, + "test_auc": 0.6025072656519384, + "val_auc": 0.6938775510204083, + "k": 1, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26677 + ] + }, + { + "test_f1": 0.5724849221757128, + "test_acc": 0.5732009925558312, + "test_auc": 0.5984064824392887, + "val_auc": 0.7248899559823929, + "k": 2, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26677, + 17459 + ] + }, + { + "test_f1": 0.6050163893966636, + "test_acc": 0.6054590570719603, + "test_auc": 0.6499433525442097, + "val_auc": 0.8151260504201681, + "k": 5, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26677, + 17459, + 7862, + 9325, + 31614 + ] + }, + { + "test_f1": 0.5872199984123812, + "test_acc": 0.5880893300248139, + "test_auc": 0.6498940938870006, + "val_auc": 0.7739095638255302, + "k": 10, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26677, + 17459, + 7862, + 9325, + 31614, + 26960, + 3195, + 28038, + 21724, + 19460 + ] + }, + { + "test_f1": 0.6321381888488691, + "test_acc": 0.6327543424317618, + "test_auc": 0.6649426136643515, + "val_auc": 0.7991196478591436, + "k": 16, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26677, + 17459, + 7862, + 9325, + 31614, + 26960, + 3195, + 28038, + 21724, + 19460, + 22932, + 12747, + 22687, + 31570, + 23507, + 24193 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bb92d6ddcffe3882d3b1343a35d0d80b59865ea8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.605531478580386, + "test_acc": 0.607940446650124, + "test_auc": 0.6494753953007242, + "val_auc": 0.7156862745098039, + "k": 1, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26179 + ] + }, + { + "test_f1": 0.6234277412829916, + "test_acc": 0.6253101736972705, + "test_auc": 0.6808654746071622, + "val_auc": 0.7631052420968387, + "k": 2, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26179, + 14082 + ] + }, + { + "test_f1": 0.6499300640094215, + "test_acc": 0.6501240694789082, + "test_auc": 0.7181296487857741, + "val_auc": 0.668267306922769, + "k": 5, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26179, + 14082, + 2223, + 4843, + 28727 + ] + }, + { + "test_f1": 0.6872062361131884, + "test_acc": 0.6873449131513648, + "test_auc": 0.7441751637850351, + "val_auc": 0.6466586634653861, + "k": 10, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26179, + 14082, + 2223, + 4843, + 28727, + 4297, + 30058, + 587, + 29780, + 18374 + ] + }, + { + "test_f1": 0.6674937965260546, + "test_acc": 0.6674937965260546, + "test_auc": 0.7585094330328557, + "val_auc": 0.6994797919167667, + "k": 16, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26179, + 14082, + 2223, + 4843, + 28727, + 4297, + 30058, + 587, + 29780, + 18374, + 21723, + 1192, + 4452, + 17981, + 28735, + 27025 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7a132d9310d7ad4ba64eabfed5071c49a8aaabac --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7102819956452633, + "test_acc": 0.7121820615796519, + "test_auc": 0.7741681122851285, + "val_auc": 0.7858585858585858, + "k": 1, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16256 + ] + }, + { + "test_f1": 0.738777926617996, + "test_acc": 0.7389558232931727, + "test_auc": 0.8139596564923801, + "val_auc": 0.846060606060606, + "k": 2, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16256, + 15495 + ] + }, + { + "test_f1": 0.7520538692122879, + "test_acc": 0.7523427041499331, + "test_auc": 0.8295974251265215, + "val_auc": 0.8937373737373737, + "k": 5, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16256, + 15495, + 4069, + 8890, + 19080 + ] + }, + { + "test_f1": 0.7758743279124481, + "test_acc": 0.7764390896921017, + "test_auc": 0.8593783601668793, + "val_auc": 0.922020202020202, + "k": 10, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16256, + 15495, + 4069, + 8890, + 19080, + 29735, + 15182, + 22863, + 638, + 8917 + ] + }, + { + "test_f1": 0.8097016640999816, + "test_acc": 0.8099062918340026, + "test_auc": 0.8992559246462416, + "val_auc": 0.9490909090909091, + "k": 16, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16256, + 15495, + 4069, + 8890, + 19080, + 29735, + 15182, + 22863, + 638, + 8917, + 13054, + 23830, + 23533, + 12157, + 1177, + 7662 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ac8acff33746d365af25631eaccb972ffa6c2c9c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6492067068935873, + "test_acc": 0.6532797858099063, + "test_auc": 0.6854023598227982, + "val_auc": 0.6844444444444444, + "k": 1, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14503 + ] + }, + { + "test_f1": 0.6773345383673495, + "test_acc": 0.677376171352075, + "test_auc": 0.7388424538716292, + "val_auc": 0.7729292929292929, + "k": 2, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14503, + 26815 + ] + }, + { + "test_f1": 0.7456237356528174, + "test_acc": 0.7456492637215528, + "test_auc": 0.8182319966738829, + "val_auc": 0.8432323232323231, + "k": 5, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14503, + 26815, + 28317, + 11462, + 23830 + ] + }, + { + "test_f1": 0.8111529295578486, + "test_acc": 0.8112449799196787, + "test_auc": 0.8624392481828218, + "val_auc": 0.8767676767676769, + "k": 10, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14503, + 26815, + 28317, + 11462, + 23830, + 31848, + 5574, + 12196, + 563, + 26848 + ] + }, + { + "test_f1": 0.82297849034258, + "test_acc": 0.8232931726907631, + "test_auc": 0.8917649926165933, + "val_auc": 0.907070707070707, + "k": 16, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14503, + 26815, + 28317, + 11462, + 23830, + 31848, + 5574, + 12196, + 563, + 26848, + 4093, + 28801, + 31715, + 5690, + 21568, + 21604 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..89bd6e4397f5861abc8b8e3715a8cf23c11e24d4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6636124839923084, + "test_acc": 0.6693440428380187, + "test_auc": 0.7070507949706814, + "val_auc": 0.7668686868686869, + "k": 1, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19907 + ] + }, + { + "test_f1": 0.7165757524285254, + "test_acc": 0.7175368139223561, + "test_auc": 0.8061640693323392, + "val_auc": 0.8626262626262626, + "k": 2, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19907, + 20419 + ] + }, + { + "test_f1": 0.7723781535840889, + "test_acc": 0.7724230254350736, + "test_auc": 0.8803458014938854, + "val_auc": 0.8917171717171717, + "k": 5, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19907, + 20419, + 32272, + 7565, + 30535 + ] + }, + { + "test_f1": 0.8458713241049357, + "test_acc": 0.8460508701472557, + "test_auc": 0.9177144413700163, + "val_auc": 0.915151515151515, + "k": 10, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19907, + 20419, + 32272, + 7565, + 30535, + 31886, + 22646, + 31295, + 22246, + 12522 + ] + }, + { + "test_f1": 0.8647925033467202, + "test_acc": 0.8647925033467202, + "test_auc": 0.9412338174363093, + "val_auc": 0.9341414141414142, + "k": 16, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19907, + 20419, + 32272, + 7565, + 30535, + 31886, + 22646, + 31295, + 22246, + 12522, + 12480, + 31512, + 10599, + 18685, + 1136, + 2860 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9a75681d510403cfdb3c6d1992563057bfe2a60a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6601897537201162, + "test_acc": 0.6656410256410257, + "test_auc": 0.7061509071935906, + "val_auc": 0.7860057119543045, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21991 + ] + }, + { + "test_f1": 0.6952865282441281, + "test_acc": 0.6953846153846154, + "test_auc": 0.7650196081731578, + "val_auc": 0.8535291717666259, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21991, + 2436 + ] + }, + { + "test_f1": 0.752562296947444, + "test_acc": 0.7528205128205128, + "test_auc": 0.8233581310802168, + "val_auc": 0.868625050999592, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21991, + 2436, + 27396, + 15985, + 27711 + ] + }, + { + "test_f1": 0.7565235089454156, + "test_acc": 0.7569230769230769, + "test_auc": 0.8438457265964252, + "val_auc": 0.8763769889840882, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21991, + 2436, + 27396, + 15985, + 27711, + 18248, + 27233, + 13859, + 15664, + 12103 + ] + }, + { + "test_f1": 0.7817785138457813, + "test_acc": 0.7825641025641026, + "test_auc": 0.8614720436260814, + "val_auc": 0.8698490412076703, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21991, + 2436, + 27396, + 15985, + 27711, + 18248, + 27233, + 13859, + 15664, + 12103, + 12663, + 31432, + 2179, + 21946, + 30131, + 14101 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b7bebef6b154b9250c501cabefd2a09eb66c93f2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7199675524477226, + "test_acc": 0.7220168067226891, + "test_auc": 0.764518750316362, + "val_auc": 0.7996788438378161, + "k": 1, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19541 + ] + }, + { + "test_f1": 0.786688982814831, + "test_acc": 0.786890756302521, + "test_auc": 0.858696290792604, + "val_auc": 0.9221196306704135, + "k": 2, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19541, + 25616 + ] + }, + { + "test_f1": 0.8263236162905293, + "test_acc": 0.8265546218487395, + "test_auc": 0.9046600556073786, + "val_auc": 0.9731031714171016, + "k": 5, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19541, + 25616, + 23529, + 3974, + 22117 + ] + }, + { + "test_f1": 0.843128976106703, + "test_acc": 0.8433613445378151, + "test_auc": 0.9233961356848962, + "val_auc": 0.967482938578884, + "k": 10, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19541, + 25616, + 23529, + 3974, + 22117, + 385, + 8067, + 22719, + 28670, + 4049 + ] + }, + { + "test_f1": 0.8472775155452953, + "test_acc": 0.8473949579831933, + "test_auc": 0.9269131758393533, + "val_auc": 0.9731031714171015, + "k": 16, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19541, + 25616, + 23529, + 3974, + 22117, + 385, + 8067, + 22719, + 28670, + 4049, + 4852, + 7539, + 16227, + 19278, + 8256, + 27684 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ad3db788551d84f4441bc1d189a79ce3fd4a292a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7901179443480568, + "test_acc": 0.7912605042016807, + "test_auc": 0.8456913320461925, + "val_auc": 0.8588920112404657, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23006 + ] + }, + { + "test_f1": 0.8080051417117977, + "test_acc": 0.8080672268907563, + "test_auc": 0.8921599200237181, + "val_auc": 0.9277398635086311, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23006, + 116 + ] + }, + { + "test_f1": 0.8112726943001087, + "test_acc": 0.8114285714285714, + "test_auc": 0.9000190269070294, + "val_auc": 0.9277398635086311, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23006, + 116, + 25843, + 19593, + 12034 + ] + }, + { + "test_f1": 0.8409569737822774, + "test_acc": 0.8410084033613445, + "test_auc": 0.9209334844639202, + "val_auc": 0.9566439181051787, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23006, + 116, + 25843, + 19593, + 12034, + 20787, + 5522, + 13405, + 32705, + 8835 + ] + }, + { + "test_f1": 0.8396194693007528, + "test_acc": 0.8396638655462185, + "test_auc": 0.927649847061631, + "val_auc": 0.9618627057406663, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23006, + 116, + 25843, + 19593, + 12034, + 20787, + 5522, + 13405, + 32705, + 8835, + 3039, + 3620, + 7175, + 3963, + 17283, + 786 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d6027269b372f9cf311af280ca0bb0c788ae0911 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7582792541891996, + "test_acc": 0.7589915966386555, + "test_auc": 0.7943717866672453, + "val_auc": 0.8317944600562023, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22683 + ] + }, + { + "test_f1": 0.8099742764605607, + "test_acc": 0.8100840336134454, + "test_auc": 0.8721256263965118, + "val_auc": 0.8703331995182657, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22683, + 191 + ] + }, + { + "test_f1": 0.8294757511820815, + "test_acc": 0.8295798319327731, + "test_auc": 0.9106535313216333, + "val_auc": 0.8900040144520274, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22683, + 191, + 21984, + 575, + 29104 + ] + }, + { + "test_f1": 0.8372836780846908, + "test_acc": 0.8373109243697479, + "test_auc": 0.9308577564700523, + "val_auc": 0.8948213568847851, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22683, + 191, + 21984, + 575, + 29104, + 1029, + 25764, + 12385, + 20928, + 16050 + ] + }, + { + "test_f1": 0.8682349963644872, + "test_acc": 0.8682352941176471, + "test_auc": 0.9490218994728508, + "val_auc": 0.9253311922922521, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22683, + 191, + 21984, + 575, + 29104, + 1029, + 25764, + 12385, + 20928, + 16050, + 2086, + 882, + 24733, + 14017, + 28725, + 28948 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e0adc0b96f759f0a854ff56b595f30f883196a28 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9213396319005319, + "test_acc": 0.921344537815126, + "test_auc": 0.9757106391594536, + "val_auc": 0.9817342432757927, + "k": 1, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1748 + ] + }, + { + "test_f1": 0.9606642423511109, + "test_acc": 0.9606722689075631, + "test_auc": 0.9930587944985575, + "val_auc": 0.9823364110798877, + "k": 2, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1748, + 32420 + ] + }, + { + "test_f1": 0.9737781884598148, + "test_acc": 0.973781512605042, + "test_auc": 0.9962425248208487, + "val_auc": 0.9883580891208351, + "k": 5, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1748, + 32420, + 23362, + 21505, + 30222 + ] + }, + { + "test_f1": 0.9936134410485875, + "test_acc": 0.9936134453781512, + "test_auc": 0.9997324482431973, + "val_auc": 0.9983942191890808, + "k": 10, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1748, + 32420, + 23362, + 21505, + 30222, + 19168, + 19479, + 3095, + 22929, + 4678 + ] + }, + { + "test_f1": 0.992604950107309, + "test_acc": 0.9926050420168068, + "test_auc": 0.9997505260646029, + "val_auc": 0.9987956643918106, + "k": 16, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1748, + 32420, + 23362, + 21505, + 30222, + 19168, + 19479, + 3095, + 22929, + 4678, + 75, + 14476, + 1719, + 22446, + 29926, + 25590 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4496ece8ffb5d3b75f664e02ab92caec725efad9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9374365366409826, + "test_acc": 0.9374789915966386, + "test_auc": 0.9830448113037002, + "val_auc": 0.9606583701324769, + "k": 1, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420 + ] + }, + { + "test_f1": 0.941762476197666, + "test_acc": 0.9418487394957983, + "test_auc": 0.9891813277798266, + "val_auc": 0.9763147330389402, + "k": 2, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420, + 19168 + ] + }, + { + "test_f1": 0.9704132483000207, + "test_acc": 0.9704201680672269, + "test_auc": 0.9971518392375498, + "val_auc": 0.9931754315535929, + "k": 5, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420, + 19168, + 28257, + 13386, + 25463 + ] + }, + { + "test_f1": 0.9936133328077134, + "test_acc": 0.9936134453781512, + "test_auc": 0.9997451027181812, + "val_auc": 0.9979927739863508, + "k": 10, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420, + 19168, + 28257, + 13386, + 25463, + 23362, + 12674, + 13544, + 15978, + 20986 + ] + }, + { + "test_f1": 0.9949579205278593, + "test_acc": 0.9949579831932773, + "test_auc": 0.9998223854046902, + "val_auc": 0.9983942191890807, + "k": 16, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420, + 19168, + 28257, + 13386, + 25463, + 23362, + 12674, + 13544, + 15978, + 20986, + 25590, + 30205, + 24802, + 1748, + 4396, + 22446 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8b8f8fa8255748817fba16de688f5df89e861fc2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9293358688376324, + "test_acc": 0.9294117647058824, + "test_auc": 0.9778786218915185, + "val_auc": 0.9614612605379367, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420 + ] + }, + { + "test_f1": 0.9234155700887897, + "test_acc": 0.9236974789915966, + "test_auc": 0.9806876893651794, + "val_auc": 0.9702930549979927, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420, + 1748 + ] + }, + { + "test_f1": 0.9488653297535855, + "test_acc": 0.9489075630252101, + "test_auc": 0.9912205060343767, + "val_auc": 0.9971898835808912, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420, + 1748, + 75, + 19168, + 23362 + ] + }, + { + "test_f1": 0.9875627301097328, + "test_acc": 0.987563025210084, + "test_auc": 0.9995756231425039, + "val_auc": 0.9983942191890808, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420, + 1748, + 75, + 19168, + 23362, + 13544, + 24802, + 13386, + 26281, + 16896 + ] + }, + { + "test_f1": 0.9939494895935167, + "test_acc": 0.9939495798319328, + "test_auc": 0.9999227173134911, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32420, + 1748, + 75, + 19168, + 23362, + 13544, + 24802, + 13386, + 26281, + 16896, + 29926, + 29921, + 19479, + 20986, + 30222, + 15978 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ce8652953d7a270f4447a198447d998de938fdba --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7327137290721708, + "test_acc": 0.7330817610062893, + "test_auc": 0.7986383322582703, + "val_auc": 0.7959183673469388, + "k": 1, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7777 + ] + }, + { + "test_f1": 0.7712616286635317, + "test_acc": 0.7715723270440251, + "test_auc": 0.8444310806965598, + "val_auc": 0.81312525010004, + "k": 2, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7777, + 18810 + ] + }, + { + "test_f1": 0.8206500113869484, + "test_acc": 0.8208805031446541, + "test_auc": 0.8941540536626907, + "val_auc": 0.876750700280112, + "k": 5, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7777, + 18810, + 17023, + 8090, + 6110 + ] + }, + { + "test_f1": 0.8444207785205643, + "test_acc": 0.8445283018867924, + "test_auc": 0.9198601270430838, + "val_auc": 0.9267707082833132, + "k": 10, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7777, + 18810, + 17023, + 8090, + 6110, + 24298, + 22531, + 5760, + 28678, + 14087 + ] + }, + { + "test_f1": 0.8592563222570098, + "test_acc": 0.8593710691823899, + "test_auc": 0.9304617336631771, + "val_auc": 0.9395758303321329, + "k": 16, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7777, + 18810, + 17023, + 8090, + 6110, + 24298, + 22531, + 5760, + 28678, + 14087, + 3714, + 9041, + 13597, + 15272, + 18882, + 9958 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6a95a06227429bd6309e8f121be22a766e3da981 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7297786783688569, + "test_acc": 0.7305863708399366, + "test_auc": 0.7819469559975889, + "val_auc": 0.8483393357342937, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19983 + ] + }, + { + "test_f1": 0.7458405350927172, + "test_acc": 0.7464342313787639, + "test_auc": 0.8503516174402249, + "val_auc": 0.8509403761504601, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19983, + 24827 + ] + }, + { + "test_f1": 0.8080181119362585, + "test_acc": 0.8082408874801902, + "test_auc": 0.8881052843078159, + "val_auc": 0.8699479791916767, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19983, + 24827, + 5910, + 4093, + 5760 + ] + }, + { + "test_f1": 0.8186593534661042, + "test_acc": 0.8193343898573693, + "test_auc": 0.9061482820976492, + "val_auc": 0.9179671868747499, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19983, + 24827, + 5910, + 4093, + 5760, + 662, + 4047, + 20972, + 21088, + 30573 + ] + }, + { + "test_f1": 0.8443273277442892, + "test_acc": 0.8446909667194928, + "test_auc": 0.9233172593932087, + "val_auc": 0.9367747098839536, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19983, + 24827, + 5910, + 4093, + 5760, + 662, + 4047, + 20972, + 21088, + 30573, + 23635, + 3107, + 2192, + 27034, + 21704, + 10426 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6686248326ddbcdceb3ded0773a871e9c37bc148 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.639002678250414, + "test_acc": 0.6398559423769508, + "test_auc": 0.67043096292197, + "val_auc": 0.7020933977455717, + "k": 1, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13472 + ] + }, + { + "test_f1": 0.6737803210995694, + "test_acc": 0.6746698679471789, + "test_auc": 0.7291378435712969, + "val_auc": 0.7177938808373592, + "k": 2, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13472, + 9512 + ] + }, + { + "test_f1": 0.7178578792883493, + "test_acc": 0.7178871548619448, + "test_auc": 0.7950562626821619, + "val_auc": 0.783011272141707, + "k": 5, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13472, + 9512, + 28764, + 31643, + 21733 + ] + }, + { + "test_f1": 0.7442490661948856, + "test_acc": 0.7442977190876351, + "test_auc": 0.824669110865154, + "val_auc": 0.8152173913043478, + "k": 10, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13472, + 9512, + 28764, + 31643, + 21733, + 13595, + 2435, + 19825, + 28334, + 26354 + ] + }, + { + "test_f1": 0.7646814646805151, + "test_acc": 0.7647058823529411, + "test_auc": 0.847981230400295, + "val_auc": 0.8301127214170693, + "k": 16, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13472, + 9512, + 28764, + 31643, + 21733, + 13595, + 2435, + 19825, + 28334, + 26354, + 6671, + 22798, + 29937, + 23387, + 7325, + 23904 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d3fc41763f29023a12c3d3a1ef0c570b9bde6a4a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6166958787564848, + "test_acc": 0.6170798898071626, + "test_auc": 0.6691305931637423, + "val_auc": 0.755502200880352, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8923 + ] + }, + { + "test_f1": 0.6519891866653652, + "test_acc": 0.6528925619834711, + "test_auc": 0.7105215226762187, + "val_auc": 0.7519007603041217, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8923, + 31900 + ] + }, + { + "test_f1": 0.7373710444891464, + "test_acc": 0.7382920110192838, + "test_auc": 0.8027745734928056, + "val_auc": 0.8255302120848339, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8923, + 31900, + 188, + 10273, + 11435 + ] + }, + { + "test_f1": 0.770688465336227, + "test_acc": 0.7713498622589532, + "test_auc": 0.8384736810151175, + "val_auc": 0.7899159663865546, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8923, + 31900, + 188, + 10273, + 11435, + 2615, + 570, + 18041, + 30957, + 3835 + ] + }, + { + "test_f1": 0.7905191383992516, + "test_acc": 0.790633608815427, + "test_auc": 0.8471252504401675, + "val_auc": 0.7971188475390155, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8923, + 31900, + 188, + 10273, + 11435, + 2615, + 570, + 18041, + 30957, + 3835, + 15043, + 15926, + 27321, + 16173, + 29696, + 1460 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e97251776e64e6e3e69c469943e407906b2327f1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6173725635151596, + "test_acc": 0.6188679245283019, + "test_auc": 0.647536325147665, + "val_auc": 0.6346538615446178, + "k": 1, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22138 + ] + }, + { + "test_f1": 0.621222153514086, + "test_acc": 0.6221383647798742, + "test_auc": 0.6608459260849444, + "val_auc": 0.6550620248099239, + "k": 2, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22138, + 17430 + ] + }, + { + "test_f1": 0.6550288979860017, + "test_acc": 0.6553459119496855, + "test_auc": 0.703127927099588, + "val_auc": 0.7202881152460985, + "k": 5, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22138, + 17430, + 30980, + 4281, + 30088 + ] + }, + { + "test_f1": 0.6568742953686459, + "test_acc": 0.6571069182389937, + "test_auc": 0.7220016120882314, + "val_auc": 0.7062825130052021, + "k": 10, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22138, + 17430, + 30980, + 4281, + 30088, + 8280, + 7168, + 21568, + 15801, + 31465 + ] + }, + { + "test_f1": 0.6713935082151044, + "test_acc": 0.6714465408805032, + "test_auc": 0.7382462869820837, + "val_auc": 0.7759103641456583, + "k": 16, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22138, + 17430, + 30980, + 4281, + 30088, + 8280, + 7168, + 21568, + 15801, + 31465, + 29657, + 31293, + 16542, + 4116, + 23084, + 22537 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fdf0e00a5b0807652d92c0c5ab3bc1c394bfe6e4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6728105126120991, + "test_acc": 0.6737106918238994, + "test_auc": 0.7319721043928393, + "val_auc": 0.7114845938375349, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29312 + ] + }, + { + "test_f1": 0.7363018704191072, + "test_acc": 0.7366037735849057, + "test_auc": 0.801022288739989, + "val_auc": 0.7879151660664265, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29312, + 24512 + ] + }, + { + "test_f1": 0.7859024376414564, + "test_acc": 0.7859119496855346, + "test_auc": 0.8566261686880214, + "val_auc": 0.8227290916366546, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29312, + 24512, + 2014, + 27875, + 7548 + ] + }, + { + "test_f1": 0.7994186320928678, + "test_acc": 0.799496855345912, + "test_auc": 0.8805566159918747, + "val_auc": 0.8867547018807523, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29312, + 24512, + 2014, + 27875, + 7548, + 25212, + 28591, + 18596, + 31010, + 11791 + ] + }, + { + "test_f1": 0.8111774713658202, + "test_acc": 0.8113207547169812, + "test_auc": 0.8996247236817989, + "val_auc": 0.9047619047619048, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29312, + 24512, + 2014, + 27875, + 7548, + 25212, + 28591, + 18596, + 31010, + 11791, + 14706, + 12097, + 32594, + 32755, + 28229, + 8717 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..354b947269d4d10e5c429b322a8f79c7c0e3a024 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5586401448470414, + "test_acc": 0.5656565656565656, + "test_auc": 0.6228571428571429, + "val_auc": 0.6463267763950221, + "k": 1, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29959 + ] + }, + { + "test_f1": 0.5552832244008714, + "test_acc": 0.5555555555555556, + "test_auc": 0.6053061224489795, + "val_auc": 0.7224006423123244, + "k": 2, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29959, + 26366 + ] + }, + { + "test_f1": 0.6565656565656566, + "test_acc": 0.6565656565656566, + "test_auc": 0.7020408163265306, + "val_auc": 0.8309915696507428, + "k": 5, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29959, + 26366, + 31591, + 25985, + 29262 + ] + }, + { + "test_f1": 0.6550865504926211, + "test_acc": 0.6565656565656566, + "test_auc": 0.7024489795918367, + "val_auc": 0.8896025692492975, + "k": 10, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29959, + 26366, + 31591, + 25985, + 29262, + 24895, + 3199, + 20741, + 31184, + 8250 + ] + }, + { + "test_f1": 0.6647495806374312, + "test_acc": 0.6666666666666666, + "test_auc": 0.7338775510204082, + "val_auc": 0.8851866720192694, + "k": 16, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 29959, + 26366, + 31591, + 25985, + 29262, + 24895, + 3199, + 20741, + 31184, + 8250, + 12799, + 30887, + 23695, + 25652, + 15018, + 28965 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fc51c0f952f261ea0a6952b1cd774e144a3de486 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6787860249233824, + "test_acc": 0.6788990825688074, + "test_auc": 0.7317195325542571, + "val_auc": 0.7612721417069244, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19232 + ] + }, + { + "test_f1": 0.7255836733655877, + "test_acc": 0.725604670558799, + "test_auc": 0.7760948803561493, + "val_auc": 0.8146135265700483, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19232, + 8361 + ] + }, + { + "test_f1": 0.7372708332039172, + "test_acc": 0.737281067556297, + "test_auc": 0.819794101279911, + "val_auc": 0.9045893719806763, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19232, + 8361, + 27647, + 4807, + 27712 + ] + }, + { + "test_f1": 0.7556143917138742, + "test_acc": 0.755629691409508, + "test_auc": 0.838906510851419, + "val_auc": 0.8969404186795491, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19232, + 8361, + 27647, + 4807, + 27712, + 29325, + 26813, + 8846, + 15281, + 15230 + ] + }, + { + "test_f1": 0.7721530318748848, + "test_acc": 0.7723102585487907, + "test_auc": 0.8509766277128548, + "val_auc": 0.932367149758454, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19232, + 8361, + 27647, + 4807, + 27712, + 29325, + 26813, + 8846, + 15281, + 15230, + 27660, + 21432, + 5997, + 7326, + 18177, + 13920 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..09428ec4491f05e629fb655702a71170e176f576 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22397 + ] + }, + { + "test_f1": 0.5512730566037737, + "test_acc": 0.5527044025157233, + "test_auc": 0.5691291685695452, + "val_auc": 0.5214085634253701, + "k": 2, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22397, + 28588 + ] + }, + { + "test_f1": 0.5964867959141555, + "test_acc": 0.5967295597484277, + "test_auc": 0.6346660739474592, + "val_auc": 0.6234493797519007, + "k": 5, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22397, + 28588, + 15653, + 15351, + 5535 + ] + }, + { + "test_f1": 0.6117815900397015, + "test_acc": 0.6118238993710692, + "test_auc": 0.6550500790348532, + "val_auc": 0.6202480992396958, + "k": 10, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22397, + 28588, + 15653, + 15351, + 5535, + 16611, + 30087, + 9388, + 22952, + 17748 + ] + }, + { + "test_f1": 0.6266474799064764, + "test_acc": 0.6266666666666667, + "test_auc": 0.6658357290193097, + "val_auc": 0.6390556222488996, + "k": 16, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22397, + 28588, + 15653, + 15351, + 5535, + 16611, + 30087, + 9388, + 22952, + 17748, + 26339, + 12049, + 12364, + 649, + 30430, + 22382 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6a85a3fd7cde1ee6fc16dcc588d735f14a8c9e2f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5115572776524306, + "test_acc": 0.5152603231597845, + "test_auc": 0.5145059163968714, + "val_auc": 0.5467683661180249, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1285 + ] + }, + { + "test_f1": 0.5208909313836534, + "test_acc": 0.5218432076600837, + "test_auc": 0.5251640260149557, + "val_auc": 0.5724608590927338, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1285, + 27899 + ] + }, + { + "test_f1": 0.5252993066636028, + "test_acc": 0.5266307600239377, + "test_auc": 0.536958141134, + "val_auc": 0.5503813729425934, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1285, + 27899, + 26653, + 16357, + 24507 + ] + }, + { + "test_f1": 0.5379174592409501, + "test_acc": 0.538001196888091, + "test_auc": 0.5511188150015758, + "val_auc": 0.6138097149739061, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1285, + 27899, + 26653, + 16357, + 24507, + 16461, + 18077, + 19993, + 4140, + 3865 + ] + }, + { + "test_f1": 0.5439601586874245, + "test_acc": 0.5439856373429084, + "test_auc": 0.5577701057215713, + "val_auc": 0.6407065435568046, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1285, + 27899, + 26653, + 16357, + 24507, + 16461, + 18077, + 19993, + 4140, + 3865, + 26487, + 9365, + 22562, + 2784, + 12866, + 26466 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b452e4e8fe1cc97d199dc71b8cb884c62c3e2c1b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5746440076331454, + "test_acc": 0.580377358490566, + "test_auc": 0.615825678783319, + "val_auc": 0.6578631452581032, + "k": 1, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5472 + ] + }, + { + "test_f1": 0.5863574339622193, + "test_acc": 0.5891823899371069, + "test_auc": 0.6357547651282633, + "val_auc": 0.6778711484593838, + "k": 2, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5472, + 13985 + ] + }, + { + "test_f1": 0.6180949348689766, + "test_acc": 0.6181132075471698, + "test_auc": 0.6764553096130886, + "val_auc": 0.7446978791516606, + "k": 5, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5472, + 13985, + 11157, + 24380, + 14345 + ] + }, + { + "test_f1": 0.6389911974404083, + "test_acc": 0.6389937106918239, + "test_auc": 0.706341217916457, + "val_auc": 0.7727090836334534, + "k": 10, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5472, + 13985, + 11157, + 24380, + 14345, + 30805, + 7125, + 5127, + 2586, + 31578 + ] + }, + { + "test_f1": 0.6356344254646761, + "test_acc": 0.6359748427672955, + "test_auc": 0.7070328361715335, + "val_auc": 0.8051220488195278, + "k": 16, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5472, + 13985, + 11157, + 24380, + 14345, + 30805, + 7125, + 5127, + 2586, + 31578, + 8196, + 30710, + 29072, + 7918, + 28425, + 20556 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1ed91b62c996554548fd986e90a866ad20b38ee3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9443912818385808, + "test_acc": 0.9444025157232704, + "test_auc": 0.9841003747700091, + "val_auc": 0.9935974389755902, + "k": 1, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7199 + ] + }, + { + "test_f1": 0.9738356400974403, + "test_acc": 0.9738364779874213, + "test_auc": 0.9957414340091886, + "val_auc": 1.0, + "k": 2, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7199, + 12876 + ] + }, + { + "test_f1": 0.9874213119999193, + "test_acc": 0.9874213836477987, + "test_auc": 0.9989499148894373, + "val_auc": 0.9995998399359743, + "k": 5, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7199, + 12876, + 2192, + 12282, + 25406 + ] + }, + { + "test_f1": 0.9904401971199387, + "test_acc": 0.9904402515723271, + "test_auc": 0.9996230528617097, + "val_auc": 0.9987995198079233, + "k": 10, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7199, + 12876, + 2192, + 12282, + 25406, + 27704, + 9580, + 18920, + 20872, + 6476 + ] + }, + { + "test_f1": 0.9939621724408219, + "test_acc": 0.9939622641509434, + "test_auc": 0.9997800086882646, + "val_auc": 0.9991996798719487, + "k": 16, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7199, + 12876, + 2192, + 12282, + 25406, + 27704, + 9580, + 18920, + 20872, + 6476, + 10719, + 17985, + 7542, + 21692, + 25335, + 26307 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1bab095525b70fd0369360f83105176aa4b82fef --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6084261515945389, + "test_acc": 0.609056603773585, + "test_auc": 0.646568768423323, + "val_auc": 0.6634653861544617, + "k": 1, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13509 + ] + }, + { + "test_f1": 0.632085701342505, + "test_acc": 0.6332075471698113, + "test_auc": 0.6856759581140592, + "val_auc": 0.7264905962384953, + "k": 2, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13509, + 15465 + ] + }, + { + "test_f1": 0.6709587778644674, + "test_acc": 0.6711949685534592, + "test_auc": 0.7390738492353214, + "val_auc": 0.760704281712685, + "k": 5, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13509, + 15465, + 30350, + 9496, + 26403 + ] + }, + { + "test_f1": 0.7084243144292207, + "test_acc": 0.7084276729559749, + "test_auc": 0.77924821196935, + "val_auc": 0.8147258903561425, + "k": 10, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13509, + 15465, + 30350, + 9496, + 26403, + 11937, + 8793, + 13099, + 8638, + 12896 + ] + }, + { + "test_f1": 0.7174823099132951, + "test_acc": 0.7174842767295597, + "test_auc": 0.7927086930237692, + "val_auc": 0.8523409363745499, + "k": 16, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13509, + 15465, + 30350, + 9496, + 26403, + 11937, + 8793, + 13099, + 8638, + 12896, + 31979, + 30184, + 12485, + 26943, + 14448, + 24919 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7f26517f84d80bd911f985906035358f3e6d719c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.642203091269332, + "test_acc": 0.6425157232704403, + "test_auc": 0.6783362479861554, + "val_auc": 0.7058823529411765, + "k": 1, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27345 + ] + }, + { + "test_f1": 0.6477986975576877, + "test_acc": 0.6477987421383647, + "test_auc": 0.7017900052554886, + "val_auc": 0.7611044417767107, + "k": 2, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27345, + 13193 + ] + }, + { + "test_f1": 0.6684272532601063, + "test_acc": 0.6684276729559748, + "test_auc": 0.7514012104838391, + "val_auc": 0.7803121248499401, + "k": 5, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27345, + 13193, + 12357, + 8497, + 30472 + ] + }, + { + "test_f1": 0.6948392522887357, + "test_acc": 0.6948427672955975, + "test_auc": 0.7781406607739036, + "val_auc": 0.7975190076030412, + "k": 10, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27345, + 13193, + 12357, + 8497, + 30472, + 19033, + 436, + 28059, + 10270, + 9065 + ] + }, + { + "test_f1": 0.7053825887851659, + "test_acc": 0.7054088050314465, + "test_auc": 0.7901855015346229, + "val_auc": 0.8319327731092436, + "k": 16, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27345, + 13193, + 12357, + 8497, + 30472, + 19033, + 436, + 28059, + 10270, + 9065, + 31939, + 22913, + 13503, + 7582, + 16551, + 24127 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3113420d00a4bf4990d25ae2a06d1a8cead9f4e6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7002967213903285, + "test_acc": 0.7009198423127464, + "test_auc": 0.7639553574438058, + "val_auc": 0.7203999999999999, + "k": 1, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2123 + ] + }, + { + "test_f1": 0.739340590265011, + "test_acc": 0.7400788436268069, + "test_auc": 0.8039644084026937, + "val_auc": 0.7869999999999999, + "k": 2, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2123, + 23112 + ] + }, + { + "test_f1": 0.7709112046640785, + "test_acc": 0.7710906701708279, + "test_auc": 0.8521314787156038, + "val_auc": 0.8412000000000001, + "k": 5, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2123, + 23112, + 9968, + 32138, + 21873 + ] + }, + { + "test_f1": 0.8262506067351926, + "test_acc": 0.8262812089356111, + "test_auc": 0.9048383950738029, + "val_auc": 0.9104000000000001, + "k": 10, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2123, + 23112, + 9968, + 32138, + 21873, + 7981, + 18333, + 23155, + 603, + 19878 + ] + }, + { + "test_f1": 0.8383667469489923, + "test_acc": 0.8383705650459922, + "test_auc": 0.9194072340258588, + "val_auc": 0.9276000000000001, + "k": 16, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2123, + 23112, + 9968, + 32138, + 21873, + 7981, + 18333, + 23155, + 603, + 19878, + 29205, + 9151, + 29748, + 5634, + 13854, + 4481 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..abb13c531427acdcb48f612d72d19d573ff93b0c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5349594626591904, + "test_acc": 0.5373584905660377, + "test_auc": 0.5510086183938052, + "val_auc": 0.6262505002000801, + "k": 1, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18645 + ] + }, + { + "test_f1": 0.5519980432599746, + "test_acc": 0.5567295597484276, + "test_auc": 0.5719705753393031, + "val_auc": 0.7006802721088435, + "k": 2, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18645, + 25246 + ] + }, + { + "test_f1": 0.5319427535043354, + "test_acc": 0.5388679245283019, + "test_auc": 0.5720624704442052, + "val_auc": 0.6678671468587435, + "k": 5, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18645, + 25246, + 1217, + 10948, + 25544 + ] + }, + { + "test_f1": 0.5361955579012182, + "test_acc": 0.5408805031446541, + "test_auc": 0.5711842772791758, + "val_auc": 0.6598639455782314, + "k": 10, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18645, + 25246, + 1217, + 10948, + 25544, + 10298, + 19411, + 24108, + 20742, + 13631 + ] + }, + { + "test_f1": 0.544265866202867, + "test_acc": 0.5461635220125786, + "test_auc": 0.5669034336871759, + "val_auc": 0.6634653861544618, + "k": 16, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18645, + 25246, + 1217, + 10948, + 25544, + 10298, + 19411, + 24108, + 20742, + 13631, + 22713, + 26028, + 406, + 12686, + 29676, + 14559 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..52584cc63c9ac480b12edc29ff2355ed3915d98c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7100637907547411, + "test_acc": 0.7101886792452831, + "test_auc": 0.7797717355972776, + "val_auc": 0.74749899959984, + "k": 1, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8372 + ] + }, + { + "test_f1": 0.7644902937896919, + "test_acc": 0.7645283018867924, + "test_auc": 0.8337025170651488, + "val_auc": 0.8095238095238095, + "k": 2, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8372, + 17875 + ] + }, + { + "test_f1": 0.8269044357468237, + "test_acc": 0.8269182389937106, + "test_auc": 0.9087825898521475, + "val_auc": 0.9027611044417767, + "k": 5, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8372, + 17875, + 32056, + 24191, + 28208 + ] + }, + { + "test_f1": 0.8515723082564541, + "test_acc": 0.8515723270440252, + "test_auc": 0.9302171863592222, + "val_auc": 0.9291716686674669, + "k": 10, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8372, + 17875, + 32056, + 24191, + 28208, + 6677, + 32649, + 29475, + 372, + 11751 + ] + }, + { + "test_f1": 0.8706893196297222, + "test_acc": 0.870691823899371, + "test_auc": 0.9424379695384182, + "val_auc": 0.9535814325730292, + "k": 16, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8372, + 17875, + 32056, + 24191, + 28208, + 6677, + 32649, + 29475, + 372, + 11751, + 9484, + 10320, + 11746, + 446, + 24588, + 28087 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..22a219083eb803a825021ceedb0a297d36f6298c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7567055391593854, + "test_acc": 0.7567295597484277, + "test_auc": 0.8447709406919625, + "val_auc": 0.8835534213685474, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12638 + ] + }, + { + "test_f1": 0.8191074073214824, + "test_acc": 0.8191194968553459, + "test_auc": 0.8962909312948653, + "val_auc": 0.9295718287314927, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12638, + 1960 + ] + }, + { + "test_f1": 0.868653642158207, + "test_acc": 0.8686792452830189, + "test_auc": 0.9467393186496939, + "val_auc": 0.9799919967987195, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12638, + 1960, + 5209, + 25466, + 22771 + ] + }, + { + "test_f1": 0.8995873715452428, + "test_acc": 0.8996226415094339, + "test_auc": 0.9646234224673658, + "val_auc": 0.9795918367346939, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12638, + 1960, + 5209, + 25466, + 22771, + 9484, + 2984, + 2600, + 3792, + 15069 + ] + }, + { + "test_f1": 0.9094210295130997, + "test_acc": 0.909433962264151, + "test_auc": 0.9704492683326936, + "val_auc": 0.9863945578231292, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12638, + 1960, + 5209, + 25466, + 22771, + 9484, + 2984, + 2600, + 3792, + 15069, + 17548, + 2723, + 4618, + 6347, + 15985, + 27308 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4637eecb1ea9f5a7815ba45c89a9c211be4caebf --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6281755205792489, + "test_acc": 0.6334177215189873, + "test_auc": 0.6865086201592361, + "val_auc": 0.6514605842336935, + "k": 1, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22372 + ] + }, + { + "test_f1": 0.6706260002909937, + "test_acc": 0.6708860759493671, + "test_auc": 0.72362422012478, + "val_auc": 0.7715086034413765, + "k": 2, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22372, + 22185 + ] + }, + { + "test_f1": 0.7184264398300215, + "test_acc": 0.7184810126582278, + "test_auc": 0.7882215768553955, + "val_auc": 0.847939175670268, + "k": 5, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22372, + 22185, + 24575, + 8929, + 17284 + ] + }, + { + "test_f1": 0.7964217721518987, + "test_acc": 0.7964556962025316, + "test_auc": 0.8687010078387458, + "val_auc": 0.9327731092436976, + "k": 10, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22372, + 22185, + 24575, + 8929, + 17284, + 28616, + 26627, + 727, + 10603, + 14125 + ] + }, + { + "test_f1": 0.8202526115761262, + "test_acc": 0.8202531645569621, + "test_auc": 0.9026073776913643, + "val_auc": 0.9575830332132853, + "k": 16, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22372, + 22185, + 24575, + 8929, + 17284, + 28616, + 26627, + 727, + 10603, + 14125, + 11956, + 5192, + 18018, + 31241, + 15986, + 22046 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dda7d911c6176ee5ff1ca867c45f12bef59106e4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.69007645111382, + "test_acc": 0.6902564102564103, + "test_auc": 0.7619963308311172, + "val_auc": 0.6509587923296614, + "k": 1, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26907 + ] + }, + { + "test_f1": 0.7321651320901117, + "test_acc": 0.7323076923076923, + "test_auc": 0.8038888140842226, + "val_auc": 0.7376580987352102, + "k": 2, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26907, + 31065 + ] + }, + { + "test_f1": 0.7825613577851706, + "test_acc": 0.7825641025641026, + "test_auc": 0.857003399872084, + "val_auc": 0.8241534067727458, + "k": 5, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26907, + 31065, + 5384, + 19340, + 4100 + ] + }, + { + "test_f1": 0.8102564102564103, + "test_acc": 0.8102564102564103, + "test_auc": 0.8790689063183762, + "val_auc": 0.8718890248878008, + "k": 10, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26907, + 31065, + 5384, + 19340, + 4100, + 26738, + 25282, + 1994, + 21934, + 22976 + ] + }, + { + "test_f1": 0.8225349804669749, + "test_acc": 0.8225641025641026, + "test_auc": 0.8889992257717035, + "val_auc": 0.8669930640554876, + "k": 16, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26907, + 31065, + 5384, + 19340, + 4100, + 26738, + 25282, + 1994, + 21934, + 22976, + 10152, + 9884, + 18132, + 20484, + 3112, + 31496 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..917d81c606d1bc3c92fa7723b60941d7177b95e4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6783553591865811, + "test_acc": 0.6784905660377358, + "test_auc": 0.7424823222171478, + "val_auc": 0.7242897158863546, + "k": 1, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14903 + ] + }, + { + "test_f1": 0.7154239342931071, + "test_acc": 0.7154716981132075, + "test_auc": 0.7886729030448417, + "val_auc": 0.7783113245298119, + "k": 2, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14903, + 25892 + ] + }, + { + "test_f1": 0.7448298872103124, + "test_acc": 0.7449056603773585, + "test_auc": 0.8288571388066699, + "val_auc": 0.8299319727891157, + "k": 5, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14903, + 25892, + 14220, + 23967, + 8877 + ] + }, + { + "test_f1": 0.7546503664048292, + "test_acc": 0.7547169811320755, + "test_auc": 0.8410042540092089, + "val_auc": 0.8519407763105241, + "k": 10, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14903, + 25892, + 14220, + 23967, + 8877, + 24869, + 19415, + 22481, + 25020, + 26167 + ] + }, + { + "test_f1": 0.7791064953846866, + "test_acc": 0.7791194968553459, + "test_auc": 0.8628859214674053, + "val_auc": 0.8467386954781914, + "k": 16, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14903, + 25892, + 14220, + 23967, + 8877, + 24869, + 19415, + 22481, + 25020, + 26167, + 9615, + 1661, + 18130, + 4506, + 11011, + 31563 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..241f0784b33e3f62813206074b47fee1dee6dc16 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6888851795486899, + "test_acc": 0.6892307692307692, + "test_auc": 0.7562064496583297, + "val_auc": 0.8376172990616075, + "k": 1, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10949 + ] + }, + { + "test_f1": 0.7126656095354029, + "test_acc": 0.7138461538461538, + "test_auc": 0.7788063419396103, + "val_auc": 0.8551611587107303, + "k": 2, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10949, + 8496 + ] + }, + { + "test_f1": 0.7762054487098496, + "test_acc": 0.7764102564102564, + "test_auc": 0.8479525027771233, + "val_auc": 0.9106487148102815, + "k": 5, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10949, + 8496, + 21895, + 28768, + 28602 + ] + }, + { + "test_f1": 0.8102524182136306, + "test_acc": 0.8102564102564103, + "test_auc": 0.8805374154239742, + "val_auc": 0.9298245614035088, + "k": 10, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10949, + 8496, + 21895, + 28768, + 28602, + 24020, + 21255, + 26449, + 32534, + 12943 + ] + }, + { + "test_f1": 0.8255397241369186, + "test_acc": 0.8256410256410256, + "test_auc": 0.8923990978557244, + "val_auc": 0.9298245614035089, + "k": 16, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10949, + 8496, + 21895, + 28768, + 28602, + 24020, + 21255, + 26449, + 32534, + 12943, + 1628, + 27076, + 25266, + 12770, + 18900, + 15573 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6edf7ffc8367b9bfb4d13d24d342ca2ef9726200 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6727472416118107, + "test_acc": 0.6738461538461539, + "test_auc": 0.7172636079038611, + "val_auc": 0.7103223174214606, + "k": 1, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26723 + ] + }, + { + "test_f1": 0.698558700155, + "test_acc": 0.7005128205128205, + "test_auc": 0.7810112094792473, + "val_auc": 0.7723378212974296, + "k": 2, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26723, + 14135 + ] + }, + { + "test_f1": 0.7804421475330291, + "test_acc": 0.7805128205128206, + "test_auc": 0.8511335712121721, + "val_auc": 0.835577315381477, + "k": 5, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26723, + 14135, + 31625, + 952, + 20126 + ] + }, + { + "test_f1": 0.8437916815035966, + "test_acc": 0.8441025641025641, + "test_auc": 0.9041345474130676, + "val_auc": 0.8780089759281925, + "k": 10, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26723, + 14135, + 31625, + 952, + 20126, + 14332, + 21635, + 6437, + 19351, + 8081 + ] + }, + { + "test_f1": 0.8376255636682123, + "test_acc": 0.837948717948718, + "test_auc": 0.9033771501666275, + "val_auc": 0.8935128518971848, + "k": 16, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26723, + 14135, + 31625, + 952, + 20126, + 14332, + 21635, + 6437, + 19351, + 8081, + 31432, + 2600, + 17221, + 18848, + 9712, + 3348 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..22cc0aa3890981f0b398e762f4ea37959fca1204 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7062845355587118, + "test_acc": 0.7076923076923077, + "test_auc": 0.7654277611337394, + "val_auc": 0.6852305181558548, + "k": 1, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2984 + ] + }, + { + "test_f1": 0.7445025044675654, + "test_acc": 0.7446153846153846, + "test_auc": 0.8232950146430135, + "val_auc": 0.8325173398612811, + "k": 2, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2984, + 25967 + ] + }, + { + "test_f1": 0.8287078580505725, + "test_acc": 0.8287179487179487, + "test_auc": 0.898437237014845, + "val_auc": 0.860873113015096, + "k": 5, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2984, + 25967, + 8639, + 27249, + 5689 + ] + }, + { + "test_f1": 0.8635854385665503, + "test_acc": 0.8635897435897436, + "test_auc": 0.9303741542397415, + "val_auc": 0.9184006527947777, + "k": 10, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2984, + 25967, + 8639, + 27249, + 5689, + 9719, + 12075, + 19811, + 2816, + 15985 + ] + }, + { + "test_f1": 0.8738342090302322, + "test_acc": 0.8738461538461538, + "test_auc": 0.9405190695795604, + "val_auc": 0.9192166462668299, + "k": 16, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2984, + 25967, + 8639, + 27249, + 5689, + 9719, + 12075, + 19811, + 2816, + 15985, + 10924, + 9328, + 17548, + 1960, + 25360, + 23504 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..230a6068f626a75b2a87cbf60b70ae9615a969c3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6763647711315054, + "test_acc": 0.676923076923077, + "test_auc": 0.741752785538762, + "val_auc": 0.8002855977152183, + "k": 1, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 914 + ] + }, + { + "test_f1": 0.7507141625194841, + "test_acc": 0.7507692307692307, + "test_auc": 0.7998788164405696, + "val_auc": 0.8392492860057119, + "k": 2, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 914, + 25374 + ] + }, + { + "test_f1": 0.7815370826827127, + "test_acc": 0.7815384615384615, + "test_auc": 0.8565615848116606, + "val_auc": 0.8331293349653203, + "k": 5, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 914, + 25374, + 31659, + 7810, + 9642 + ] + }, + { + "test_f1": 0.8071754303624701, + "test_acc": 0.8071794871794872, + "test_auc": 0.8860032988857846, + "val_auc": 0.8882088943288454, + "k": 10, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 914, + 25374, + 31659, + 7810, + 9642, + 664, + 12641, + 8922, + 27100, + 21727 + ] + }, + { + "test_f1": 0.8133321551499237, + "test_acc": 0.8133333333333334, + "test_auc": 0.8926726024169389, + "val_auc": 0.8812729498164015, + "k": 16, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 914, + 25374, + 31659, + 7810, + 9642, + 664, + 12641, + 8922, + 27100, + 21727, + 21437, + 9363, + 17593, + 30573, + 27240, + 22950 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cb9af288966aa2f19f9974e1a1095e931455c91c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7817233053947233, + "test_acc": 0.7818181818181819, + "test_auc": 0.8543348157909186, + "val_auc": 0.868080808080808, + "k": 1, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22124 + ] + }, + { + "test_f1": 0.8405793465745001, + "test_acc": 0.8406417112299466, + "test_auc": 0.896996650743974, + "val_auc": 0.9002020202020202, + "k": 2, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22124, + 14502 + ] + }, + { + "test_f1": 0.8554989428060759, + "test_acc": 0.8556149732620321, + "test_auc": 0.9212490162704297, + "val_auc": 0.9373737373737374, + "k": 5, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22124, + 14502, + 23590, + 30474, + 29233 + ] + }, + { + "test_f1": 0.8715119613202587, + "test_acc": 0.8716577540106952, + "test_auc": 0.9325939347352625, + "val_auc": 0.945858585858586, + "k": 10, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22124, + 14502, + 23590, + 30474, + 29233, + 25575, + 10867, + 25410, + 10200, + 16795 + ] + }, + { + "test_f1": 0.8780104002501677, + "test_acc": 0.8780748663101604, + "test_auc": 0.9391734841413641, + "val_auc": 0.9454545454545455, + "k": 16, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22124, + 14502, + 23590, + 30474, + 29233, + 25575, + 10867, + 25410, + 10200, + 16795, + 3604, + 31458, + 11078, + 6946, + 23058, + 28627 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a3006f06b3803fb64bddc2598cc1465213c2ff32 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8333979742408861, + "test_acc": 0.8335632183908046, + "test_auc": 0.9031189120082256, + "val_auc": 0.882711038961039, + "k": 1, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22854 + ] + }, + { + "test_f1": 0.8776771016428161, + "test_acc": 0.8777011494252873, + "test_auc": 0.9416871854537583, + "val_auc": 0.9543425324675325, + "k": 2, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22854, + 24800 + ] + }, + { + "test_f1": 0.9158601124866035, + "test_acc": 0.9158620689655173, + "test_auc": 0.9704136283078089, + "val_auc": 0.9744318181818182, + "k": 5, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22854, + 24800, + 18373, + 17819, + 3827 + ] + }, + { + "test_f1": 0.949423876119365, + "test_acc": 0.9494252873563218, + "test_auc": 0.9848290627198442, + "val_auc": 1.0, + "k": 10, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22854, + 24800, + 18373, + 17819, + 3827, + 28039, + 10409, + 25524, + 27732, + 14903 + ] + }, + { + "test_f1": 0.95678060422363, + "test_acc": 0.9567816091954023, + "test_auc": 0.9908857689810054, + "val_auc": 1.0, + "k": 16, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22854, + 24800, + 18373, + 17819, + 3827, + 28039, + 10409, + 25524, + 27732, + 14903, + 16260, + 13625, + 23839, + 17256, + 29120, + 3672 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fdebc546f37f6fe1868ce4c16e4252ac9915b185 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8431427765087466, + "test_acc": 0.8432871476349737, + "test_auc": 0.9158060633951871, + "val_auc": 0.9081939081939081, + "k": 1, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16345 + ] + }, + { + "test_f1": 0.9383594809666759, + "test_acc": 0.9383659818442427, + "test_auc": 0.9833011006590805, + "val_auc": 0.9982839982839983, + "k": 2, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16345, + 27140 + ] + }, + { + "test_f1": 0.9708550321310305, + "test_acc": 0.9708552317247969, + "test_auc": 0.9934174122184664, + "val_auc": 1.0, + "k": 5, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16345, + 27140, + 9907, + 17605, + 3615 + ] + }, + { + "test_f1": 0.9765884043013552, + "test_acc": 0.9765886287625418, + "test_auc": 0.9953824183088895, + "val_auc": 0.9982839982839983, + "k": 10, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16345, + 27140, + 9907, + 17605, + 3615, + 26578, + 4395, + 29567, + 9656, + 27282 + ] + }, + { + "test_f1": 0.9827988665017662, + "test_acc": 0.9827998088867654, + "test_auc": 0.9964197077692616, + "val_auc": 0.9991419991419992, + "k": 16, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16345, + 27140, + 9907, + 17605, + 3615, + 26578, + 4395, + 29567, + 9656, + 27282, + 10020, + 25627, + 19064, + 3347, + 16833, + 17343 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2ceb2d1953ec7d1a60db811a57b63d1665af7a9b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9512608442101776, + "test_acc": 0.9512643678160919, + "test_auc": 0.9883279668813247, + "val_auc": 0.9756493506493505, + "k": 1, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19157 + ] + }, + { + "test_f1": 0.9641373852052999, + "test_acc": 0.9641379310344828, + "test_auc": 0.995671606959251, + "val_auc": 0.9902597402597403, + "k": 2, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19157, + 13068 + ] + }, + { + "test_f1": 0.9848275797923417, + "test_acc": 0.9848275862068966, + "test_auc": 0.9989937902483901, + "val_auc": 0.9983766233766234, + "k": 5, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19157, + 13068, + 10563, + 8322, + 7706 + ] + }, + { + "test_f1": 0.987586059943069, + "test_acc": 0.9875862068965517, + "test_auc": 0.999561157259592, + "val_auc": 1.0, + "k": 10, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19157, + 13068, + 10563, + 8322, + 7706, + 2814, + 16865, + 22232, + 14397, + 25577 + ] + }, + { + "test_f1": 0.9926436750508345, + "test_acc": 0.9926436781609196, + "test_auc": 0.9998029858217437, + "val_auc": 1.0, + "k": 16, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19157, + 13068, + 10563, + 8322, + 7706, + 2814, + 16865, + 22232, + 14397, + 25577, + 22140, + 21398, + 26196, + 17658, + 18460, + 12515 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3eb59e903345a7f1cb7c981a3642b53cdc02fa40 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8537908781446845, + "test_acc": 0.8537931034482759, + "test_auc": 0.9330261716001947, + "val_auc": 0.939935064935065, + "k": 1, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30753 + ] + }, + { + "test_f1": 0.8960559186638465, + "test_acc": 0.8960919540229885, + "test_auc": 0.9662657611342607, + "val_auc": 0.9614448051948052, + "k": 2, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30753, + 17029 + ] + }, + { + "test_f1": 0.9526436180973611, + "test_acc": 0.9526436781609195, + "test_auc": 0.9899793346501433, + "val_auc": 0.976461038961039, + "k": 5, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30753, + 17029, + 18293, + 12419, + 16539 + ] + }, + { + "test_f1": 0.9595402298850575, + "test_acc": 0.9595402298850575, + "test_auc": 0.9921363439580063, + "val_auc": 0.9801136363636364, + "k": 10, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30753, + 17029, + 18293, + 12419, + 16539, + 30260, + 10996, + 5332, + 7492, + 32549 + ] + }, + { + "test_f1": 0.9613792613609029, + "test_acc": 0.9613793103448276, + "test_auc": 0.9933522512040696, + "val_auc": 0.9809253246753247, + "k": 16, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30753, + 17029, + 18293, + 12419, + 16539, + 30260, + 10996, + 5332, + 7492, + 32549, + 8156, + 21159, + 16356, + 7677, + 27628, + 5256 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9817a3fd82d259b48ce94dc09c83cfcf249bfa72 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9103334549548772, + "test_acc": 0.9103448275862069, + "test_auc": 0.9637172601331241, + "val_auc": 0.9809253246753247, + "k": 1, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3947 + ] + }, + { + "test_f1": 0.926430871997599, + "test_acc": 0.9264367816091954, + "test_auc": 0.9771615753016938, + "val_auc": 0.9853896103896105, + "k": 2, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3947, + 2421 + ] + }, + { + "test_f1": 0.9737903000033774, + "test_acc": 0.9737931034482759, + "test_auc": 0.9953223929866335, + "val_auc": 0.987012987012987, + "k": 5, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3947, + 2421, + 9398, + 25886, + 29768 + ] + }, + { + "test_f1": 0.9797700636260362, + "test_acc": 0.9797701149425287, + "test_auc": 0.9973356580442665, + "val_auc": 0.9931006493506493, + "k": 10, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3947, + 2421, + 9398, + 25886, + 29768, + 5161, + 21362, + 15413, + 2088, + 26434 + ] + }, + { + "test_f1": 0.9908045938135429, + "test_acc": 0.9908045977011494, + "test_auc": 0.9992280088208236, + "val_auc": 0.994724025974026, + "k": 16, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3947, + 2421, + 9398, + 25886, + 29768, + 5161, + 21362, + 15413, + 2088, + 26434, + 19325, + 460, + 30603, + 7233, + 193, + 10333 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..560d0e4a0321f8e5816ccc4b813bd5ab9e0d4dd5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8037869561656528, + "test_acc": 0.8040251572327044, + "test_auc": 0.8752785965921346, + "val_auc": 0.8871548619447778, + "k": 1, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6588 + ] + }, + { + "test_f1": 0.8148216014810898, + "test_acc": 0.8148427672955975, + "test_auc": 0.8957815843222393, + "val_auc": 0.9279711884753902, + "k": 2, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6588, + 17771 + ] + }, + { + "test_f1": 0.8572161630717212, + "test_acc": 0.8573584905660377, + "test_auc": 0.9345856214286221, + "val_auc": 0.9683873549419768, + "k": 5, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6588, + 17771, + 24464, + 13633, + 398 + ] + }, + { + "test_f1": 0.8711394625576627, + "test_acc": 0.8711949685534591, + "test_auc": 0.9426660617960405, + "val_auc": 0.9627851140456183, + "k": 10, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6588, + 17771, + 24464, + 13633, + 398, + 16455, + 12692, + 14074, + 19678, + 29051 + ] + }, + { + "test_f1": 0.8937513742510691, + "test_acc": 0.8938364779874214, + "test_auc": 0.953089953915744, + "val_auc": 0.9771908763505402, + "k": 16, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6588, + 17771, + 24464, + 13633, + 398, + 16455, + 12692, + 14074, + 19678, + 29051, + 9070, + 26677, + 24721, + 16352, + 14006, + 20838 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..71bb7eaf038feb5b3f4e29b846b6d1a20b1ceae6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9475737726374909, + "test_acc": 0.9475862068965517, + "test_auc": 0.9824454448292657, + "val_auc": 0.9823457792207791, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7072 + ] + }, + { + "test_f1": 0.9742487582871985, + "test_acc": 0.9742528735632184, + "test_auc": 0.996217835110125, + "val_auc": 0.9983766233766234, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7072, + 23306 + ] + }, + { + "test_f1": 0.9802249106223324, + "test_acc": 0.9802298850574712, + "test_auc": 0.9993353942312896, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7072, + 23306, + 25059, + 21598, + 15824 + ] + }, + { + "test_f1": 0.9917240854480806, + "test_acc": 0.9917241379310345, + "test_auc": 0.9996896815303857, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7072, + 23306, + 25059, + 21598, + 15824, + 1554, + 6365, + 532, + 827, + 24058 + ] + }, + { + "test_f1": 0.9917240854480806, + "test_acc": 0.9917241379310345, + "test_auc": 0.9997793102981763, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7072, + 23306, + 25059, + 21598, + 15824, + 1554, + 6365, + 532, + 827, + 24058, + 14701, + 18975, + 19679, + 14876, + 17981, + 22736 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..09c1a0002ba0c98c5ffcfcacd868cc7bec1af858 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8333729463963617, + "test_acc": 0.8335632183908046, + "test_auc": 0.8985326248714758, + "val_auc": 0.9062500000000001, + "k": 1, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26284 + ] + }, + { + "test_f1": 0.9217805483619173, + "test_acc": 0.9218390804597701, + "test_auc": 0.9736013684452621, + "val_auc": 0.9914772727272727, + "k": 2, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26284, + 2003 + ] + }, + { + "test_f1": 0.9742525687717765, + "test_acc": 0.9742528735632184, + "test_auc": 0.9941741300936198, + "val_auc": 0.9995941558441559, + "k": 5, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26284, + 2003, + 8213, + 2042, + 4694 + ] + }, + { + "test_f1": 0.9894252873563218, + "test_acc": 0.9894252873563218, + "test_auc": 0.9968815953244222, + "val_auc": 1.0, + "k": 10, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26284, + 2003, + 8213, + 2042, + 4694, + 9438, + 17692, + 26298, + 26033, + 20532 + ] + }, + { + "test_f1": 0.9903448030942082, + "test_acc": 0.9903448275862069, + "test_auc": 0.9970574706423507, + "val_auc": 1.0, + "k": 16, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26284, + 2003, + 8213, + 2042, + 4694, + 9438, + 17692, + 26298, + 26033, + 20532, + 12534, + 4175, + 9658, + 6859, + 1400, + 3635 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8961461188f6078e5c153f7eeefa20440a6e5961 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7565934088832684, + "test_acc": 0.7586206896551724, + "test_auc": 0.8242126197305049, + "val_auc": 0.8498376623376624, + "k": 1, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25233 + ] + }, + { + "test_f1": 0.8592966035818567, + "test_acc": 0.8593103448275862, + "test_auc": 0.9221523418475026, + "val_auc": 0.9285714285714286, + "k": 2, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25233, + 21252 + ] + }, + { + "test_f1": 0.9144758886972089, + "test_acc": 0.9144827586206896, + "test_auc": 0.969383320255425, + "val_auc": 0.9898538961038961, + "k": 5, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25233, + 21252, + 20834, + 10079, + 15413 + ] + }, + { + "test_f1": 0.9411436773152865, + "test_acc": 0.9411494252873563, + "test_auc": 0.984878950430218, + "val_auc": 0.994724025974026, + "k": 10, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25233, + 21252, + 20834, + 10079, + 15413, + 4784, + 4628, + 17996, + 4375, + 19746 + ] + }, + { + "test_f1": 0.9526402543064824, + "test_acc": 0.9526436781609195, + "test_auc": 0.9919765341739273, + "val_auc": 0.9987824675324676, + "k": 16, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25233, + 21252, + 20834, + 10079, + 15413, + 4784, + 4628, + 17996, + 4375, + 19746, + 30345, + 23641, + 853, + 18532, + 31591, + 26634 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3a22092a14e94a62b7a80d285ff4b2ab0e6d4ca7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8000743484577604, + "test_acc": 0.8004597701149425, + "test_auc": 0.8729262778018292, + "val_auc": 0.8303571428571428, + "k": 1, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 24817 + ] + }, + { + "test_f1": 0.9218265876575663, + "test_acc": 0.9218390804597701, + "test_auc": 0.9640284241571514, + "val_auc": 0.9707792207792209, + "k": 2, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 24817, + 9997 + ] + }, + { + "test_f1": 0.952642957389076, + "test_acc": 0.9526436781609195, + "test_auc": 0.9816032726338006, + "val_auc": 0.9878246753246753, + "k": 5, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 24817, + 9997, + 27799, + 5345, + 15329 + ] + }, + { + "test_f1": 0.9577001658358933, + "test_acc": 0.9577011494252874, + "test_auc": 0.9879187185453758, + "val_auc": 0.9922889610389611, + "k": 10, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 24817, + 9997, + 27799, + 5345, + 15329, + 26519, + 18917, + 16810, + 11013, + 780 + ] + }, + { + "test_f1": 0.9613775958291783, + "test_acc": 0.9613793103448276, + "test_auc": 0.9886196831538503, + "val_auc": 0.9959415584415584, + "k": 16, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 24817, + 9997, + 27799, + 5345, + 15329, + 26519, + 18917, + 16810, + 11013, + 780, + 518, + 3737, + 9364, + 14927, + 3495, + 19410 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..49f8945e048a864bffd8e7d26bc6de74e66b4449 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8937162217190885, + "test_acc": 0.8937931034482759, + "test_auc": 0.9459462430326316, + "val_auc": 0.9468344155844155, + "k": 1, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19694 + ] + }, + { + "test_f1": 0.9117155159729298, + "test_acc": 0.9117241379310345, + "test_auc": 0.9612131507386763, + "val_auc": 0.9813311688311689, + "k": 2, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19694, + 18972 + ] + }, + { + "test_f1": 0.9420689655172414, + "test_acc": 0.9420689655172414, + "test_auc": 0.9836512054223714, + "val_auc": 0.9955357142857143, + "k": 5, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19694, + 18972, + 15795, + 12495, + 28005 + ] + }, + { + "test_f1": 0.9540145314599284, + "test_acc": 0.9540229885057471, + "test_auc": 0.9880337139455598, + "val_auc": 0.9983766233766234, + "k": 10, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19694, + 18972, + 15795, + 12495, + 28005, + 22043, + 9650, + 21081, + 11030, + 22639 + ] + }, + { + "test_f1": 0.9599996448645189, + "test_acc": 0.96, + "test_auc": 0.9918116510633693, + "val_auc": 0.9967532467532468, + "k": 16, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19694, + 18972, + 15795, + 12495, + 28005, + 22043, + 9650, + 21081, + 11030, + 22639, + 1439, + 30960, + 25689, + 8383, + 1107, + 24089 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b47217aa10a56d852c03d4f946a7443884f129fb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8878070797963384, + "test_acc": 0.887816091954023, + "test_auc": 0.9527774771362086, + "val_auc": 0.9508928571428572, + "k": 1, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2303 + ] + }, + { + "test_f1": 0.9342522898347841, + "test_acc": 0.9342528735632184, + "test_auc": 0.9812481397802911, + "val_auc": 0.9801136363636364, + "k": 2, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2303, + 3047 + ] + }, + { + "test_f1": 0.9567725445676236, + "test_acc": 0.9567816091954023, + "test_auc": 0.9877487621083392, + "val_auc": 0.9943181818181818, + "k": 5, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2303, + 3047, + 27837, + 28006, + 13111 + ] + }, + { + "test_f1": 0.9668896086049882, + "test_acc": 0.9668965517241379, + "test_auc": 0.9894686197846203, + "val_auc": 0.9967532467532467, + "k": 10, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2303, + 3047, + 27837, + 28006, + 13111, + 4191, + 19284, + 5243, + 7574, + 9332 + ] + }, + { + "test_f1": 0.9664251420686641, + "test_acc": 0.9664367816091954, + "test_auc": 0.9904672195465123, + "val_auc": 0.9935064935064934, + "k": 16, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2303, + 3047, + 27837, + 28006, + 13111, + 4191, + 19284, + 5243, + 7574, + 9332, + 17500, + 22645, + 21838, + 15822, + 14588, + 23079 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..769c6ad523120b8443cf7fc9bb9800e3d44817cb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7772477864050259, + "test_acc": 0.7788505747126436, + "test_auc": 0.8318978637913307, + "val_auc": 0.7660308441558441, + "k": 1, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30973 + ] + }, + { + "test_f1": 0.8510707971533675, + "test_acc": 0.8519540229885058, + "test_auc": 0.9226799678012879, + "val_auc": 0.9297889610389611, + "k": 2, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30973, + 13335 + ] + }, + { + "test_f1": 0.9245617255529028, + "test_acc": 0.9245977011494253, + "test_auc": 0.9732128361924346, + "val_auc": 0.9963474025974026, + "k": 5, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30973, + 13335, + 15507, + 200, + 13166 + ] + }, + { + "test_f1": 0.9443218441481841, + "test_acc": 0.944367816091954, + "test_auc": 0.9866977379728341, + "val_auc": 1.0, + "k": 10, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30973, + 13335, + 15507, + 200, + 13166, + 28228, + 30039, + 26428, + 28721, + 27333 + ] + }, + { + "test_f1": 0.9466367514208895, + "test_acc": 0.9466666666666667, + "test_auc": 0.9896351940040045, + "val_auc": 0.9987824675324676, + "k": 16, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 30973, + 13335, + 15507, + 200, + 13166, + 28228, + 30039, + 26428, + 28721, + 27333, + 27616, + 6037, + 1981, + 29258, + 1505, + 31018 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..41a1573e8e18bad9a395789b3fa4ab6a55f67350 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8762260786245778, + "test_acc": 0.8763218390804598, + "test_auc": 0.9212175814438012, + "val_auc": 0.942775974025974, + "k": 1, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14876 + ] + }, + { + "test_f1": 0.8560131936136478, + "test_acc": 0.8583908045977011, + "test_auc": 0.9549078514800584, + "val_auc": 0.989448051948052, + "k": 2, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14876, + 14698 + ] + }, + { + "test_f1": 0.9839079031064156, + "test_acc": 0.9839080459770115, + "test_auc": 0.9968274798419828, + "val_auc": 1.0, + "k": 5, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14876, + 14698, + 19125, + 31068, + 32645 + ] + }, + { + "test_f1": 0.9889654705898647, + "test_acc": 0.9889655172413793, + "test_auc": 0.9983739988635748, + "val_auc": 1.0, + "k": 10, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14876, + 14698, + 19125, + 31068, + 32645, + 17067, + 9929, + 22626, + 25178, + 17585 + ] + }, + { + "test_f1": 0.9954022288724742, + "test_acc": 0.9954022988505747, + "test_auc": 0.9994072663564045, + "val_auc": 1.0, + "k": 16, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14876, + 14698, + 19125, + 31068, + 32645, + 17067, + 9929, + 22626, + 25178, + 17585, + 19176, + 16480, + 3629, + 14207, + 16715, + 25759 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a80746cbda88ae64c17c5d9fcde2ecdf95ad5bc3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8225023957660248, + "test_acc": 0.8225287356321839, + "test_auc": 0.9028753923372477, + "val_auc": 0.9025974025974026, + "k": 1, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7502 + ] + }, + { + "test_f1": 0.8864256897185104, + "test_acc": 0.8864367816091954, + "test_auc": 0.9452723361653769, + "val_auc": 0.9159902597402597, + "k": 2, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7502, + 19463 + ] + }, + { + "test_f1": 0.9163208485082732, + "test_acc": 0.9163218390804597, + "test_auc": 0.9779919097353753, + "val_auc": 0.9776785714285714, + "k": 5, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7502, + 19463, + 9261, + 8998, + 2946 + ] + }, + { + "test_f1": 0.9379304834104605, + "test_acc": 0.9379310344827586, + "test_auc": 0.9853448509118459, + "val_auc": 0.9910714285714286, + "k": 10, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7502, + 19463, + 9261, + 8998, + 2946, + 2532, + 13256, + 11567, + 9540, + 6092 + ] + }, + { + "test_f1": 0.9517158501260109, + "test_acc": 0.9517241379310345, + "test_auc": 0.9892453934195574, + "val_auc": 0.997564935064935, + "k": 16, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7502, + 19463, + 9261, + 8998, + 2946, + 2532, + 13256, + 11567, + 9540, + 6092, + 19296, + 16427, + 4469, + 24534, + 32419, + 17014 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2156b18ef7300a7994f61a694e7005e7f9172ba4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.846651096827726, + "test_acc": 0.8473563218390805, + "test_auc": 0.9013673460414524, + "val_auc": 0.8555194805194806, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2987 + ] + }, + { + "test_f1": 0.8758183150016509, + "test_acc": 0.8763218390804598, + "test_auc": 0.9346800760322528, + "val_auc": 0.9034090909090909, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2987, + 30207 + ] + }, + { + "test_f1": 0.9282750130070915, + "test_acc": 0.9282758620689655, + "test_auc": 0.9762619054061367, + "val_auc": 0.9768668831168831, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2987, + 30207, + 942, + 27585, + 32048 + ] + }, + { + "test_f1": 0.9563218206143173, + "test_acc": 0.9563218390804598, + "test_auc": 0.9890044104118189, + "val_auc": 0.9955357142857143, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2987, + 30207, + 942, + 27585, + 32048, + 25029, + 22371, + 18816, + 10640, + 13475 + ] + }, + { + "test_f1": 0.9664362707708986, + "test_acc": 0.9664367816091954, + "test_auc": 0.9909424211266843, + "val_auc": 0.9959415584415584, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2987, + 30207, + 942, + 27585, + 32048, + 25029, + 22371, + 18816, + 10640, + 13475, + 25732, + 19272, + 23552, + 2360, + 28167, + 12985 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..286b0165788b7ad09bc3377e5cbb3422a54e7f94 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.588825281071327, + "test_acc": 0.5919496855345912, + "test_auc": 0.6187823974546828, + "val_auc": 0.6066426570628252, + "k": 1, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2354 + ] + }, + { + "test_f1": 0.6135407335257527, + "test_acc": 0.6135849056603774, + "test_auc": 0.6417012391409351, + "val_auc": 0.6394557823129252, + "k": 2, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2354, + 21736 + ] + }, + { + "test_f1": 0.6103117524983919, + "test_acc": 0.6103144654088051, + "test_auc": 0.6423793136271073, + "val_auc": 0.640656262505002, + "k": 5, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2354, + 21736, + 28041, + 10918, + 26526 + ] + }, + { + "test_f1": 0.6281447529861074, + "test_acc": 0.6281761006289308, + "test_auc": 0.6684820802013896, + "val_auc": 0.6690676270508203, + "k": 10, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2354, + 21736, + 28041, + 10918, + 26526, + 20933, + 13220, + 10594, + 31036, + 15418 + ] + }, + { + "test_f1": 0.6341345993543928, + "test_acc": 0.6342138364779875, + "test_auc": 0.6740205956422987, + "val_auc": 0.6830732292917167, + "k": 16, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2354, + 21736, + 28041, + 10918, + 26526, + 20933, + 13220, + 10594, + 31036, + 15418, + 6606, + 18244, + 13330, + 6770, + 12172, + 32437 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ade726a73f2ef77f994f2e9d924bded7a0eea251 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8444342611544348, + "test_acc": 0.8445977011494253, + "test_auc": 0.918288158179555, + "val_auc": 0.9107142857142857, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28197 + ] + }, + { + "test_f1": 0.9066575506593426, + "test_acc": 0.9066666666666666, + "test_auc": 0.9608457573461768, + "val_auc": 0.9809253246753247, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28197, + 10549 + ] + }, + { + "test_f1": 0.9535628061070852, + "test_acc": 0.9535632183908046, + "test_auc": 0.9856035905622599, + "val_auc": 1.0, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28197, + 10549, + 10533, + 16741, + 8509 + ] + }, + { + "test_f1": 0.9687347597828802, + "test_acc": 0.968735632183908, + "test_auc": 0.9899970912928189, + "val_auc": 0.9995941558441559, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28197, + 10549, + 10533, + 16741, + 8509, + 15296, + 13169, + 28391, + 32346, + 15029 + ] + }, + { + "test_f1": 0.9691953892753606, + "test_acc": 0.9691954022988506, + "test_auc": 0.991300090643433, + "val_auc": 1.0, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28197, + 10549, + 10533, + 16741, + 8509, + 15296, + 13169, + 28391, + 32346, + 15029, + 16209, + 16283, + 1743, + 8887, + 22051, + 31106 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..949a033b91f3455ceeddad190266a065e9a85976 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8151560026796654, + "test_acc": 0.8151724137931035, + "test_auc": 0.8962166513339467, + "val_auc": 0.9598214285714285, + "k": 1, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 31486 + ] + }, + { + "test_f1": 0.8891951212240097, + "test_acc": 0.8891954022988505, + "test_auc": 0.95539785026246, + "val_auc": 0.9906655844155844, + "k": 2, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 31486, + 19442 + ] + }, + { + "test_f1": 0.9236776769100759, + "test_acc": 0.9236781609195402, + "test_auc": 0.980398357595108, + "val_auc": 1.0, + "k": 5, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 31486, + 19442, + 6694, + 21966, + 11793 + ] + }, + { + "test_f1": 0.9618390320590385, + "test_acc": 0.9618390804597701, + "test_auc": 0.9930512338329996, + "val_auc": 0.9987824675324676, + "k": 10, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 31486, + 19442, + 6694, + 21966, + 11793, + 4010, + 17734, + 7308, + 11367, + 938 + ] + }, + { + "test_f1": 0.9714941805639601, + "test_acc": 0.9714942528735632, + "test_auc": 0.9951346799069214, + "val_auc": 0.9991883116883117, + "k": 16, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 31486, + 19442, + 6694, + 21966, + 11793, + 4010, + 17734, + 7308, + 11367, + 938, + 10752, + 4578, + 25445, + 4052, + 25732, + 17797 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d1f0f0311f143762443b5a20a071c29a3f1d5c2d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7648072998228598, + "test_acc": 0.7655172413793103, + "test_auc": 0.8381980897234699, + "val_auc": 0.8547077922077921, + "k": 1, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10009 + ] + }, + { + "test_f1": 0.8353830903755175, + "test_acc": 0.8354022988505747, + "test_auc": 0.9112455354726987, + "val_auc": 0.9642857142857143, + "k": 2, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10009, + 7027 + ] + }, + { + "test_f1": 0.9172051507554179, + "test_acc": 0.9172413793103448, + "test_auc": 0.9672068631960604, + "val_auc": 0.9829545454545454, + "k": 5, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10009, + 7027, + 28081, + 14581, + 17801 + ] + }, + { + "test_f1": 0.9411337464012105, + "test_acc": 0.9411494252873563, + "test_auc": 0.9863062462795606, + "val_auc": 0.9955357142857143, + "k": 10, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10009, + 7027, + 28081, + 14581, + 17801, + 22058, + 31228, + 16671, + 10339, + 13177 + ] + }, + { + "test_f1": 0.9484949467188707, + "test_acc": 0.9485057471264368, + "test_auc": 0.9876346122625682, + "val_auc": 0.9983766233766235, + "k": 16, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10009, + 7027, + 28081, + 14581, + 17801, + 22058, + 31228, + 16671, + 10339, + 13177, + 31468, + 20194, + 1132, + 23839, + 13169, + 1463 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..763897eab124d8b9efb0a891941ee7ea86621172 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9029880956033405, + "test_acc": 0.9029885057471264, + "test_auc": 0.9515269021592077, + "val_auc": 0.9521103896103896, + "k": 1, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10881 + ] + }, + { + "test_f1": 0.9388498508306736, + "test_acc": 0.9388505747126437, + "test_auc": 0.9695668055630717, + "val_auc": 0.9679383116883117, + "k": 2, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10881, + 8062 + ] + }, + { + "test_f1": 0.9609128310543897, + "test_acc": 0.960919540229885, + "test_auc": 0.9825443746956004, + "val_auc": 0.9837662337662338, + "k": 5, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10881, + 8062, + 12004, + 6564, + 7540 + ] + }, + { + "test_f1": 0.9618258464616319, + "test_acc": 0.9618390804597701, + "test_auc": 0.9843885288706098, + "val_auc": 0.984577922077922, + "k": 10, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10881, + 8062, + 12004, + 6564, + 7540, + 19520, + 29267, + 16339, + 26339, + 5852 + ] + }, + { + "test_f1": 0.9696479736322314, + "test_acc": 0.9696551724137931, + "test_auc": 0.9875948712051518, + "val_auc": 0.9862012987012987, + "k": 16, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10881, + 8062, + 12004, + 6564, + 7540, + 19520, + 29267, + 16339, + 26339, + 5852, + 27498, + 25203, + 23380, + 2138, + 15551, + 20435 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9ae87ee0dbbebbded97de5990b1b04462b82789a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8116829377519034, + "test_acc": 0.8124137931034483, + "test_auc": 0.8830074002922237, + "val_auc": 0.9257305194805194, + "k": 1, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16716 + ] + }, + { + "test_f1": 0.8771265236767172, + "test_acc": 0.8772413793103448, + "test_auc": 0.9460874506196223, + "val_auc": 0.9659090909090908, + "k": 2, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16716, + 15714 + ] + }, + { + "test_f1": 0.9028202643639827, + "test_acc": 0.9029885057471264, + "test_auc": 0.9633113940148276, + "val_auc": 0.9756493506493507, + "k": 5, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16716, + 15714, + 2224, + 14662, + 23572 + ] + }, + { + "test_f1": 0.9207603145451091, + "test_acc": 0.9209195402298851, + "test_auc": 0.9719056090697548, + "val_auc": 0.9724025974025974, + "k": 10, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16716, + 15714, + 2224, + 14662, + 23572, + 28519, + 26931, + 6861, + 15191, + 27445 + ] + }, + { + "test_f1": 0.923994458237231, + "test_acc": 0.9241379310344827, + "test_auc": 0.9737210143947184, + "val_auc": 0.9679383116883117, + "k": 16, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16716, + 15714, + 2224, + 14662, + 23572, + 28519, + 26931, + 6861, + 15191, + 27445, + 11900, + 6225, + 27983, + 12154, + 29558, + 15923 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6c123c09c810fabefc84b43d49d93d6d3940124a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.842758221820554, + "test_acc": 0.8427586206896551, + "test_auc": 0.9258512196276855, + "val_auc": 0.9443993506493507, + "k": 1, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18767 + ] + }, + { + "test_f1": 0.9062021378960444, + "test_acc": 0.9062068965517242, + "test_auc": 0.9634399182856215, + "val_auc": 0.9874188311688312, + "k": 2, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18767, + 15310 + ] + }, + { + "test_f1": 0.9622971768808645, + "test_acc": 0.9622988505747126, + "test_auc": 0.9932321824774067, + "val_auc": 0.9951298701298701, + "k": 5, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18767, + 15310, + 5713, + 17721, + 13305 + ] + }, + { + "test_f1": 0.9779310344827586, + "test_acc": 0.9779310344827586, + "test_auc": 0.997505614481303, + "val_auc": 0.9987824675324676, + "k": 10, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18767, + 15310, + 5713, + 17721, + 13305, + 29759, + 24363, + 25732, + 5528, + 31911 + ] + }, + { + "test_f1": 0.984827522061186, + "test_acc": 0.9848275862068966, + "test_auc": 0.9985515652903296, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18767, + 15310, + 5713, + 17721, + 13305, + 29759, + 24363, + 25732, + 5528, + 31911, + 3529, + 24802, + 2987, + 20185, + 3763, + 6413 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ec19adc912efe54f995608bbea2af3084380862a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6252857337309052, + "test_acc": 0.6259119496855345, + "test_auc": 0.662133217017252, + "val_auc": 0.6862745098039216, + "k": 1, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19965 + ] + }, + { + "test_f1": 0.6401614052443688, + "test_acc": 0.6405031446540881, + "test_auc": 0.6884585064488592, + "val_auc": 0.7675070028011204, + "k": 2, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19965, + 32036 + ] + }, + { + "test_f1": 0.664809010803607, + "test_acc": 0.6649056603773585, + "test_auc": 0.7255661548556562, + "val_auc": 0.7627050820328132, + "k": 5, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19965, + 32036, + 26823, + 27055, + 31651 + ] + }, + { + "test_f1": 0.6706670207411408, + "test_acc": 0.6706918238993711, + "test_auc": 0.7361275099008748, + "val_auc": 0.787515006002401, + "k": 10, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19965, + 32036, + 26823, + 27055, + 31651, + 5703, + 23262, + 10662, + 9712, + 19484 + ] + }, + { + "test_f1": 0.6817434826979464, + "test_acc": 0.6817610062893081, + "test_auc": 0.7478743877456993, + "val_auc": 0.82953181272509, + "k": 16, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19965, + 32036, + 26823, + 27055, + 31651, + 5703, + 23262, + 10662, + 9712, + 19484, + 20442, + 28044, + 27527, + 29572, + 5427, + 16755 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c0006debd96e25c05fbf44cf86d3310169849d0f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6475722825089967, + "test_acc": 0.6485693323550991, + "test_auc": 0.6863182055025171, + "val_auc": 0.7818509615384615, + "k": 1, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19154 + ] + }, + { + "test_f1": 0.6664713839249339, + "test_acc": 0.6669112252384446, + "test_auc": 0.7186559355097084, + "val_auc": 0.7548076923076924, + "k": 2, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19154, + 30564 + ] + }, + { + "test_f1": 0.6896097278903344, + "test_acc": 0.6896551724137931, + "test_auc": 0.7485003509587849, + "val_auc": 0.8125, + "k": 5, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19154, + 30564, + 17957, + 27470, + 1535 + ] + }, + { + "test_f1": 0.6851819455705153, + "test_acc": 0.6852531181217901, + "test_auc": 0.7635657412550976, + "val_auc": 0.8253205128205129, + "k": 10, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19154, + 30564, + 17957, + 27470, + 1535, + 28203, + 8522, + 28783, + 13267, + 13597 + ] + }, + { + "test_f1": 0.6896291102144237, + "test_acc": 0.6896551724137931, + "test_auc": 0.7737155554407225, + "val_auc": 0.8277243589743589, + "k": 16, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 19154, + 30564, + 17957, + 27470, + 1535, + 28203, + 8522, + 28783, + 13267, + 13597, + 29422, + 17568, + 30801, + 13070, + 17720, + 27692 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b9fdbcad195be1f88dad611fbf770a67d175e537 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6629569825875439, + "test_acc": 0.6633962264150943, + "test_auc": 0.7126133752692299, + "val_auc": 0.6692677070828331, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14236 + ] + }, + { + "test_f1": 0.6813281581029257, + "test_acc": 0.6817610062893081, + "test_auc": 0.7548712000234927, + "val_auc": 0.7394957983193278, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14236, + 1803 + ] + }, + { + "test_f1": 0.7406136637737634, + "test_acc": 0.74062893081761, + "test_auc": 0.8162219416144577, + "val_auc": 0.8219287715086034, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14236, + 1803, + 6143, + 24350, + 8072 + ] + }, + { + "test_f1": 0.767038422512211, + "test_acc": 0.7670440251572327, + "test_auc": 0.8397579741154526, + "val_auc": 0.8487394957983194, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14236, + 1803, + 6143, + 24350, + 8072, + 18959, + 17769, + 7122, + 17323, + 14987 + ] + }, + { + "test_f1": 0.7760998353923555, + "test_acc": 0.7761006289308177, + "test_auc": 0.8498514995357145, + "val_auc": 0.8683473389355743, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14236, + 1803, + 6143, + 24350, + 8072, + 18959, + 17769, + 7122, + 17323, + 14987, + 23349, + 25397, + 24119, + 27937, + 9330, + 15990 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8ecc739180d1fbc747815545f04d34ef07718c6b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7370191690165715, + "test_acc": 0.7396226415094339, + "test_auc": 0.7982780933208715, + "val_auc": 0.7839135654261704, + "k": 1, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25105 + ] + }, + { + "test_f1": 0.7511722331680932, + "test_acc": 0.7532075471698113, + "test_auc": 0.823242170688955, + "val_auc": 0.8477390956382553, + "k": 2, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25105, + 15920 + ] + }, + { + "test_f1": 0.7948602067668534, + "test_acc": 0.7952201257861635, + "test_auc": 0.8701243697717254, + "val_auc": 0.8719487795118047, + "k": 5, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25105, + 15920, + 28284, + 12736, + 18389 + ] + }, + { + "test_f1": 0.7952121958299487, + "test_acc": 0.7954716981132075, + "test_auc": 0.8743548356064926, + "val_auc": 0.8695478191276511, + "k": 10, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25105, + 15920, + 28284, + 12736, + 18389, + 14591, + 18341, + 23556, + 20663, + 25026 + ] + }, + { + "test_f1": 0.7990956014147897, + "test_acc": 0.7992452830188679, + "test_auc": 0.8770415649407264, + "val_auc": 0.8763505402160865, + "k": 16, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25105, + 15920, + 28284, + 12736, + 18389, + 14591, + 18341, + 23556, + 20663, + 25026, + 28329, + 5927, + 17689, + 23714, + 22964, + 20225 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..62023637020b88a05516caf4c9401fdfdaad4aca --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6849266384564368, + "test_acc": 0.6850314465408806, + "test_auc": 0.742741805640081, + "val_auc": 0.7200880352140856, + "k": 1, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11800 + ] + }, + { + "test_f1": 0.7197447029045837, + "test_acc": 0.719748427672956, + "test_auc": 0.7782382518563824, + "val_auc": 0.8127250900360143, + "k": 2, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11800, + 17155 + ] + }, + { + "test_f1": 0.77260712396386, + "test_acc": 0.7728301886792452, + "test_auc": 0.8594964350775007, + "val_auc": 0.8443377350940376, + "k": 5, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11800, + 17155, + 32542, + 3220, + 13062 + ] + }, + { + "test_f1": 0.8147402693860071, + "test_acc": 0.8148427672955975, + "test_auc": 0.9005216502841914, + "val_auc": 0.9151660664265706, + "k": 10, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11800, + 17155, + 32542, + 3220, + 13062, + 25563, + 11049, + 7489, + 23934, + 18248 + ] + }, + { + "test_f1": 0.8431961900418122, + "test_acc": 0.8432704402515723, + "test_auc": 0.9179779228972222, + "val_auc": 0.9431772709083633, + "k": 16, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11800, + 17155, + 32542, + 3220, + 13062, + 25563, + 11049, + 7489, + 23934, + 18248, + 11996, + 19331, + 31320, + 1499, + 20270, + 12255 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3856f37c43f7b0126dfe3ed8e8a13870e6fb9f4c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.717051435581095, + "test_acc": 0.7177358490566038, + "test_auc": 0.8056178034487752, + "val_auc": 0.8253301320528212, + "k": 1, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13513 + ] + }, + { + "test_f1": 0.7574317426256083, + "test_acc": 0.7574842767295598, + "test_auc": 0.837349714796074, + "val_auc": 0.9027611044417767, + "k": 2, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13513, + 18155 + ] + }, + { + "test_f1": 0.8095584226218739, + "test_acc": 0.809559748427673, + "test_auc": 0.8978164406671533, + "val_auc": 0.9559823929571828, + "k": 5, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13513, + 18155, + 15141, + 12063, + 9958 + ] + }, + { + "test_f1": 0.831189562567151, + "test_acc": 0.8311949685534591, + "test_auc": 0.9200061972236033, + "val_auc": 0.9635854341736694, + "k": 10, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13513, + 18155, + 15141, + 12063, + 9958, + 11009, + 5529, + 13765, + 4213, + 31163 + ] + }, + { + "test_f1": 0.8550908546080276, + "test_acc": 0.8550943396226415, + "test_auc": 0.9316847233375086, + "val_auc": 0.9663865546218486, + "k": 16, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 13513, + 18155, + 15141, + 12063, + 9958, + 11009, + 5529, + 13765, + 4213, + 31163, + 30131, + 6271, + 28004, + 10619, + 9413, + 3041 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dde6473b961f76e8dc18ef1da6943664dfba42a0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6117998163452709, + "test_acc": 0.6161616161616161, + "test_auc": 0.6853061224489796, + "val_auc": 0.6999198717948718, + "k": 1, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20883 + ] + }, + { + "test_f1": 0.6656426011264721, + "test_acc": 0.6666666666666666, + "test_auc": 0.7287755102040816, + "val_auc": 0.7109375, + "k": 2, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20883, + 1108 + ] + }, + { + "test_f1": 0.7768207091422842, + "test_acc": 0.7777777777777778, + "test_auc": 0.7830612244897959, + "val_auc": 0.7776442307692307, + "k": 5, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20883, + 1108, + 20131, + 10612, + 28834 + ] + }, + { + "test_f1": 0.7669630250275413, + "test_acc": 0.7676767676767676, + "test_auc": 0.826938775510204, + "val_auc": 0.8261217948717949, + "k": 10, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20883, + 1108, + 20131, + 10612, + 28834, + 29025, + 15030, + 16784, + 30591, + 20269 + ] + }, + { + "test_f1": 0.747165058866643, + "test_acc": 0.7474747474747475, + "test_auc": 0.8142857142857143, + "val_auc": 0.8513621794871795, + "k": 16, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20883, + 1108, + 20131, + 10612, + 28834, + 29025, + 15030, + 16784, + 30591, + 20269, + 27044, + 28998, + 14351, + 19149, + 15183, + 23101 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..701049e8c5329ea97fc6c7d1b34c35d00c0b7f3b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-1-tokens-0/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9040380843149427, + "test_acc": 0.9040511727078892, + "test_auc": 0.9585651936715767, + "val_auc": 0.9648, + "k": 1, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17074 + ] + }, + { + "test_f1": 0.9552177726803065, + "test_acc": 0.9552238805970149, + "test_auc": 0.9780869248954356, + "val_auc": 0.9808, + "k": 2, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17074, + 25923 + ] + }, + { + "test_f1": 0.9552234734618076, + "test_acc": 0.9552238805970149, + "test_auc": 0.9812693216948536, + "val_auc": 0.986, + "k": 5, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17074, + 25923, + 28701, + 30786, + 27433 + ] + }, + { + "test_f1": 0.9616169790195993, + "test_acc": 0.9616204690831557, + "test_auc": 0.9891252955082741, + "val_auc": 0.9863999999999999, + "k": 10, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17074, + 25923, + 28701, + 30786, + 27433, + 16789, + 16618, + 6572, + 11863, + 21445 + ] + }, + { + "test_f1": 0.9637477207412005, + "test_acc": 0.9637526652452025, + "test_auc": 0.9891616657574105, + "val_auc": 0.9876, + "k": 16, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17074, + 25923, + 28701, + 30786, + 27433, + 16789, + 16618, + 6572, + 11863, + 21445, + 814, + 1118, + 14648, + 31335, + 20768, + 19405 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..59dfd192ddfb0be9a6611f1d35b845ef741369e4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9804259729647118, + "test_acc": 0.9804295942720763, + "test_auc": 0.9831529743286891, + "val_auc": 0.9858429858429859, + "k": 1, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431 + ] + }, + { + "test_f1": 0.9804259729647118, + "test_acc": 0.9804295942720763, + "test_auc": 0.9947204663269101, + "val_auc": 0.9931359931359931, + "k": 2, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928 + ] + }, + { + "test_f1": 0.9804264636129559, + "test_acc": 0.9804295942720763, + "test_auc": 0.9979134313232281, + "val_auc": 0.9987129987129988, + "k": 5, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1472, + 1758 + ] + }, + { + "test_f1": 0.9890213746557301, + "test_acc": 0.9890214797136038, + "test_auc": 0.9992886801256953, + "val_auc": 0.9961389961389961, + "k": 10, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1472, + 1758, + 126, + 5951, + 1411, + 542, + 786 + ] + }, + { + "test_f1": 0.9952267281351689, + "test_acc": 0.9952267303102625, + "test_auc": 0.9994919143754967, + "val_auc": 0.9978549978549978, + "k": 16, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1472, + 1758, + 126, + 5951, + 1411, + 542, + 786, + 461, + 1319, + 1634, + 1848, + 468, + 1872 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..434b0d2a3f6428e43c603544de063a33ba9c4936 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9287031176088997, + "test_acc": 0.9287103046014258, + "test_auc": 0.9568254672284833, + "val_auc": 0.9479123604795371, + "k": 1, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838 + ] + }, + { + "test_f1": 0.9552784996608675, + "test_acc": 0.9552819183408944, + "test_auc": 0.9658474627527671, + "val_auc": 0.9586606035551881, + "k": 2, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441 + ] + }, + { + "test_f1": 0.9481438008896271, + "test_acc": 0.948152948801037, + "test_auc": 0.9688648750361216, + "val_auc": 0.9689954526663911, + "k": 5, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 1500, + 943, + 4110 + ] + }, + { + "test_f1": 0.95331941945702, + "test_acc": 0.9533376539209333, + "test_auc": 0.969657869801012, + "val_auc": 0.9710624224886317, + "k": 10, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 1500, + 943, + 4110, + 1697, + 1180, + 654, + 1417, + 908 + ] + }, + { + "test_f1": 0.9656419483228559, + "test_acc": 0.9656513285806869, + "test_auc": 0.9866534948892159, + "val_auc": 0.9904919388176933, + "k": 16, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 1500, + 943, + 4110, + 1697, + 1180, + 654, + 1417, + 908, + 601, + 698, + 1411, + 920, + 786, + 1619 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f7bb70e4fe7ec7915578e2b380fd077a1fdc3840 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5590781667090372, + "test_acc": 0.5801257861635221, + "test_auc": 0.585981034673061, + "val_auc": 0.6024409763905562, + "k": 1, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921 + ] + }, + { + "test_f1": 0.565400485274269, + "test_acc": 0.5828930817610063, + "test_auc": 0.6041191790906486, + "val_auc": 0.5950380152060826, + "k": 2, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777 + ] + }, + { + "test_f1": 0.580661212514527, + "test_acc": 0.5914465408805032, + "test_auc": 0.619147699483261, + "val_auc": 0.6034413765506204, + "k": 5, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 1438, + 7284, + 2046 + ] + }, + { + "test_f1": 0.5800035259187938, + "test_acc": 0.5881761006289308, + "test_auc": 0.6219142990808464, + "val_auc": 0.633453381352541, + "k": 10, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 1438, + 7284, + 2046, + 505, + 969, + 1380, + 1335, + 1178 + ] + }, + { + "test_f1": 0.5893337000370263, + "test_acc": 0.5959748427672956, + "test_auc": 0.6328109573394063, + "val_auc": 0.6642657062825129, + "k": 16, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 1438, + 7284, + 2046, + 505, + 969, + 1380, + 1335, + 1178, + 31036, + 1533, + 918, + 521, + 1310, + 988 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..01e7b6ee82127849e31dba0fbb8a3980d4eba4f2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6872541710522294, + "test_acc": 0.6918238993710691, + "test_auc": 0.7058552877405347, + "val_auc": 0.7739095638255302, + "k": 1, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567 + ] + }, + { + "test_f1": 0.7099007950634865, + "test_acc": 0.7111949685534591, + "test_auc": 0.743777460940783, + "val_auc": 0.7595038015206081, + "k": 2, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470 + ] + }, + { + "test_f1": 0.7211270862999681, + "test_acc": 0.7225157232704402, + "test_auc": 0.7706315902460561, + "val_auc": 0.7438975590236093, + "k": 5, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 1164, + 1345, + 1629 + ] + }, + { + "test_f1": 0.7302304748204251, + "test_acc": 0.7318238993710692, + "test_auc": 0.7863110975870321, + "val_auc": 0.7971188475390156, + "k": 10, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 1164, + 1345, + 1629, + 6411, + 1731, + 628, + 4196, + 4505 + ] + }, + { + "test_f1": 0.7296655020484264, + "test_acc": 0.730817610062893, + "test_auc": 0.7843959833485057, + "val_auc": 0.8023209283713484, + "k": 16, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 1164, + 1345, + 1629, + 6411, + 1731, + 628, + 4196, + 4505, + 433, + 1335, + 794, + 993, + 534, + 1551 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c0911d0ea73404d3bada3041f9557493d6bd1c6d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7475733521873025, + "test_acc": 0.7489308176100629, + "test_auc": 0.8158130970017387, + "val_auc": 0.8541416566626652, + "k": 1, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411 + ] + }, + { + "test_f1": 0.7328480365677116, + "test_acc": 0.7350943396226415, + "test_auc": 0.8219170331500831, + "val_auc": 0.867547018807523, + "k": 2, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786 + ] + }, + { + "test_f1": 0.762683186217541, + "test_acc": 0.7630188679245283, + "test_auc": 0.8420313020549063, + "val_auc": 0.8655462184873949, + "k": 5, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1758 + ] + }, + { + "test_f1": 0.7904590019238134, + "test_acc": 0.7906918238993711, + "test_auc": 0.8870725611849254, + "val_auc": 0.9289715886354541, + "k": 10, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1758, + 1417, + 1335, + 1360, + 181, + 5054 + ] + }, + { + "test_f1": 0.804519616976982, + "test_acc": 0.8045283018867925, + "test_auc": 0.8998707899131073, + "val_auc": 0.9269707883153262, + "k": 16, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1758, + 1417, + 1335, + 1360, + 181, + 5054, + 126, + 1183, + 19471, + 191, + 1108, + 1703 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a1e2a7b01394410d9be0a50285a1ad19ad622f19 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8436226360493567, + "test_acc": 0.8437735849056603, + "test_auc": 0.8927839811896037, + "val_auc": 0.9487795118047219, + "k": 1, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8744333666094517, + "test_acc": 0.8744654088050314, + "test_auc": 0.9372707052582229, + "val_auc": 0.9635854341736694, + "k": 2, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8799869205761098, + "test_acc": 0.88, + "test_auc": 0.9406512046612843, + "val_auc": 0.9627851140456183, + "k": 5, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 4675, + 100, + 505 + ] + }, + { + "test_f1": 0.8840223680040541, + "test_acc": 0.8840251572327044, + "test_auc": 0.9439974016216067, + "val_auc": 0.9615846338535414, + "k": 10, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 4675, + 100, + 505, + 821, + 22141, + 6745, + 6319, + 7284 + ] + }, + { + "test_f1": 0.8860373319465121, + "test_acc": 0.8860377358490567, + "test_auc": 0.9459277051336705, + "val_auc": 0.9611844737895159, + "k": 16, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 4675, + 100, + 505, + 821, + 22141, + 6745, + 6319, + 7284, + 1697, + 2632, + 7143, + 1438, + 1108, + 3654 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f6c2dc7a860a1bc96685eb78b05b1607abab1905 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5242349138883963, + "test_acc": 0.558974358974359, + "test_auc": 0.5618414851718451, + "val_auc": 0.5552835577315381, + "k": 1, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945 + ] + }, + { + "test_f1": 0.5455884447124025, + "test_acc": 0.5456410256410257, + "test_auc": 0.5728616151075504, + "val_auc": 0.6048551611587107, + "k": 2, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 238 + ] + }, + { + "test_f1": 0.5579920427728778, + "test_acc": 0.5661538461538461, + "test_auc": 0.5847948295014643, + "val_auc": 0.5993472052223582, + "k": 5, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 238, + 1264, + 470, + 6892 + ] + }, + { + "test_f1": 0.5640933931934753, + "test_acc": 0.5641025641025641, + "test_auc": 0.6124229979466119, + "val_auc": 0.6234190126478989, + "k": 10, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 238, + 1264, + 470, + 6892, + 1523, + 1851, + 4531, + 7073, + 1588 + ] + }, + { + "test_f1": 0.5741841959533149, + "test_acc": 0.5764102564102564, + "test_auc": 0.6212277409364796, + "val_auc": 0.6507547939616483, + "k": 16, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 238, + 1264, + 470, + 6892, + 1523, + 1851, + 4531, + 7073, + 1588, + 551, + 913, + 26171, + 1895, + 1551, + 1548 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cadee36d360233b733f612b8ce45862339a3ffa3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3339032896927634, + "test_acc": 0.5005128205128205, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239 + ] + }, + { + "test_f1": 0.48325684281280096, + "test_acc": 0.5251282051282051, + "test_auc": 0.5120762783182414, + "val_auc": 0.5071399428804569, + "k": 2, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 6341 + ] + }, + { + "test_f1": 0.5261528492378124, + "test_acc": 0.5261538461538462, + "test_auc": 0.5320084491870603, + "val_auc": 0.5063239494084046, + "k": 5, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 6341, + 1409, + 1264, + 238 + ] + }, + { + "test_f1": 0.5012696477960669, + "test_acc": 0.5035897435897436, + "test_auc": 0.515192126434847, + "val_auc": 0.5299877600979193, + "k": 10, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 6341, + 1409, + 1264, + 238, + 456, + 326, + 400, + 827, + 538 + ] + }, + { + "test_f1": 0.5099650369914062, + "test_acc": 0.5107692307692308, + "test_auc": 0.5151226983539233, + "val_auc": 0.5450836393308853, + "k": 16, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 6341, + 1409, + 1264, + 238, + 456, + 326, + 400, + 827, + 538, + 1478, + 1158, + 16809, + 4110, + 6821, + 4201 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8d21b5246f8828698c8069455b17c651731e85e8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5327340553925142, + "test_acc": 0.5343589743589744, + "test_auc": 0.5359027333625072, + "val_auc": 0.5711954304365565, + "k": 1, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1156 + ] + }, + { + "test_f1": 0.5468314147982929, + "test_acc": 0.5476923076923077, + "test_auc": 0.5587950651361633, + "val_auc": 0.6325989392084863, + "k": 2, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1156, + 1815 + ] + }, + { + "test_f1": 0.5398049654049655, + "test_acc": 0.5425641025641026, + "test_auc": 0.5527779647894435, + "val_auc": 0.6332109343125254, + "k": 5, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1156, + 1815, + 12594, + 698, + 1770 + ] + }, + { + "test_f1": 0.5447358280093297, + "test_acc": 0.5466666666666666, + "test_auc": 0.5554688288955465, + "val_auc": 0.6511627906976745, + "k": 10, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1156, + 1815, + 12594, + 698, + 1770, + 191, + 1409, + 16809, + 1684, + 1611 + ] + }, + { + "test_f1": 0.5373513685964538, + "test_acc": 0.5425641025641026, + "test_auc": 0.5473604705961558, + "val_auc": 0.6336189310485516, + "k": 16, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1156, + 1815, + 12594, + 698, + 1770, + 191, + 1409, + 16809, + 1684, + 1611, + 360, + 538, + 400, + 1265, + 7616, + 132 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a61c5bc5943b47ead7ab6f5dc647aeb18bb2b755 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8279318478603097, + "test_acc": 0.8282828282828283, + "test_auc": 0.873061224489796, + "val_auc": 0.8537660256410255, + "k": 1, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997 + ] + }, + { + "test_f1": 0.8378878804410719, + "test_acc": 0.8383838383838383, + "test_auc": 0.8828571428571429, + "val_auc": 0.8850160256410257, + "k": 2, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335 + ] + }, + { + "test_f1": 0.8378878804410719, + "test_acc": 0.8383838383838383, + "test_auc": 0.880204081632653, + "val_auc": 0.8914262820512819, + "k": 5, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 314, + 1156 + ] + }, + { + "test_f1": 0.7968197147852603, + "test_acc": 0.797979797979798, + "test_auc": 0.8542857142857143, + "val_auc": 0.9415064102564102, + "k": 10, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 314, + 1156, + 851, + 3526, + 2412, + 1499, + 1279 + ] + }, + { + "test_f1": 0.8176238654962059, + "test_acc": 0.8181818181818182, + "test_auc": 0.8855102040816326, + "val_auc": 0.9407051282051282, + "k": 16, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 314, + 1156, + 851, + 3526, + 2412, + 1499, + 1279, + 1473, + 4201, + 1602, + 235, + 12946, + 771 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9150fcbc39da64093eaa04ab4dd73c07fa03d1a5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8684184718931782, + "test_acc": 0.8686868686868687, + "test_auc": 0.8979591836734693, + "val_auc": 0.9352964743589745, + "k": 1, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.858151895385938, + "test_acc": 0.8585858585858586, + "test_auc": 0.9032653061224488, + "val_auc": 0.953125, + "k": 2, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335 + ] + }, + { + "test_f1": 0.8786393345216874, + "test_acc": 0.8787878787878788, + "test_auc": 0.9481632653061225, + "val_auc": 0.9733573717948718, + "k": 5, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4201, + 1116 + ] + }, + { + "test_f1": 0.8886617839096123, + "test_acc": 0.8888888888888888, + "test_auc": 0.953061224489796, + "val_auc": 0.9859775641025641, + "k": 10, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4201, + 1116, + 4889, + 3526, + 1924, + 2832, + 581 + ] + }, + { + "test_f1": 0.8886617839096123, + "test_acc": 0.8888888888888888, + "test_auc": 0.9514285714285715, + "val_auc": 0.9899839743589743, + "k": 16, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4201, + 1116, + 4889, + 3526, + 1924, + 2832, + 581, + 12946, + 4531, + 2000, + 7481, + 2900, + 1851 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8af8fa36c06898b5e00d4599c7cefdc9398b5f97 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8880672805906451, + "test_acc": 0.8888888888888888, + "test_auc": 0.8971428571428571, + "val_auc": 0.8321314102564104, + "k": 1, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388 + ] + }, + { + "test_f1": 0.8980546202768425, + "test_acc": 0.898989898989899, + "test_auc": 0.9320408163265306, + "val_auc": 0.891025641025641, + "k": 2, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335 + ] + }, + { + "test_f1": 0.8382848088730442, + "test_acc": 0.8383838383838383, + "test_auc": 0.9438775510204082, + "val_auc": 0.9509214743589743, + "k": 5, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 1596, + 7384 + ] + }, + { + "test_f1": 0.9090352334202494, + "test_acc": 0.9090909090909091, + "test_auc": 0.9722448979591838, + "val_auc": 0.9567307692307693, + "k": 10, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 1596, + 7384, + 4201, + 1499, + 2900, + 2832, + 3526 + ] + }, + { + "test_f1": 0.8989280055456527, + "test_acc": 0.898989898989899, + "test_auc": 0.9795918367346939, + "val_auc": 0.9575320512820513, + "k": 16, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 1596, + 7384, + 4201, + 1499, + 2900, + 2832, + 3526, + 1769, + 3825, + 1847, + 5888, + 5463, + 2000 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f3f290e75a232c66211c0d4594044d5fad6fb408 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6405808026412557, + "test_acc": 0.6623320329432163, + "test_auc": 0.6752379821458903, + "val_auc": 0.7558, + "k": 1, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.6528718985791029, + "test_acc": 0.6775032509752926, + "test_auc": 0.7722691614520781, + "val_auc": 0.8508, + "k": 2, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388 + ] + }, + { + "test_f1": 0.7480441928265256, + "test_acc": 0.7485912440398786, + "test_auc": 0.836553275984133, + "val_auc": 0.9338, + "k": 5, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 5352, + 789, + 2335 + ] + }, + { + "test_f1": 0.7671581855827996, + "test_acc": 0.7680970957954053, + "test_auc": 0.8605033061217742, + "val_auc": 0.9476, + "k": 10, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 5352, + 789, + 2335, + 7384, + 2000, + 337, + 4201, + 12946 + ] + }, + { + "test_f1": 0.792768296521491, + "test_acc": 0.7928045080190724, + "test_auc": 0.8873528629255908, + "val_auc": 0.9500000000000001, + "k": 16, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 5352, + 789, + 2335, + 7384, + 2000, + 337, + 4201, + 12946, + 7954, + 239, + 7481, + 2659, + 2330, + 1614 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9e598b105ad43915df9d040c462e58dbed91cbf6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6383025806014451, + "test_acc": 0.6597312527091461, + "test_auc": 0.6744029966284923, + "val_auc": 0.7162, + "k": 1, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.7089106271077424, + "test_acc": 0.7121803207628955, + "test_auc": 0.7558261095687386, + "val_auc": 0.7807999999999999, + "k": 2, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384 + ] + }, + { + "test_f1": 0.7742327983102606, + "test_acc": 0.7754659731252709, + "test_auc": 0.8684991755363523, + "val_auc": 0.9136, + "k": 5, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 4201, + 3526, + 5388 + ] + }, + { + "test_f1": 0.8123465062306635, + "test_acc": 0.8131772865192891, + "test_auc": 0.9093811487176097, + "val_auc": 0.9219999999999999, + "k": 10, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 4201, + 3526, + 5388, + 7481, + 1924, + 1614, + 2900, + 1091 + ] + }, + { + "test_f1": 0.8260140726048085, + "test_acc": 0.8261811876896402, + "test_auc": 0.9213745770584159, + "val_auc": 0.9308000000000001, + "k": 16, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 4201, + 3526, + 5388, + 7481, + 1924, + 1614, + 2900, + 1091, + 3842, + 6244, + 2832, + 2050, + 119, + 2659 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bc173236044acec8d9b57d8010cf02925b876e46 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8208681713437324, + "test_acc": 0.8218465539661899, + "test_auc": 0.8524845140624789, + "val_auc": 0.8703999999999998, + "k": 1, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.823957972388373, + "test_acc": 0.8253142609449502, + "test_auc": 0.9010482036913726, + "val_auc": 0.9076, + "k": 2, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789 + ] + }, + { + "test_f1": 0.8255163710892757, + "test_acc": 0.8279150411790204, + "test_auc": 0.9032904892819725, + "val_auc": 0.9259999999999999, + "k": 5, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789, + 12946, + 4531, + 4889 + ] + }, + { + "test_f1": 0.8292051780270234, + "test_acc": 0.8313827481577807, + "test_auc": 0.9237284696241137, + "val_auc": 0.9527999999999999, + "k": 10, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789, + 12946, + 4531, + 4889, + 14154, + 2335, + 1851, + 1924, + 737 + ] + }, + { + "test_f1": 0.8233731497940172, + "test_acc": 0.8257477243172952, + "test_auc": 0.9190650266579086, + "val_auc": 0.9528000000000001, + "k": 16, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789, + 12946, + 4531, + 4889, + 14154, + 2335, + 1851, + 1924, + 737, + 5352, + 3526, + 4201, + 470, + 5388, + 1091 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cd9225f70151af4c0882525467073f061cbf1f9b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9199164336017792, + "test_acc": 0.92, + "test_auc": 0.9647031844346451, + "val_auc": 0.9791921664626683, + "k": 1, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958 + ] + }, + { + "test_f1": 0.9219478320990856, + "test_acc": 0.9220512820512821, + "test_auc": 0.9647031844346451, + "val_auc": 0.9791921664626683, + "k": 2, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818 + ] + }, + { + "test_f1": 0.9620510425317389, + "test_acc": 0.9620512820512821, + "test_auc": 0.9887589625340829, + "val_auc": 0.9946960424316605, + "k": 5, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 4586 + ] + }, + { + "test_f1": 0.9661531340597287, + "test_acc": 0.9661538461538461, + "test_auc": 0.9918853300568891, + "val_auc": 0.9969400244798041, + "k": 10, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 4586, + 2780, + 1158, + 386, + 1279, + 1499 + ] + }, + { + "test_f1": 0.9661531340597287, + "test_acc": 0.9661538461538461, + "test_auc": 0.992261924798869, + "val_auc": 0.9969400244798041, + "k": 16, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 4586, + 2780, + 1158, + 386, + 1279, + 1499, + 5463, + 1596, + 3475, + 851, + 730, + 779 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5645e3e5e757328ed6e02047f7d9effe598bea3e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9237370894583812, + "test_acc": 0.9241025641025641, + "test_auc": 0.9264903894704952, + "val_auc": 0.9534883720930232, + "k": 1, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065 + ] + }, + { + "test_f1": 0.925776755025442, + "test_acc": 0.9261538461538461, + "test_auc": 0.9736846534486822, + "val_auc": 0.9861281109751121, + "k": 2, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508 + ] + }, + { + "test_f1": 0.9496567421760119, + "test_acc": 0.9497435897435897, + "test_auc": 0.9875239842461373, + "val_auc": 0.9967360261117911, + "k": 5, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 5406 + ] + }, + { + "test_f1": 0.9764088666776238, + "test_acc": 0.9764102564102564, + "test_auc": 0.9969178139832363, + "val_auc": 1.0, + "k": 10, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 5406, + 1279, + 581, + 2708, + 3958, + 992 + ] + }, + { + "test_f1": 0.9743586506798301, + "test_acc": 0.9743589743589743, + "test_auc": 0.9971850069007304, + "val_auc": 1.0, + "k": 16, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 5406, + 1279, + 581, + 2708, + 3958, + 992, + 17569, + 1499, + 3475, + 778, + 1482, + 15024 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a1d60888f4ac351a15061beee7b83309c3847f1b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9620404209528223, + "test_acc": 0.9620512820512821, + "test_auc": 0.976129363449692, + "val_auc": 0.9987760097919217, + "k": 1, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037 + ] + }, + { + "test_f1": 0.9630675990675991, + "test_acc": 0.963076923076923, + "test_auc": 0.9770866294139429, + "val_auc": 0.9987760097919217, + "k": 2, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508 + ] + }, + { + "test_f1": 0.9712747992747993, + "test_acc": 0.9712820512820513, + "test_auc": 0.9874356212340527, + "val_auc": 0.9995920032639739, + "k": 5, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 1253, + 5472 + ] + }, + { + "test_f1": 0.9784590459118746, + "test_acc": 0.9784615384615385, + "test_auc": 0.9964023630794089, + "val_auc": 1.0, + "k": 10, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 1253, + 5472, + 493, + 1039, + 3537, + 3958, + 26034 + ] + }, + { + "test_f1": 0.9794859710430901, + "test_acc": 0.9794871794871794, + "test_auc": 0.9970756050762446, + "val_auc": 1.0, + "k": 16, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 1253, + 5472, + 493, + 1039, + 3537, + 3958, + 26034, + 778, + 1279, + 582, + 581, + 3475, + 12813 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1b978d6bb01098a50eb9de18a43185ebc00af306 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5980428247780579, + "test_acc": 0.623753792804508, + "test_auc": 0.6298045487545864, + "val_auc": 0.6582, + "k": 1, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351 + ] + }, + { + "test_f1": 0.6490498176734419, + "test_acc": 0.6536627654963155, + "test_auc": 0.6869337918864359, + "val_auc": 0.7512, + "k": 2, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 559 + ] + }, + { + "test_f1": 0.6779067475658432, + "test_acc": 0.6805374945817079, + "test_auc": 0.7382474473192531, + "val_auc": 0.7959999999999999, + "k": 5, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 559, + 1894, + 1441, + 3078 + ] + }, + { + "test_f1": 0.6967382114424031, + "test_acc": 0.6996098829648895, + "test_auc": 0.7671622968339694, + "val_auc": 0.7992, + "k": 10, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 559, + 1894, + 1441, + 3078, + 619, + 19018, + 654, + 7595, + 1960 + ] + }, + { + "test_f1": 0.7119529767462748, + "test_acc": 0.7126137841352406, + "test_auc": 0.7887689562756188, + "val_auc": 0.8256000000000001, + "k": 16, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 559, + 1894, + 1441, + 3078, + 619, + 19018, + 654, + 7595, + 1960, + 1602, + 52, + 3274, + 4510, + 121, + 1305 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1a4ecc9e9baac487fe25190ac4270ce96872451e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6108183897450026, + "test_acc": 0.6376246207195492, + "test_auc": 0.6437821010971303, + "val_auc": 0.6344, + "k": 1, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274 + ] + }, + { + "test_f1": 0.6123149585604465, + "test_acc": 0.6384915474642393, + "test_auc": 0.6533385892577723, + "val_auc": 0.6526, + "k": 2, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121 + ] + }, + { + "test_f1": 0.6806277045630823, + "test_acc": 0.682271348071088, + "test_auc": 0.7544860743054439, + "val_auc": 0.7408, + "k": 5, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 1602, + 19018 + ] + }, + { + "test_f1": 0.7458537999290492, + "test_acc": 0.7459904638058084, + "test_auc": 0.8238458636275499, + "val_auc": 0.8024, + "k": 10, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 1602, + 19018, + 1524, + 698, + 52, + 3078, + 4510 + ] + }, + { + "test_f1": 0.7511919307766238, + "test_acc": 0.7511920242739488, + "test_auc": 0.831158187292287, + "val_auc": 0.8428, + "k": 16, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 1602, + 19018, + 1524, + 698, + 52, + 3078, + 4510, + 1895, + 32479, + 1158, + 649, + 1499, + 1960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a61fa9a5bd8fddb582a723e4f80ed9de3e09b554 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.570345877010146, + "test_acc": 0.5934113567403554, + "test_auc": 0.5989424017821041, + "val_auc": 0.5464, + "k": 1, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.570880531268104, + "test_acc": 0.600346770697876, + "test_auc": 0.6626151205280175, + "val_auc": 0.6798000000000001, + "k": 2, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508 + ] + }, + { + "test_f1": 0.6546889597208375, + "test_acc": 0.6575639358474209, + "test_auc": 0.713260261453431, + "val_auc": 0.7168, + "k": 5, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121 + ] + }, + { + "test_f1": 0.6665119990617179, + "test_acc": 0.6688339835283919, + "test_auc": 0.7387761712719888, + "val_auc": 0.7924, + "k": 10, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 567, + 52, + 925, + 3301, + 51 + ] + }, + { + "test_f1": 0.6700447579177848, + "test_acc": 0.6718682271348071, + "test_auc": 0.7485442241699373, + "val_auc": 0.7736, + "k": 16, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 567, + 52, + 925, + 3301, + 51, + 1441, + 628, + 3274, + 4654, + 4510, + 7893 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0efd109897eb2af3eca9815bcf431ced5bf103c3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6173765750380024, + "test_acc": 0.6183368869936035, + "test_auc": 0.6303509729041644, + "val_auc": 0.6432, + "k": 1, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7129 + ] + }, + { + "test_f1": 0.6268453137707765, + "test_acc": 0.6268656716417911, + "test_auc": 0.6707128568830696, + "val_auc": 0.7083999999999999, + "k": 2, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7129, + 1048 + ] + }, + { + "test_f1": 0.69467115409936, + "test_acc": 0.6950959488272921, + "test_auc": 0.7445899254409893, + "val_auc": 0.7448, + "k": 5, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7129, + 1048, + 1504, + 696, + 246 + ] + }, + { + "test_f1": 0.7267409576953303, + "test_acc": 0.7270788912579957, + "test_auc": 0.8027823240589198, + "val_auc": 0.7827999999999999, + "k": 10, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7129, + 1048, + 1504, + 696, + 246, + 927, + 1818, + 4672, + 1927, + 741 + ] + }, + { + "test_f1": 0.7288338221721588, + "test_acc": 0.7292110874200426, + "test_auc": 0.7925440989270776, + "val_auc": 0.7908, + "k": 16, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7129, + 1048, + 1504, + 696, + 246, + 927, + 1818, + 4672, + 1927, + 741, + 1902, + 6905, + 437, + 29085, + 1499, + 499 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6b1813180d009db35bb8831e7cdb8c7d6099542b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9393815708101421, + "test_acc": 0.9393939393939394, + "test_auc": 0.9710204081632654, + "val_auc": 0.9524959742351048, + "k": 1, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299 + ] + }, + { + "test_f1": 0.9494949494949495, + "test_acc": 0.9494949494949495, + "test_auc": 0.9767346938775511, + "val_auc": 0.9569243156199679, + "k": 2, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 23633 + ] + }, + { + "test_f1": 0.9797979797979798, + "test_acc": 0.9797979797979798, + "test_auc": 0.9987755102040816, + "val_auc": 0.9923510466988729, + "k": 5, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 23633, + 114, + 1915, + 6055 + ] + }, + { + "test_f1": 1.0, + "test_acc": 1.0, + "test_auc": 1.0000000000000002, + "val_auc": 1.0, + "k": 10, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 23633, + 114, + 1915, + 6055, + 432, + 19890, + 6877, + 1500, + 1232 + ] + }, + { + "test_f1": 1.0, + "test_acc": 1.0, + "test_auc": 1.0, + "val_auc": 1.0, + "k": 16, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 23633, + 114, + 1915, + 6055, + 432, + 19890, + 6877, + 1500, + 1232, + 25757, + 1574, + 27804, + 448, + 373, + 1960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9065530751e54e84bb33dad18b9a4c10137b55ec --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7929166350218981, + "test_acc": 0.797979797979798, + "test_auc": 0.8, + "val_auc": 0.8055555555555556, + "k": 1, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429 + ] + }, + { + "test_f1": 0.7877488779805815, + "test_acc": 0.7878787878787878, + "test_auc": 0.8583673469387756, + "val_auc": 0.8651368760064413, + "k": 2, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429, + 23791 + ] + }, + { + "test_f1": 0.8888888888888888, + "test_acc": 0.8888888888888888, + "test_auc": 0.9710204081632653, + "val_auc": 0.9601449275362319, + "k": 5, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429, + 23791, + 278, + 1657, + 1411 + ] + }, + { + "test_f1": 0.9190265998776638, + "test_acc": 0.9191919191919192, + "test_auc": 0.986530612244898, + "val_auc": 0.9790660225442834, + "k": 10, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429, + 23791, + 278, + 1657, + 1411, + 6088, + 1396, + 1067, + 114, + 521 + ] + }, + { + "test_f1": 0.9595712022182611, + "test_acc": 0.9595959595959596, + "test_auc": 0.9881632653061225, + "val_auc": 0.9738325281803542, + "k": 16, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429, + 23791, + 278, + 1657, + 1411, + 6088, + 1396, + 1067, + 114, + 521, + 16, + 1299, + 4314, + 5387, + 1420, + 5001 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a93e5856db12a4be5a344d7856ce947473ea64a4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.898989898989899, + "test_acc": 0.898989898989899, + "test_auc": 0.9506122448979591, + "val_auc": 0.9223027375201288, + "k": 1, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9494330117733285, + "test_acc": 0.9494949494949495, + "test_auc": 0.9934693877551021, + "val_auc": 0.9734299516908212, + "k": 2, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610 + ] + }, + { + "test_f1": 0.9696784111400831, + "test_acc": 0.9696969696969697, + "test_auc": 0.9975510204081632, + "val_auc": 0.9826892109500805, + "k": 5, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610, + 177, + 1501, + 6662 + ] + }, + { + "test_f1": 0.9696784111400831, + "test_acc": 0.9696969696969697, + "test_auc": 0.9985714285714286, + "val_auc": 0.9893317230273753, + "k": 10, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610, + 177, + 1501, + 6662, + 1550, + 419, + 907, + 7577, + 1915 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 0.996734693877551, + "val_auc": 1.0, + "k": 16, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610, + 177, + 1501, + 6662, + 1550, + 419, + 907, + 7577, + 1915, + 4314, + 635, + 1417, + 1360, + 1960, + 3395 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c32270dc4f53dfd8a943bf204f8391a8ea24a511 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9288804594067752, + "test_acc": 0.9288888888888889, + "test_auc": 0.9461125158027813, + "val_auc": 0.9220779220779222, + "k": 1, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752 + ] + }, + { + "test_f1": 0.9244414592977442, + "test_acc": 0.9244444444444444, + "test_auc": 0.9562262958280657, + "val_auc": 0.9358766233766234, + "k": 2, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309 + ] + }, + { + "test_f1": 0.9466645596291614, + "test_acc": 0.9466666666666667, + "test_auc": 0.9607695954487989, + "val_auc": 0.9833603896103896, + "k": 5, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580 + ] + }, + { + "test_f1": 0.9511111111111111, + "test_acc": 0.9511111111111111, + "test_auc": 0.9636140960809102, + "val_auc": 0.9862012987012988, + "k": 10, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 10260, + 1295, + 5563, + 3581, + 27804 + ] + }, + { + "test_f1": 0.9555555555555556, + "test_acc": 0.9555555555555556, + "test_auc": 0.9776390644753478, + "val_auc": 0.9857954545454546, + "k": 16, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 10260, + 1295, + 5563, + 3581, + 27804, + 1541, + 1182, + 969, + 126, + 1327, + 943 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cad115f85e698c93c08f343bb2c4ea085cce25ea --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8172146616218644, + "test_acc": 0.8177777777777778, + "test_auc": 0.8311472819216182, + "val_auc": 0.8482142857142858, + "k": 1, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406 + ] + }, + { + "test_f1": 0.8121985590167409, + "test_acc": 0.8133333333333334, + "test_auc": 0.8623182680151706, + "val_auc": 0.8567370129870129, + "k": 2, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854 + ] + }, + { + "test_f1": 0.8385270205643233, + "test_acc": 0.84, + "test_auc": 0.8949115044247787, + "val_auc": 0.8705357142857143, + "k": 5, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1182, + 1804, + 1915 + ] + }, + { + "test_f1": 0.8617964514701901, + "test_acc": 0.8622222222222222, + "test_auc": 0.9179045512010114, + "val_auc": 0.8835227272727273, + "k": 10, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1182, + 1804, + 1915, + 1232, + 1570, + 311, + 6662, + 15467 + ] + }, + { + "test_f1": 0.8619222607070207, + "test_acc": 0.8622222222222222, + "test_auc": 0.9266750948166877, + "val_auc": 0.8835227272727273, + "k": 16, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1182, + 1804, + 1915, + 1232, + 1570, + 311, + 6662, + 15467, + 25940, + 322, + 3481, + 19890, + 1420, + 800 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3e99cc983d15fd06e7627295fb5f5f6df2bc1f43 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9155221842403964, + "test_acc": 0.9155555555555556, + "test_auc": 0.9554756637168141, + "val_auc": 0.9415584415584415, + "k": 1, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9243368516357579, + "test_acc": 0.9244444444444444, + "test_auc": 0.9668141592920355, + "val_auc": 0.9577922077922078, + "k": 2, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16 + ] + }, + { + "test_f1": 0.9466350461133068, + "test_acc": 0.9466666666666667, + "test_auc": 0.9845132743362831, + "val_auc": 0.9699675324675324, + "k": 5, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 21467, + 114 + ] + }, + { + "test_f1": 0.9555292050944225, + "test_acc": 0.9555555555555556, + "test_auc": 0.9927307206068268, + "val_auc": 0.9951298701298701, + "k": 10, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 21467, + 114, + 1835, + 1915, + 1960, + 969, + 4678 + ] + }, + { + "test_f1": 0.9644360149623308, + "test_acc": 0.9644444444444444, + "test_auc": 0.9939159292035398, + "val_auc": 0.9967532467532467, + "k": 16, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 21467, + 114, + 1835, + 1915, + 1960, + 969, + 4678, + 1417, + 15752, + 1420, + 79, + 1232, + 7095 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b4a567e6a46d7c884ae6e52887a3cc11e9de89cc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6594289164198677, + "test_acc": 0.6644993498049415, + "test_auc": 0.6834645811318826, + "val_auc": 0.7234, + "k": 1, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649 + ] + }, + { + "test_f1": 0.7048566405828927, + "test_acc": 0.7104464672735153, + "test_auc": 0.7605865040486651, + "val_auc": 0.8028000000000001, + "k": 2, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394 + ] + }, + { + "test_f1": 0.7232391455067397, + "test_acc": 0.7247507585609017, + "test_auc": 0.7817636457376658, + "val_auc": 0.8502, + "k": 5, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 943, + 7246, + 1928 + ] + }, + { + "test_f1": 0.7321557893079398, + "test_acc": 0.7334200260078023, + "test_auc": 0.7906771725030477, + "val_auc": 0.8855999999999999, + "k": 10, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 943, + 7246, + 1928, + 1560, + 7614, + 28053, + 9605, + 32305 + ] + }, + { + "test_f1": 0.7328927724439724, + "test_acc": 0.7334200260078023, + "test_auc": 0.8071100782977418, + "val_auc": 0.8892, + "k": 16, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 943, + 7246, + 1928, + 1560, + 7614, + 28053, + 9605, + 32305, + 17225, + 584, + 2041, + 1551, + 253, + 1848 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fb8071bd969a0dfffd4640b2bd56994326c03bc1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6016470111213743, + "test_acc": 0.6016471608149111, + "test_auc": 0.630380996902061, + "val_auc": 0.6295999999999999, + "k": 1, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943 + ] + }, + { + "test_f1": 0.6286981391720032, + "test_acc": 0.6293888166449935, + "test_auc": 0.6633463904725974, + "val_auc": 0.6974, + "k": 2, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253 + ] + }, + { + "test_f1": 0.6324927904732357, + "test_acc": 0.6337234503684439, + "test_auc": 0.69978625573254, + "val_auc": 0.7303999999999999, + "k": 5, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 32305, + 1011, + 7649 + ] + }, + { + "test_f1": 0.6553074710044289, + "test_acc": 0.6562635457303858, + "test_auc": 0.7227348293427889, + "val_auc": 0.7804, + "k": 10, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 32305, + 1011, + 7649, + 5394, + 1551, + 28053, + 7614, + 844 + ] + }, + { + "test_f1": 0.6747703550318936, + "test_acc": 0.6749024707412223, + "test_auc": 0.7469640648087049, + "val_auc": 0.79, + "k": 16, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 32305, + 1011, + 7649, + 5394, + 1551, + 28053, + 7614, + 844, + 1928, + 1411, + 319, + 1875, + 2041, + 1417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9d6226dc8b7c152a6edcd075c418d8a6287a7da6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6062340696233182, + "test_acc": 0.6207195491980928, + "test_auc": 0.6405842042685721, + "val_auc": 0.686, + "k": 1, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253 + ] + }, + { + "test_f1": 0.6126210861510584, + "test_acc": 0.6276549631556133, + "test_auc": 0.7001007844805428, + "val_auc": 0.7352, + "k": 2, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394 + ] + }, + { + "test_f1": 0.7200777176222631, + "test_acc": 0.7221499783268314, + "test_auc": 0.7868137674155734, + "val_auc": 0.7824, + "k": 5, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 844, + 1099, + 794 + ] + }, + { + "test_f1": 0.7317249116638467, + "test_acc": 0.7338534893801474, + "test_auc": 0.8143619012116685, + "val_auc": 0.7844, + "k": 10, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 844, + 1099, + 794, + 7649, + 1541, + 9605, + 7246, + 4675 + ] + }, + { + "test_f1": 0.7337140905122372, + "test_acc": 0.7355873428695275, + "test_auc": 0.8218444536970092, + "val_auc": 0.7984, + "k": 16, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 844, + 1099, + 794, + 7649, + 1541, + 9605, + 7246, + 4675, + 24246, + 28053, + 1737, + 100, + 1960, + 1560 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2ec075ab287852779cc020759b9480e257d05a1d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8076391072864068, + "test_acc": 0.8108474576271186, + "test_auc": 0.822181590201248, + "val_auc": 0.7745571658615137, + "k": 1, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.8066602996498157, + "test_acc": 0.8094915254237288, + "test_auc": 0.8752909510099172, + "val_auc": 0.8433977455716586, + "k": 2, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 1847 + ] + }, + { + "test_f1": 0.8425620730741159, + "test_acc": 0.8427118644067797, + "test_auc": 0.9231521623221659, + "val_auc": 0.9456521739130435, + "k": 5, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 1847, + 266, + 776, + 1433 + ] + }, + { + "test_f1": 0.8476705615800415, + "test_acc": 0.8481355932203389, + "test_auc": 0.9266215118053488, + "val_auc": 0.9468599033816425, + "k": 10, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 1847, + 266, + 776, + 1433, + 508, + 5020, + 32278, + 6838, + 1697 + ] + }, + { + "test_f1": 0.8636705385114931, + "test_acc": 0.8637288135593221, + "test_auc": 0.9424753541972325, + "val_auc": 0.9545088566827697, + "k": 16, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 1847, + 266, + 776, + 1433, + 508, + 5020, + 32278, + 6838, + 1697, + 6436, + 1180, + 4674, + 1628, + 27673, + 4413 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e3e4fce01e16fcaafaf73f9636fbb3abbf8596ae --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7647319404453181, + "test_acc": 0.7742372881355932, + "test_auc": 0.7784847014006097, + "val_auc": 0.663647342995169, + "k": 1, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528 + ] + }, + { + "test_f1": 0.7787747765418429, + "test_acc": 0.7810169491525424, + "test_auc": 0.8331844105415274, + "val_auc": 0.7622785829307568, + "k": 2, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 1815 + ] + }, + { + "test_f1": 0.8092615936732289, + "test_acc": 0.8108474576271186, + "test_auc": 0.8853220960974875, + "val_auc": 0.8184380032206119, + "k": 5, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 1815, + 1628, + 266, + 4185 + ] + }, + { + "test_f1": 0.8171841958401527, + "test_acc": 0.8183050847457627, + "test_auc": 0.9054450585211415, + "val_auc": 0.8393719806763285, + "k": 10, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 1815, + 1628, + 266, + 4185, + 4791, + 1789, + 1866, + 1433, + 416 + ] + }, + { + "test_f1": 0.8145383111934908, + "test_acc": 0.8149152542372882, + "test_auc": 0.9102418432596808, + "val_auc": 0.8464170692431562, + "k": 16, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 1815, + 1628, + 266, + 4185, + 4791, + 1789, + 1866, + 1433, + 416, + 1103, + 31149, + 20997, + 6436, + 203, + 16223 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..21a554e607cc0694636308b9255a044274546dc3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6471029736660664, + "test_acc": 0.6535593220338983, + "test_auc": 0.6796394965306505, + "val_auc": 0.6878019323671498, + "k": 1, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020 + ] + }, + { + "test_f1": 0.6754866928050837, + "test_acc": 0.6820338983050848, + "test_auc": 0.757458090184701, + "val_auc": 0.8049516908212561, + "k": 2, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266 + ] + }, + { + "test_f1": 0.7120119913400922, + "test_acc": 0.7152542372881356, + "test_auc": 0.8127672796402319, + "val_auc": 0.8711755233494364, + "k": 5, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 895, + 431 + ] + }, + { + "test_f1": 0.7829581740446301, + "test_acc": 0.7844067796610169, + "test_auc": 0.8642127499972422, + "val_auc": 0.9068035426731078, + "k": 10, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 895, + 431, + 776, + 32278, + 373, + 1628, + 27673 + ] + }, + { + "test_f1": 0.7945380294404495, + "test_acc": 0.7959322033898305, + "test_auc": 0.8731784536298552, + "val_auc": 0.9279388083735911, + "k": 16, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 895, + 431, + 776, + 32278, + 373, + 1628, + 27673, + 5460, + 1847, + 3927, + 18827, + 16223, + 800 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..37cbcca4c6b830f50f10d473a5a3ff5137dbcf5d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.58152537452829, + "test_acc": 0.5860058309037901, + "test_auc": 0.5988372093023256, + "val_auc": 0.5290116046418567, + "k": 1, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381 + ] + }, + { + "test_f1": 0.5750986108729206, + "test_acc": 0.5801749271137027, + "test_auc": 0.6109751121991024, + "val_auc": 0.5742296918767507, + "k": 2, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 993 + ] + }, + { + "test_f1": 0.5722981782091157, + "test_acc": 0.5830903790087464, + "test_auc": 0.6088161294709642, + "val_auc": 0.6050420168067226, + "k": 5, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 993, + 219, + 1122, + 1039 + ] + }, + { + "test_f1": 0.628879710048719, + "test_acc": 0.6297376093294461, + "test_auc": 0.6809805521555828, + "val_auc": 0.7302921168467387, + "k": 10, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 993, + 219, + 1122, + 1039, + 799, + 1799, + 1508, + 275, + 5897 + ] + }, + { + "test_f1": 0.681723318941124, + "test_acc": 0.6822157434402333, + "test_auc": 0.7260981912144704, + "val_auc": 0.789515806322529, + "k": 16, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 993, + 219, + 1122, + 1039, + 799, + 1799, + 1508, + 275, + 5897, + 1411, + 1697, + 801, + 1239, + 506, + 304 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b35004b5ff1b1dfe30da6d566fad4333c2ca0bd6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7223035219313956, + "test_acc": 0.7376093294460642, + "test_auc": 0.7429110567115463, + "val_auc": 0.697078831532613, + "k": 1, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7754599468877698, + "test_acc": 0.7813411078717201, + "test_auc": 0.7950836393308854, + "val_auc": 0.8041216486594639, + "k": 2, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275 + ] + }, + { + "test_f1": 0.8276366749856373, + "test_acc": 0.8279883381924198, + "test_auc": 0.8807119543043656, + "val_auc": 0.8679471788715486, + "k": 5, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 8006, + 191, + 534 + ] + }, + { + "test_f1": 0.7573863685074042, + "test_acc": 0.7580174927113703, + "test_auc": 0.8578131374949001, + "val_auc": 0.8791516606642658, + "k": 10, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 8006, + 191, + 534, + 1928, + 433, + 1122, + 506, + 799 + ] + }, + { + "test_f1": 0.8770219973235196, + "test_acc": 0.8775510204081632, + "test_auc": 0.9475044199646403, + "val_auc": 0.933173269307723, + "k": 16, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 8006, + 191, + 534, + 1928, + 433, + 1122, + 506, + 799, + 1799, + 801, + 1960, + 786, + 1411, + 821 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5de4c6685755efe5824cdad4d9193c73b1809bd9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6957690165777936, + "test_acc": 0.7026239067055393, + "test_auc": 0.6961444308445532, + "val_auc": 0.6940776310524209, + "k": 1, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.6965608935852123, + "test_acc": 0.6967930029154519, + "test_auc": 0.7453760369917041, + "val_auc": 0.7438975590236094, + "k": 2, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191 + ] + }, + { + "test_f1": 0.7551020408163265, + "test_acc": 0.7551020408163265, + "test_auc": 0.8053855569155446, + "val_auc": 0.793517406962785, + "k": 5, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 1239, + 534, + 275 + ] + }, + { + "test_f1": 0.769620551505346, + "test_acc": 0.7696793002915452, + "test_auc": 0.864613083095335, + "val_auc": 0.8475390156062425, + "k": 10, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 1239, + 534, + 275, + 433, + 5897, + 8006, + 1464, + 1195 + ] + }, + { + "test_f1": 0.7898087237229635, + "test_acc": 0.7900874635568513, + "test_auc": 0.860873113015096, + "val_auc": 0.8471388555422169, + "k": 16, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 1239, + 534, + 275, + 433, + 5897, + 8006, + 1464, + 1195, + 3233, + 1596, + 181, + 836, + 14404, + 1411 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..00a67cfb3f6576a2b339fee32806d1c6537b81da --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5412375481440526, + "test_acc": 0.5454545454545454, + "test_auc": 0.5648979591836735, + "val_auc": 0.5434343434343434, + "k": 1, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779 + ] + }, + { + "test_f1": 0.5141595992659822, + "test_acc": 0.5151515151515151, + "test_auc": 0.5502040816326531, + "val_auc": 0.5119191919191919, + "k": 2, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779, + 320 + ] + }, + { + "test_f1": 0.49391624923539823, + "test_acc": 0.494949494949495, + "test_auc": 0.5134693877551021, + "val_auc": 0.5133333333333333, + "k": 5, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779, + 320, + 1073, + 1703, + 5883 + ] + }, + { + "test_f1": 0.5440580924451892, + "test_acc": 0.5454545454545454, + "test_auc": 0.5318367346938775, + "val_auc": 0.5414141414141415, + "k": 10, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779, + 320, + 1073, + 1703, + 5883, + 859, + 987, + 1408, + 6804, + 1011 + ] + }, + { + "test_f1": 0.5004586635346351, + "test_acc": 0.5050505050505051, + "test_auc": 0.4893877551020408, + "val_auc": 0.5915151515151515, + "k": 16, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779, + 320, + 1073, + 1703, + 5883, + 859, + 987, + 1408, + 6804, + 1011, + 1030, + 2548, + 1377, + 4264, + 113, + 15517 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ae6e89523fd91fe1ce63f89f5c2a1e46cba45ee0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5603877760657421, + "test_acc": 0.5757575757575758, + "test_auc": 0.5616326530612245, + "val_auc": 0.5709090909090909, + "k": 1, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520 + ] + }, + { + "test_f1": 0.5597195942023528, + "test_acc": 0.5656565656565656, + "test_auc": 0.5563265306122449, + "val_auc": 0.6343434343434343, + "k": 2, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 2046 + ] + }, + { + "test_f1": 0.5547354106809319, + "test_acc": 0.5656565656565656, + "test_auc": 0.5495918367346939, + "val_auc": 0.6527272727272727, + "k": 5, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 2046, + 5129, + 30370, + 1818 + ] + }, + { + "test_f1": 0.5636261691770523, + "test_acc": 0.5757575757575758, + "test_auc": 0.5889795918367347, + "val_auc": 0.6565656565656566, + "k": 10, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 2046, + 5129, + 30370, + 1818, + 1030, + 320, + 6804, + 5883, + 2489 + ] + }, + { + "test_f1": 0.5458089668615984, + "test_acc": 0.5555555555555556, + "test_auc": 0.5697959183673469, + "val_auc": 0.6856565656565656, + "k": 16, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 2046, + 5129, + 30370, + 1818, + 1030, + 320, + 6804, + 5883, + 2489, + 859, + 5653, + 4712, + 1194, + 31245, + 1865 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4319c26bed15d1c95902c95c716bc7a94a3b5576 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.60581934482108, + "test_acc": 0.6060606060606061, + "test_auc": 0.590204081632653, + "val_auc": 0.5874747474747474, + "k": 1, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666 + ] + }, + { + "test_f1": 0.5655678969958852, + "test_acc": 0.5656565656565656, + "test_auc": 0.6006122448979592, + "val_auc": 0.6044444444444445, + "k": 2, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245 + ] + }, + { + "test_f1": 0.5655678969958852, + "test_acc": 0.5656565656565656, + "test_auc": 0.6006122448979593, + "val_auc": 0.6622222222222222, + "k": 5, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 2595 + ] + }, + { + "test_f1": 0.5656565656565656, + "test_acc": 0.5656565656565656, + "test_auc": 0.6061224489795919, + "val_auc": 0.6715151515151515, + "k": 10, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 2595, + 111, + 27174, + 4985, + 1116, + 7026 + ] + }, + { + "test_f1": 0.564322177225403, + "test_acc": 0.5656565656565656, + "test_auc": 0.5657142857142857, + "val_auc": 0.6226262626262626, + "k": 16, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 2595, + 111, + 27174, + 4985, + 1116, + 7026, + 5883, + 373, + 30370, + 680, + 1180, + 1011 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..87e40e74f2c442ce3c58e25897f2c10d3dccfa44 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5001917751225241, + "test_acc": 0.5532203389830509, + "test_auc": 0.5539743999882333, + "val_auc": 0.6183574879227053, + "k": 1, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217 + ] + }, + { + "test_f1": 0.5657685569379061, + "test_acc": 0.5816949152542373, + "test_auc": 0.6007435108272386, + "val_auc": 0.5599838969404187, + "k": 2, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1541 + ] + }, + { + "test_f1": 0.5789168587220413, + "test_acc": 0.5789830508474576, + "test_auc": 0.6165421966295647, + "val_auc": 0.6151368760064412, + "k": 5, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1541, + 186, + 1960, + 1319 + ] + }, + { + "test_f1": 0.5762672911744462, + "test_acc": 0.576271186440678, + "test_auc": 0.6201770158814207, + "val_auc": 0.6127214170692431, + "k": 10, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1541, + 186, + 1960, + 1319, + 539, + 311, + 20, + 1616, + 100 + ] + }, + { + "test_f1": 0.5726051562512278, + "test_acc": 0.5728813559322034, + "test_auc": 0.6109732196372168, + "val_auc": 0.6203703703703705, + "k": 16, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1541, + 186, + 1960, + 1319, + 539, + 311, + 20, + 1616, + 100, + 3780, + 22873, + 1812, + 253, + 836, + 26270 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..112e9469b33748be2e1b4c6b53dbe642d259e64a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6175603213749017, + "test_acc": 0.6176271186440678, + "test_auc": 0.6392520031034774, + "val_auc": 0.6302334943639292, + "k": 1, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319 + ] + }, + { + "test_f1": 0.6460802219249627, + "test_acc": 0.6461016949152543, + "test_auc": 0.7048736362533232, + "val_auc": 0.6690821256038647, + "k": 2, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541 + ] + }, + { + "test_f1": 0.6724721095133531, + "test_acc": 0.6732203389830509, + "test_auc": 0.7240148113828493, + "val_auc": 0.6940418679549114, + "k": 5, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1874, + 1960 + ] + }, + { + "test_f1": 0.6605869830218918, + "test_acc": 0.6610169491525424, + "test_auc": 0.7268259588973094, + "val_auc": 0.7391304347826088, + "k": 10, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1874, + 1960, + 186, + 1335, + 100, + 1255, + 908 + ] + }, + { + "test_f1": 0.6621657120544301, + "test_acc": 0.6630508474576271, + "test_auc": 0.7296830702364012, + "val_auc": 0.7493961352657005, + "k": 16, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1874, + 1960, + 186, + 1335, + 100, + 1255, + 908, + 253, + 311, + 1500, + 6478, + 217, + 765 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4eaf8ee4bb1f77daa761c20598d8c9a555d1edd7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5962311744930189, + "test_acc": 0.6040677966101695, + "test_auc": 0.6123714391825058, + "val_auc": 0.6479468599033817, + "k": 1, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.5943248813654936, + "test_acc": 0.6033898305084746, + "test_auc": 0.6123714391825058, + "val_auc": 0.6479468599033817, + "k": 2, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255 + ] + }, + { + "test_f1": 0.6076391665585991, + "test_acc": 0.6128813559322034, + "test_auc": 0.633352086573783, + "val_auc": 0.6829710144927535, + "k": 5, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 12485, + 20, + 521 + ] + }, + { + "test_f1": 0.6117636219363933, + "test_acc": 0.615593220338983, + "test_auc": 0.6484475258592477, + "val_auc": 0.7045088566827697, + "k": 10, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 12485, + 20, + 521, + 186, + 567, + 17442, + 1319, + 4511 + ] + }, + { + "test_f1": 0.6228911490099195, + "test_acc": 0.6250847457627119, + "test_auc": 0.6556316716491453, + "val_auc": 0.6936392914653784, + "k": 16, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 12485, + 20, + 521, + 186, + 567, + 17442, + 1319, + 4511, + 1942, + 5144, + 6212, + 311, + 1500, + 1596 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..38794b89e476ad9e495fc6259228c39ac7894caf --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5706085637243745, + "test_acc": 0.5707196029776674, + "test_auc": 0.5891335402196937, + "val_auc": 0.5782312925170068, + "k": 1, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067 + ] + }, + { + "test_f1": 0.5864343954485649, + "test_acc": 0.5880893300248139, + "test_auc": 0.6421604847051869, + "val_auc": 0.60984393757503, + "k": 2, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268 + ] + }, + { + "test_f1": 0.6029043187034281, + "test_acc": 0.6029776674937966, + "test_auc": 0.6513472242746663, + "val_auc": 0.6234493797519008, + "k": 5, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 24630, + 2, + 1441 + ] + }, + { + "test_f1": 0.5979851850040949, + "test_acc": 0.598014888337469, + "test_auc": 0.6355967686320871, + "val_auc": 0.6070428171268508, + "k": 10, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 24630, + 2, + 1441, + 124, + 8013, + 2480, + 1619, + 1101 + ] + }, + { + "test_f1": 0.6148153524520289, + "test_acc": 0.6153846153846154, + "test_auc": 0.6497463179153736, + "val_auc": 0.6526610644257702, + "k": 16, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 24630, + 2, + 1441, + 124, + 8013, + 2480, + 1619, + 1101, + 5731, + 12744, + 1955, + 710, + 1551, + 1120 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7cec08afbb793255dcc5afb690f0c687bd6b3e9a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5570576057084091, + "test_acc": 0.598014888337469, + "test_auc": 0.6031722575242598, + "val_auc": 0.6162464985994398, + "k": 1, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575 + ] + }, + { + "test_f1": 0.6159454959265206, + "test_acc": 0.6178660049627791, + "test_auc": 0.6490566967144475, + "val_auc": 0.6384553821528611, + "k": 2, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711 + ] + }, + { + "test_f1": 0.6367776841131992, + "test_acc": 0.6377171215880894, + "test_auc": 0.6786857790256636, + "val_auc": 0.6538615446178472, + "k": 5, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 1625, + 583, + 24630 + ] + }, + { + "test_f1": 0.6470515055171582, + "test_acc": 0.6476426799007444, + "test_auc": 0.6897936062262943, + "val_auc": 0.6634653861544618, + "k": 10, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 1625, + 583, + 24630, + 1381, + 1619, + 2, + 1551, + 765 + ] + }, + { + "test_f1": 0.6300679399815874, + "test_acc": 0.630272952853598, + "test_auc": 0.7010492093985518, + "val_auc": 0.682673069227691, + "k": 16, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 1625, + 583, + 24630, + 1381, + 1619, + 2, + 1551, + 765, + 434, + 1799, + 680, + 10, + 4802, + 433 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..70b4d1fe7c30e1d5bad2019ff47a9593d720d551 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5500205717200584, + "test_acc": 0.5508684863523573, + "test_auc": 0.5761538840451209, + "val_auc": 0.54421768707483, + "k": 1, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619 + ] + }, + { + "test_f1": 0.5630876214123848, + "test_acc": 0.56575682382134, + "test_auc": 0.6006600660066006, + "val_auc": 0.6052420968387355, + "k": 2, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 32279 + ] + }, + { + "test_f1": 0.6200441792768254, + "test_acc": 0.6228287841191067, + "test_auc": 0.6800773360918181, + "val_auc": 0.6824729891956782, + "k": 5, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 32279, + 1042, + 575, + 125 + ] + }, + { + "test_f1": 0.6773796242439268, + "test_acc": 0.6774193548387096, + "test_auc": 0.7438796118417812, + "val_auc": 0.740296118447379, + "k": 10, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 32279, + 1042, + 575, + 125, + 1067, + 347, + 1662, + 433, + 2480 + ] + }, + { + "test_f1": 0.6823703993448154, + "test_acc": 0.6823821339950372, + "test_auc": 0.7390768927639033, + "val_auc": 0.7603041216486595, + "k": 16, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 32279, + 1042, + 575, + 125, + 1067, + 347, + 1662, + 433, + 2480, + 1625, + 772, + 24630, + 124, + 1559, + 10 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c323235a1a75113bceb1b55af4617e29259951c6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8345124727743573, + "test_acc": 0.8366800535475234, + "test_auc": 0.8505720348095368, + "val_auc": 0.8426262626262626, + "k": 1, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731 + ] + }, + { + "test_f1": 0.8280884883921245, + "test_acc": 0.8299866131191432, + "test_auc": 0.8892560680133619, + "val_auc": 0.9050505050505051, + "k": 2, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226 + ] + }, + { + "test_f1": 0.855359950866758, + "test_acc": 0.856760374832664, + "test_auc": 0.9512623474932259, + "val_auc": 0.9563636363636364, + "k": 5, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 737, + 1274 + ] + }, + { + "test_f1": 0.8566733333519457, + "test_acc": 0.85809906291834, + "test_auc": 0.9537676879184528, + "val_auc": 0.9624242424242424, + "k": 10, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 737, + 1274, + 2790, + 4148, + 23697, + 6567, + 5872 + ] + }, + { + "test_f1": 0.8627230875847964, + "test_acc": 0.8634538152610441, + "test_auc": 0.9557640750670241, + "val_auc": 0.9632323232323232, + "k": 16, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 737, + 1274, + 2790, + 4148, + 23697, + 6567, + 5872, + 8053, + 292, + 13701, + 416, + 1101, + 5964 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2637bc789952396740f13d85c6447fc0e8ba3009 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8653148318811644, + "test_acc": 0.8674698795180723, + "test_auc": 0.8685932818167482, + "val_auc": 0.8777777777777778, + "k": 1, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226 + ] + }, + { + "test_f1": 0.8653148318811644, + "test_acc": 0.8674698795180723, + "test_auc": 0.9122879958710269, + "val_auc": 0.9171717171717172, + "k": 2, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731 + ] + }, + { + "test_f1": 0.8664745511881633, + "test_acc": 0.8674698795180723, + "test_auc": 0.9626887069719431, + "val_auc": 0.958989898989899, + "k": 5, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4505, + 1065, + 4648 + ] + }, + { + "test_f1": 0.8664113677918267, + "test_acc": 0.8674698795180723, + "test_auc": 0.9631833235365802, + "val_auc": 0.9589898989898991, + "k": 10, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4505, + 1065, + 4648, + 32, + 1380, + 6567, + 1682, + 3825 + ] + }, + { + "test_f1": 0.8664113677918267, + "test_acc": 0.8674698795180723, + "test_auc": 0.9631833235365802, + "val_auc": 0.9589898989898991, + "k": 16, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4505, + 1065, + 4648, + 32, + 1380, + 6567, + 1682, + 3825, + 22775, + 805, + 2046, + 4148, + 4478, + 7481 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d2a894e7660c67b8eac1b85ed1923d4a00166e1a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8738885933440323, + "test_acc": 0.8755020080321285, + "test_auc": 0.8793673208986251, + "val_auc": 0.8606060606060606, + "k": 1, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648 + ] + }, + { + "test_f1": 0.8738885933440323, + "test_acc": 0.8755020080321285, + "test_auc": 0.9120048458086623, + "val_auc": 0.8846464646464647, + "k": 2, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505 + ] + }, + { + "test_f1": 0.8886909980619866, + "test_acc": 0.8888888888888888, + "test_auc": 0.9740362145345587, + "val_auc": 0.9630303030303031, + "k": 5, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32 + ] + }, + { + "test_f1": 0.8981973091728039, + "test_acc": 0.8982597054886211, + "test_auc": 0.9757852934008114, + "val_auc": 0.9648484848484848, + "k": 10, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32, + 1682, + 1065, + 2832, + 1101, + 1643 + ] + }, + { + "test_f1": 0.8914588881706655, + "test_acc": 0.891566265060241, + "test_auc": 0.9737494802941894, + "val_auc": 0.9676767676767677, + "k": 16, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32, + 1682, + 1065, + 2832, + 1101, + 1643, + 1305, + 1425, + 23, + 2790, + 1380, + 2302 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1ff4503710e77b025f513a0c7216d1771654c320 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.725434624778192, + "test_acc": 0.7425641025641025, + "test_auc": 0.7423881576732757, + "val_auc": 0.6923704610363117, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518 + ] + }, + { + "test_f1": 0.757071995447694, + "test_acc": 0.7661538461538462, + "test_auc": 0.7801991113205642, + "val_auc": 0.7384740922072623, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541 + ] + }, + { + "test_f1": 0.7593820864376043, + "test_acc": 0.7671794871794871, + "test_auc": 0.8570538930218466, + "val_auc": 0.861485108119135, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 821 + ] + }, + { + "test_f1": 0.7692907895317534, + "test_acc": 0.7743589743589744, + "test_auc": 0.8776003972127782, + "val_auc": 0.8714810281517748, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 821, + 1335, + 253, + 2632, + 505, + 1030 + ] + }, + { + "test_f1": 0.7649489054092634, + "test_acc": 0.7702564102564102, + "test_auc": 0.8775709428754166, + "val_auc": 0.8639330885352918, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 821, + 1335, + 253, + 2632, + 505, + 1030, + 374, + 2581, + 26067, + 549, + 373, + 4558 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bfd669c9dd334ac5867d275234baa7924c246168 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8422685801123586, + "test_acc": 0.8450420168067226, + "test_auc": 0.8505940372113876, + "val_auc": 0.8396226415094339, + "k": 1, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852 + ] + }, + { + "test_f1": 0.8956118546458784, + "test_acc": 0.8961344537815126, + "test_auc": 0.9069403468049259, + "val_auc": 0.9339622641509434, + "k": 2, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782 + ] + }, + { + "test_f1": 0.9014901293591318, + "test_acc": 0.9018487394957984, + "test_auc": 0.9615477959520142, + "val_auc": 0.9855479727017262, + "k": 5, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 173 + ] + }, + { + "test_f1": 0.907953876483314, + "test_acc": 0.908235294117647, + "test_auc": 0.9726141343254442, + "val_auc": 0.9791248494580489, + "k": 10, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 173, + 1275, + 8815, + 11077, + 1521, + 23246 + ] + }, + { + "test_f1": 0.9160934650413246, + "test_acc": 0.9163025210084034, + "test_auc": 0.9775050889067256, + "val_auc": 0.9813327980730631, + "k": 16, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 173, + 1275, + 8815, + 11077, + 1521, + 23246, + 1688, + 259, + 6919, + 1152, + 1283, + 19562 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d7955f783acade8a243eaa980eac7a1b7490387b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.745761953429119, + "test_acc": 0.7519327731092437, + "test_auc": 0.7625846493987317, + "val_auc": 0.7834203131272581, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021 + ] + }, + { + "test_f1": 0.7711394454346422, + "test_acc": 0.7717647058823529, + "test_auc": 0.8791025355952303, + "val_auc": 0.8900040144520274, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914 + ] + }, + { + "test_f1": 0.777972914680097, + "test_acc": 0.7791596638655462, + "test_auc": 0.8939525619888495, + "val_auc": 0.9281413087113609, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 1180, + 4790 + ] + }, + { + "test_f1": 0.7963601111480981, + "test_acc": 0.7976470588235294, + "test_auc": 0.9096242253653528, + "val_auc": 0.9514251304696909, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 1180, + 4790, + 6919, + 1152, + 1275, + 1067, + 2962 + ] + }, + { + "test_f1": 0.825880188890273, + "test_acc": 0.8258823529411765, + "test_auc": 0.9250922872782756, + "val_auc": 0.9702930549979929, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 1180, + 4790, + 6919, + 1152, + 1275, + 1067, + 2962, + 1500, + 1091, + 984, + 98, + 347, + 8815 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..067c0fb13ed670f033cce60cf7242344188179e7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7652516882980215, + "test_acc": 0.772436974789916, + "test_auc": 0.7901169454266727, + "val_auc": 0.7462866318747491, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235 + ] + }, + { + "test_f1": 0.804054927331362, + "test_acc": 0.8077310924369748, + "test_auc": 0.831229526867258, + "val_auc": 0.8012846246487355, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 764 + ] + }, + { + "test_f1": 0.8424795734533714, + "test_acc": 0.8440336134453782, + "test_auc": 0.8474014939511609, + "val_auc": 0.8450421517462867, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 764, + 5189, + 8189, + 1216 + ] + }, + { + "test_f1": 0.8576623663498278, + "test_acc": 0.8584873949579832, + "test_auc": 0.9189404046539542, + "val_auc": 0.9201124046567644, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 764, + 5189, + 8189, + 1216, + 692, + 576, + 42, + 3494, + 964 + ] + }, + { + "test_f1": 0.8487852294728858, + "test_acc": 0.8494117647058823, + "test_auc": 0.9323414936619159, + "val_auc": 0.937173825772782, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 764, + 5189, + 8189, + 1216, + 692, + 576, + 42, + 3494, + 964, + 25737, + 2046, + 1980, + 1021, + 197, + 30916 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f5d376f2ad29a53dd56bfddf04a9afd7b42617a0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9425208135722886, + "test_acc": 0.9425210084033614, + "test_auc": 0.9847432226247549, + "val_auc": 0.9839421918908069, + "k": 1, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9875628706353862, + "test_acc": 0.987563025210084, + "test_auc": 0.9994653484319298, + "val_auc": 0.9991971095945404, + "k": 2, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847 + ] + }, + { + "test_f1": 0.9868900067916023, + "test_acc": 0.986890756302521, + "test_auc": 0.9991968927840569, + "val_auc": 0.9991971095945403, + "k": 5, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1758, + 431 + ] + }, + { + "test_f1": 0.9969747755599927, + "test_acc": 0.9969747899159664, + "test_auc": 0.9993568815034961, + "val_auc": 1.0, + "k": 10, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1758, + 431, + 786, + 1411, + 461, + 1697, + 279 + ] + }, + { + "test_f1": 0.9973109073552354, + "test_acc": 0.9973109243697479, + "test_auc": 0.9999624885205833, + "val_auc": 1.0, + "k": 16, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1758, + 431, + 786, + 1411, + 461, + 1697, + 279, + 4148, + 27673, + 4674, + 1634, + 126, + 1360 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..decb25425d63b59b53cc99e7b5c9541752e2a52e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9378127230556642, + "test_acc": 0.9378151260504202, + "test_auc": 0.9856218047450666, + "val_auc": 0.9863508631071859, + "k": 1, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.959653581409563, + "test_acc": 0.9596638655462185, + "test_auc": 0.9951976267436058, + "val_auc": 0.9967884383781613, + "k": 2, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9616572698854934, + "test_acc": 0.9616806722689075, + "test_auc": 0.9960161001077438, + "val_auc": 0.9971898835808912, + "k": 5, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1758, + 1411 + ] + }, + { + "test_f1": 0.9922684655437655, + "test_acc": 0.9922689075630252, + "test_auc": 0.9996176540772718, + "val_auc": 1.0, + "k": 10, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1758, + 1411, + 431, + 1847, + 126, + 279, + 461 + ] + }, + { + "test_f1": 0.9979831928215676, + "test_acc": 0.9979831932773109, + "test_auc": 0.9999850857973404, + "val_auc": 1.0, + "k": 16, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1758, + 1411, + 431, + 1847, + 126, + 279, + 461, + 1551, + 1360, + 1628, + 1914, + 1634, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7a2358d8bb8ba714f84d7e975791d0e6a4872f97 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9233595260756386, + "test_acc": 0.9233613445378152, + "test_auc": 0.9778948919307836, + "val_auc": 0.9763147330389402, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9640293216694888, + "test_acc": 0.9640336134453782, + "test_auc": 0.9954950069057278, + "val_auc": 0.9963869931754316, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9815106622384483, + "test_acc": 0.9815126050420168, + "test_auc": 0.9989908056200332, + "val_auc": 1.0, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 461, + 1758 + ] + }, + { + "test_f1": 0.9973109073552354, + "test_acc": 0.9973109243697479, + "test_auc": 0.9996949367637807, + "val_auc": 1.0, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 461, + 1758, + 1411, + 1847, + 431, + 279, + 27673 + ] + }, + { + "test_f1": 0.9973109237620901, + "test_acc": 0.9973109243697479, + "test_auc": 0.9999855377428756, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 461, + 1758, + 1411, + 1847, + 431, + 279, + 27673, + 4148, + 1360, + 1634, + 1697, + 1914, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..27618140cb45f171a3050644989763484714c59a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7420501751367465, + "test_acc": 0.7421383647798742, + "test_auc": 0.8119284402945098, + "val_auc": 0.8643457382953181, + "k": 1, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585 + ] + }, + { + "test_f1": 0.7787120476922823, + "test_acc": 0.7788679245283019, + "test_auc": 0.844581581081861, + "val_auc": 0.9199679871948779, + "k": 2, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585, + 776 + ] + }, + { + "test_f1": 0.8205859865377818, + "test_acc": 0.82062893081761, + "test_auc": 0.9019613402609922, + "val_auc": 0.9575830332132853, + "k": 5, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585, + 776, + 559, + 20310, + 32700 + ] + }, + { + "test_f1": 0.8469170091298825, + "test_acc": 0.8470440251572326, + "test_auc": 0.924410200508537, + "val_auc": 0.9595838335334133, + "k": 10, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585, + 776, + 559, + 20310, + 32700, + 508, + 30020, + 1866, + 1499, + 1327 + ] + }, + { + "test_f1": 0.8613546386141514, + "test_acc": 0.8613836477987421, + "test_auc": 0.9335752815838159, + "val_auc": 0.9663865546218487, + "k": 16, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585, + 776, + 559, + 20310, + 32700, + 508, + 30020, + 1866, + 1499, + 1327, + 400, + 431, + 16910, + 10363, + 28521, + 27673 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e713a8f6b59701f254bd60b4d031961a8deeb5a9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7158594049946407, + "test_acc": 0.716323296354992, + "test_auc": 0.7326602370906169, + "val_auc": 0.7406962785114046, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.722662440570523, + "test_acc": 0.722662440570523, + "test_auc": 0.8038276069921639, + "val_auc": 0.8225290116046419, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 431 + ] + }, + { + "test_f1": 0.7953672902773752, + "test_acc": 0.7955625990491284, + "test_auc": 0.8645670082378943, + "val_auc": 0.8967587034813925, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 431, + 1327, + 508, + 1042 + ] + }, + { + "test_f1": 0.7952282862810539, + "test_acc": 0.7955625990491284, + "test_auc": 0.8806007635121559, + "val_auc": 0.8995598239295719, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 431, + 1327, + 508, + 1042, + 28521, + 16910, + 5186, + 21219, + 468 + ] + }, + { + "test_f1": 0.8175754913325436, + "test_acc": 0.8177496038034865, + "test_auc": 0.899291742013261, + "val_auc": 0.9095638255302121, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 431, + 1327, + 508, + 1042, + 28521, + 16910, + 5186, + 21219, + 468, + 1866, + 1681, + 913, + 11098, + 272, + 435 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..992909469a6bc986e804828d177ed862d169a1bc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6981914314720956, + "test_acc": 0.6986794717887155, + "test_auc": 0.7061024257517062, + "val_auc": 0.7334943639291466, + "k": 1, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6902359334066399, + "test_acc": 0.6902761104441777, + "test_auc": 0.7404912608374838, + "val_auc": 0.7357085346215781, + "k": 2, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447 + ] + }, + { + "test_f1": 0.7250377284442667, + "test_acc": 0.7250900360144058, + "test_auc": 0.796065071020107, + "val_auc": 0.8059581320450886, + "k": 5, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 2837, + 5186 + ] + }, + { + "test_f1": 0.73690298755624, + "test_acc": 0.737094837935174, + "test_auc": 0.8312609527762405, + "val_auc": 0.8305152979066023, + "k": 10, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 2837, + 5186, + 347, + 776, + 1417, + 1866, + 1042 + ] + }, + { + "test_f1": 0.7970732263672826, + "test_acc": 0.7971188475390156, + "test_auc": 0.8633093525179857, + "val_auc": 0.8144122383252818, + "k": 16, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 2837, + 5186, + 347, + 776, + 1417, + 1866, + 1042, + 1528, + 1377, + 32700, + 21219, + 30737, + 1623 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2cdae81f878826fca5025e726852e9cba77ff24f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7189186679482901, + "test_acc": 0.71900826446281, + "test_auc": 0.7513205027017181, + "val_auc": 0.7174869947979191, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7355251471495883, + "test_acc": 0.7355371900826446, + "test_auc": 0.7635541254325784, + "val_auc": 0.7707082833133253, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186 + ] + }, + { + "test_f1": 0.7272685876715779, + "test_acc": 0.7272727272727273, + "test_auc": 0.7650112318620607, + "val_auc": 0.7763105242096838, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 1377, + 3507, + 508 + ] + }, + { + "test_f1": 0.7244007867092648, + "test_acc": 0.7245179063360881, + "test_auc": 0.7828000728553214, + "val_auc": 0.7779111644657863, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 1377, + 3507, + 508, + 6501, + 1681, + 1447, + 3301, + 848 + ] + }, + { + "test_f1": 0.743801652892562, + "test_acc": 0.743801652892562, + "test_auc": 0.8191973772084269, + "val_auc": 0.7879151660664265, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 1377, + 3507, + 508, + 6501, + 1681, + 1447, + 3301, + 848, + 12612, + 601, + 3078, + 4790, + 1585, + 1688 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7252dc505c31c921554928d9cd7457f0bc4e2749 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6194619700976733, + "test_acc": 0.6259119496855345, + "test_auc": 0.6349678341817386, + "val_auc": 0.6590636254501799, + "k": 1, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.622876949790739, + "test_acc": 0.6256603773584906, + "test_auc": 0.683568699565283, + "val_auc": 0.7382953181272509, + "k": 2, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 4605 + ] + }, + { + "test_f1": 0.6653960990367009, + "test_acc": 0.6654088050314465, + "test_auc": 0.7137082687367283, + "val_auc": 0.7458983593437376, + "k": 5, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 4605, + 2837, + 776, + 1585 + ] + }, + { + "test_f1": 0.6790191818062118, + "test_acc": 0.6792452830188679, + "test_auc": 0.7368478611983932, + "val_auc": 0.8027210884353742, + "k": 10, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 4605, + 2837, + 776, + 1585, + 27673, + 4674, + 126, + 32700, + 1499 + ] + }, + { + "test_f1": 0.6848890531694913, + "test_acc": 0.6850314465408806, + "test_auc": 0.7394924402985603, + "val_auc": 0.8339335734293717, + "k": 16, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 4605, + 2837, + 776, + 1585, + 27673, + 4674, + 126, + 32700, + 1499, + 3832, + 1804, + 559, + 7976, + 11717, + 2135 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9426a34d3b4ca028687f1f8ed0335f37aa276650 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7963513676993724, + "test_acc": 0.7972327044025157, + "test_auc": 0.834122500478462, + "val_auc": 0.8259303721488597, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293 + ] + }, + { + "test_f1": 0.8026581130416162, + "test_acc": 0.8027672955974843, + "test_auc": 0.8859524788388103, + "val_auc": 0.8857543017206883, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062 + ] + }, + { + "test_f1": 0.819255533829998, + "test_acc": 0.8193710691823899, + "test_auc": 0.9075016277838142, + "val_auc": 0.9247699079631853, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1166, + 1441 + ] + }, + { + "test_f1": 0.8387009779492156, + "test_acc": 0.8387421383647798, + "test_auc": 0.9167880964701142, + "val_auc": 0.9327731092436975, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1166, + 1441, + 1221, + 1116, + 1313, + 1305, + 1928 + ] + }, + { + "test_f1": 0.8334006008588818, + "test_acc": 0.8334591194968554, + "test_auc": 0.9197439291005216, + "val_auc": 0.9499799919967987, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1166, + 1441, + 1221, + 1116, + 1313, + 1305, + 1928, + 409, + 524, + 14, + 765, + 334, + 61 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1a163b9ad9e620dd933838a4087e93187ba28792 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6628869875957805, + "test_acc": 0.6666666666666666, + "test_auc": 0.676734693877551, + "val_auc": 0.6073865917302288, + "k": 1, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377 + ] + }, + { + "test_f1": 0.6723599632690541, + "test_acc": 0.6767676767676768, + "test_auc": 0.7073469387755101, + "val_auc": 0.6567643516659976, + "k": 2, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027 + ] + }, + { + "test_f1": 0.6963497495412388, + "test_acc": 0.696969696969697, + "test_auc": 0.7461224489795919, + "val_auc": 0.7035327177840225, + "k": 5, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339 + ] + }, + { + "test_f1": 0.6133183688739244, + "test_acc": 0.6161616161616161, + "test_auc": 0.7208163265306122, + "val_auc": 0.723404255319149, + "k": 10, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339, + 681, + 1551, + 524, + 56, + 1790 + ] + }, + { + "test_f1": 0.6444313734033361, + "test_acc": 0.6464646464646465, + "test_auc": 0.7171428571428571, + "val_auc": 0.7784022480931352, + "k": 16, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339, + 681, + 1551, + 524, + 56, + 1790, + 6086, + 505, + 957, + 1221, + 11386, + 1915 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..43e4a56aed651abd5a9770dbdacad517a588cf71 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6776449286148082, + "test_acc": 0.6797331109257715, + "test_auc": 0.7071090706733444, + "val_auc": 0.803341384863124, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6796904328491036, + "test_acc": 0.6839032527105922, + "test_auc": 0.7232595993322204, + "val_auc": 0.8522544283413849, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114 + ] + }, + { + "test_f1": 0.7435598297078191, + "test_acc": 0.7456213511259383, + "test_auc": 0.8046048970506399, + "val_auc": 0.891304347826087, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 363 + ] + }, + { + "test_f1": 0.7615264389559353, + "test_acc": 0.7631359466221852, + "test_auc": 0.8399554813578185, + "val_auc": 0.8856682769726247, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 363, + 3276, + 238, + 30704, + 6086, + 1335 + ] + }, + { + "test_f1": 0.780637656397324, + "test_acc": 0.7814845704753962, + "test_auc": 0.8674874791318864, + "val_auc": 0.894927536231884, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 363, + 3276, + 238, + 30704, + 6086, + 1335, + 801, + 756, + 4534, + 11788, + 4481, + 1670 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d9fc86aa2a1d5263e707c6e66b406be302059515 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.5732098851035412, + "test_acc": 0.5939622641509434, + "test_auc": 0.643749512677474, + "val_auc": 0.5522208883553421, + "k": 2, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727 + ] + }, + { + "test_f1": 0.610459057922914, + "test_acc": 0.6115723270440252, + "test_auc": 0.6657162400674809, + "val_auc": 0.5878351340536214, + "k": 5, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1928 + ] + }, + { + "test_f1": 0.631491082103987, + "test_acc": 0.6322012578616352, + "test_auc": 0.6938665207146251, + "val_auc": 0.5930372148859544, + "k": 10, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1928, + 1438, + 7143, + 486, + 30704, + 293 + ] + }, + { + "test_f1": 0.6313694111973354, + "test_acc": 0.6314465408805031, + "test_auc": 0.6935299517284887, + "val_auc": 0.5704281712685074, + "k": 16, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1928, + 1438, + 7143, + 486, + 30704, + 293, + 2034, + 765, + 8609, + 140, + 1608, + 4281 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dac5df36bb3ca26471179bde7a082ddc93ce3d3c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5202061768302847, + "test_acc": 0.542788749251945, + "test_auc": 0.5481369796292582, + "val_auc": 0.5692492974708953, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498 + ] + }, + { + "test_f1": 0.5431188517217843, + "test_acc": 0.5433871932974267, + "test_auc": 0.5824699882531588, + "val_auc": 0.5283018867924528, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727 + ] + }, + { + "test_f1": 0.5593919119567209, + "test_acc": 0.5607420706163974, + "test_auc": 0.5831747987279031, + "val_auc": 0.5499799277398635, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 5392, + 293, + 7400 + ] + }, + { + "test_f1": 0.5552079307695571, + "test_acc": 0.5553560742070617, + "test_auc": 0.5839283156175687, + "val_auc": 0.5660377358490566, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 5392, + 293, + 7400, + 150, + 860, + 19092, + 672, + 1641 + ] + }, + { + "test_f1": 0.5517217421463793, + "test_acc": 0.5517654099341711, + "test_auc": 0.5801234850872418, + "val_auc": 0.616619831393015, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 5392, + 293, + 7400, + 150, + 860, + 19092, + 672, + 1641, + 1336, + 100, + 26677, + 4913, + 2363, + 125 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a3990985269e6f2fa6a6d42a85b0fa86df42392d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.559669373211052, + "test_acc": 0.5949685534591195, + "test_auc": 0.6024339291916572, + "val_auc": 0.5982392957182873, + "k": 1, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659 + ] + }, + { + "test_f1": 0.5846464342109482, + "test_acc": 0.6002515723270441, + "test_auc": 0.6082793945353046, + "val_auc": 0.661064425770308, + "k": 2, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794 + ] + }, + { + "test_f1": 0.5930932930841596, + "test_acc": 0.6110691823899371, + "test_auc": 0.6261416004836265, + "val_auc": 0.7014805922368949, + "k": 5, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 777 + ] + }, + { + "test_f1": 0.5967839873007303, + "test_acc": 0.609056603773585, + "test_auc": 0.6638783379694372, + "val_auc": 0.7821128451380552, + "k": 10, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 777, + 7143, + 1377, + 6400, + 4252, + 3053 + ] + }, + { + "test_f1": 0.6204316901811385, + "test_acc": 0.6279245283018868, + "test_auc": 0.7057279509973784, + "val_auc": 0.8093237294917968, + "k": 16, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 777, + 7143, + 1377, + 6400, + 4252, + 3053, + 1224, + 1550, + 1596, + 448, + 121, + 15353 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5e23fdd32f772279a26da357615b53f1757870dc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9952200531821748, + "test_acc": 0.9952201257861635, + "test_auc": 0.9994425536611719, + "val_auc": 1.0, + "k": 1, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9987421359765241, + "test_acc": 0.9987421383647799, + "test_auc": 0.99998784858117, + "val_auc": 1.0, + "k": 2, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.9999648115163047, + "val_auc": 1.0, + "k": 5, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1525 + ] + }, + { + "test_f1": 0.9964779406110091, + "test_acc": 0.9964779874213836, + "test_auc": 0.9999706340711607, + "val_auc": 1.0, + "k": 10, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1525, + 347, + 1551, + 6718, + 1255, + 1232 + ] + }, + { + "test_f1": 0.9992452824456872, + "test_acc": 0.999245283018868, + "test_auc": 0.9999969621452924, + "val_auc": 1.0, + "k": 16, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1525, + 347, + 1551, + 6718, + 1255, + 1232, + 1042, + 15467, + 1335, + 1758, + 1472, + 374 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..22dcc5937525da046aa4a52072e8c29d4df820d6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6435217418462762, + "test_acc": 0.6435220125786163, + "test_auc": 0.6846792886154369, + "val_auc": 0.7555022008803521, + "k": 1, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6420618547142569, + "test_acc": 0.6422641509433962, + "test_auc": 0.6924840436681489, + "val_auc": 0.7440976390556222, + "k": 2, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928 + ] + }, + { + "test_f1": 0.6789377097019825, + "test_acc": 0.6789937106918239, + "test_auc": 0.7494947035003174, + "val_auc": 0.8267306922769108, + "k": 5, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 6400 + ] + }, + { + "test_f1": 0.6895586481732209, + "test_acc": 0.689559748427673, + "test_auc": 0.7702924897143303, + "val_auc": 0.8403361344537815, + "k": 10, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 6400, + 1616, + 56, + 5827, + 16883, + 152 + ] + }, + { + "test_f1": 0.7184693630323427, + "test_acc": 0.7184905660377359, + "test_auc": 0.7984261887378625, + "val_auc": 0.8663465386154462, + "k": 16, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 6400, + 1616, + 56, + 5827, + 16883, + 152, + 1691, + 1327, + 30704, + 406, + 1255, + 1135 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a4458e7994b4b0a947bb5382a4b4dcc5626de561 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.665964642317666, + "test_acc": 0.6674213836477988, + "test_auc": 0.716092478373006, + "val_auc": 0.7222889155662265, + "k": 1, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6650597288794977, + "test_acc": 0.6671698113207547, + "test_auc": 0.716092478373006, + "val_auc": 0.7222889155662265, + "k": 2, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143 + ] + }, + { + "test_f1": 0.6817549817090655, + "test_acc": 0.6830188679245283, + "test_auc": 0.7472384634935936, + "val_auc": 0.7242897158863545, + "k": 5, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 13614, + 1011 + ] + }, + { + "test_f1": 0.6998313262132867, + "test_acc": 0.7001257861635221, + "test_auc": 0.7741658810436853, + "val_auc": 0.7438975590236094, + "k": 10, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 13614, + 1011, + 19940, + 1928, + 30704, + 16883, + 253 + ] + }, + { + "test_f1": 0.7080992484713896, + "test_acc": 0.7081761006289308, + "test_auc": 0.7827554152291707, + "val_auc": 0.7535014005602242, + "k": 16, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 13614, + 1011, + 19940, + 1928, + 30704, + 16883, + 253, + 598, + 1135, + 1019, + 6400, + 9390, + 5827 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..951efcac2258904bff08a4dcb77567c4a32cf570 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7909637254140568, + "test_acc": 0.7910643889618922, + "test_auc": 0.8179178318809252, + "val_auc": 0.774, + "k": 1, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062 + ] + }, + { + "test_f1": 0.8024350066040075, + "test_acc": 0.8031537450722733, + "test_auc": 0.8916912970996594, + "val_auc": 0.8240000000000001, + "k": 2, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448 + ] + }, + { + "test_f1": 0.8304622098259071, + "test_acc": 0.8307490144546649, + "test_auc": 0.9157431981049348, + "val_auc": 0.8411999999999998, + "k": 5, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 6778, + 1030, + 811 + ] + }, + { + "test_f1": 0.8464838936705362, + "test_acc": 0.8465177398160315, + "test_auc": 0.924235655362914, + "val_auc": 0.8847999999999999, + "k": 10, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 6778, + 1030, + 811, + 7319, + 140, + 296, + 186, + 1662 + ] + }, + { + "test_f1": 0.8504453974680206, + "test_acc": 0.8504599211563731, + "test_auc": 0.9290792721437677, + "val_auc": 0.9032, + "k": 16, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 6778, + 1030, + 811, + 7319, + 140, + 296, + 186, + 1662, + 2831, + 2046, + 817, + 24630, + 628, + 1726 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..811c4ac7224c93791370a42d78b93de2ee8fa1ce --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5187672743492777, + "test_acc": 0.5240251572327044, + "test_auc": 0.535953516772502, + "val_auc": 0.6032412965186076, + "k": 1, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011 + ] + }, + { + "test_f1": 0.5204626862135454, + "test_acc": 0.5235220125786163, + "test_auc": 0.5408361087511481, + "val_auc": 0.6462585034013606, + "k": 2, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411 + ] + }, + { + "test_f1": 0.5266152905725795, + "test_acc": 0.529559748427673, + "test_auc": 0.5373967509131286, + "val_auc": 0.6436574629851941, + "k": 5, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928 + ] + }, + { + "test_f1": 0.5266152905725795, + "test_acc": 0.529559748427673, + "test_auc": 0.5373967509131286, + "val_auc": 0.6436574629851941, + "k": 10, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928, + 1703, + 786, + 4880, + 6718, + 13614 + ] + }, + { + "test_f1": 0.5266152905725795, + "test_acc": 0.529559748427673, + "test_auc": 0.5373967509131286, + "val_auc": 0.6436574629851941, + "k": 16, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928, + 1703, + 786, + 4880, + 6718, + 13614, + 1596, + 1417, + 448, + 1643, + 334, + 1202 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c93771ba38c7505f031f051aa11744b54c92a85e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9919493798264469, + "test_acc": 0.9919496855345912, + "test_auc": 0.9968998692709856, + "val_auc": 1.0, + "k": 1, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.9939622221174144, + "test_acc": 0.9939622641509434, + "test_auc": 0.9986835962934122, + "val_auc": 1.0, + "k": 2, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988 + ] + }, + { + "test_f1": 0.991194711092179, + "test_acc": 0.9911949685534591, + "test_auc": 0.999299015026242, + "val_auc": 1.0, + "k": 5, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 152, + 1744, + 487 + ] + }, + { + "test_f1": 0.9904398885439057, + "test_acc": 0.9904402515723271, + "test_auc": 0.9990244689070508, + "val_auc": 1.0, + "k": 10, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 152, + 1744, + 487, + 6935, + 292, + 3417, + 559, + 347 + ] + }, + { + "test_f1": 0.9932073098646004, + "test_acc": 0.9932075471698113, + "test_auc": 0.9992509156600397, + "val_auc": 1.0, + "k": 16, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 152, + 1744, + 487, + 6935, + 292, + 3417, + 559, + 347, + 821, + 6475, + 1960, + 771, + 7368, + 1132 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d9153e6ad69b8ca836d2a2e5ac14ffc110109eb5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8552720201832801, + "test_acc": 0.8555974842767295, + "test_auc": 0.8873597650320646, + "val_auc": 0.9115646258503401, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487 + ] + }, + { + "test_f1": 0.8950175587748503, + "test_acc": 0.8950943396226415, + "test_auc": 0.9394623908524118, + "val_auc": 0.9711884753901561, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744 + ] + }, + { + "test_f1": 0.8984228339369482, + "test_acc": 0.8986163522012579, + "test_auc": 0.9402224874156869, + "val_auc": 0.9693877551020408, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 771, + 559 + ] + }, + { + "test_f1": 0.9285021361963331, + "test_acc": 0.9285534591194968, + "test_auc": 0.9761777762701017, + "val_auc": 0.9739895958383353, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 771, + 559, + 7368, + 24947, + 1524, + 1028, + 152 + ] + }, + { + "test_f1": 0.9421274866049234, + "test_acc": 0.9421383647798742, + "test_auc": 0.985335262708612, + "val_auc": 0.9867947178871549, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 771, + 559, + 7368, + 24947, + 1524, + 1028, + 152, + 878, + 6475, + 6935, + 1288, + 6065, + 7517 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..280e2083d6a51fa1a074b21bce1074aba1af136c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8616031455028565, + "test_acc": 0.8627848101265823, + "test_auc": 0.8717030916079068, + "val_auc": 0.8499399759903962, + "k": 1, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104 + ] + }, + { + "test_f1": 0.8643171286634241, + "test_acc": 0.8653164556962025, + "test_auc": 0.8968790634524118, + "val_auc": 0.8979591836734694, + "k": 2, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292 + ] + }, + { + "test_f1": 0.8595603875028988, + "test_acc": 0.8607594936708861, + "test_auc": 0.9333224632776707, + "val_auc": 0.9155662264905962, + "k": 5, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 4555, + 1708, + 337 + ] + }, + { + "test_f1": 0.8677807847716442, + "test_acc": 0.8688607594936709, + "test_auc": 0.93915486342698, + "val_auc": 0.9407763105242097, + "k": 10, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 4555, + 1708, + 337, + 1815, + 918, + 1851, + 968, + 1627 + ] + }, + { + "test_f1": 0.8766841596882038, + "test_acc": 0.8774683544303797, + "test_auc": 0.9445760473196083, + "val_auc": 0.9503801520608244, + "k": 16, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 4555, + 1708, + 337, + 1815, + 918, + 1851, + 968, + 1627, + 1745, + 4649, + 6747, + 677, + 1380, + 1810 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7b3f3cb35057ead3839e826e3611b50609f8a9cd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7493676143480152, + "test_acc": 0.7548717948717949, + "test_auc": 0.7731342781162689, + "val_auc": 0.7839657282741739, + "k": 1, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468 + ] + }, + { + "test_f1": 0.8345030835466858, + "test_acc": 0.8358974358974359, + "test_auc": 0.866414060659104, + "val_auc": 0.8449612403100775, + "k": 2, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351 + ] + }, + { + "test_f1": 0.8585705014393235, + "test_acc": 0.8594871794871795, + "test_auc": 0.8993713602854547, + "val_auc": 0.901672786617707, + "k": 5, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 1132, + 1474 + ] + }, + { + "test_f1": 0.8599914799190095, + "test_acc": 0.8605128205128205, + "test_auc": 0.9105114114518464, + "val_auc": 0.8961648306813546, + "k": 10, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 1132, + 1474, + 3417, + 26988, + 343, + 1305, + 805 + ] + }, + { + "test_f1": 0.8682272752178911, + "test_acc": 0.8687179487179487, + "test_auc": 0.9132064833204296, + "val_auc": 0.8998368013055895, + "k": 16, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 1132, + 1474, + 3417, + 26988, + 343, + 1305, + 805, + 930, + 27324, + 1473, + 21810, + 1551, + 52 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b43ad1e0ea2f31b98c354f1a3bc1397162464260 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8739831506262811, + "test_acc": 0.8749685534591195, + "test_auc": 0.8846612640108391, + "val_auc": 0.9103641456582633, + "k": 1, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.8909147603173027, + "test_acc": 0.8913207547169811, + "test_auc": 0.9415021837112256, + "val_auc": 0.9653861544617846, + "k": 2, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1447 + ] + }, + { + "test_f1": 0.8904637008134998, + "test_acc": 0.8905660377358491, + "test_auc": 0.9595208391769845, + "val_auc": 0.9739895958383353, + "k": 5, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1447, + 1279, + 8129, + 11381 + ] + }, + { + "test_f1": 0.907668538256634, + "test_acc": 0.9076729559748428, + "test_auc": 0.9724833652139309, + "val_auc": 0.9783913565426171, + "k": 10, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1447, + 1279, + 8129, + 11381, + 238, + 1556, + 27068, + 7740, + 2860 + ] + }, + { + "test_f1": 0.9202392048057211, + "test_acc": 0.920251572327044, + "test_auc": 0.9764389051976683, + "val_auc": 0.9823929571828732, + "k": 16, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1447, + 1279, + 8129, + 11381, + 238, + 1556, + 27068, + 7740, + 2860, + 6828, + 1478, + 88, + 3301, + 7119, + 4877 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6631c701126ab5da4695722ba5807c6231a30450 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8269692735038598, + "test_acc": 0.8307692307692308, + "test_auc": 0.8298317736560407, + "val_auc": 0.8488372093023255, + "k": 1, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628 + ] + }, + { + "test_f1": 0.8959549789350462, + "test_acc": 0.8964102564102564, + "test_auc": 0.90775322314606, + "val_auc": 0.9135046919624644, + "k": 2, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416 + ] + }, + { + "test_f1": 0.8929866328971435, + "test_acc": 0.8933333333333333, + "test_auc": 0.9122092436126165, + "val_auc": 0.9175846593227255, + "k": 5, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 9381, + 5964, + 4211 + ] + }, + { + "test_f1": 0.8914534418063254, + "test_acc": 0.8923076923076924, + "test_auc": 0.9087210085165113, + "val_auc": 0.9202366381068952, + "k": 10, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 9381, + 5964, + 4211, + 238, + 805, + 930, + 737, + 334 + ] + }, + { + "test_f1": 0.8914534418063254, + "test_acc": 0.8923076923076924, + "test_auc": 0.9087252162789915, + "val_auc": 0.9202366381068952, + "k": 16, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 9381, + 5964, + 4211, + 238, + 805, + 930, + 737, + 334, + 1132, + 1524, + 2062, + 1158, + 1274, + 878 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8b6a3767903cd2a2602630347ebbfd109210fc2e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8686541165468195, + "test_acc": 0.8687179487179487, + "test_auc": 0.917473154475376, + "val_auc": 0.9392084863321093, + "k": 1, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524 + ] + }, + { + "test_f1": 0.8696613267801492, + "test_acc": 0.8697435897435898, + "test_auc": 0.917473154475376, + "val_auc": 0.9392084863321093, + "k": 2, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947 + ] + }, + { + "test_f1": 0.8758527731254057, + "test_acc": 0.8758974358974358, + "test_auc": 0.9290024236711887, + "val_auc": 0.9394124847001224, + "k": 5, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 888 + ] + }, + { + "test_f1": 0.8717382050539987, + "test_acc": 0.8717948717948718, + "test_auc": 0.9336141313495137, + "val_auc": 0.9436964504283967, + "k": 10, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 888, + 969, + 1474, + 7309, + 1214, + 1682 + ] + }, + { + "test_f1": 0.8768995092881732, + "test_acc": 0.8769230769230769, + "test_auc": 0.9344472683205979, + "val_auc": 0.9457364341085271, + "k": 16, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 888, + 969, + 1474, + 7309, + 1214, + 1682, + 776, + 1260, + 3203, + 1839, + 5173, + 848 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ba556a4791cd02f82a1389f83cfec4a3b70ac79d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8185855694692904, + "test_acc": 0.8205128205128205, + "test_auc": 0.8422594001413809, + "val_auc": 0.8123215014279885, + "k": 1, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.8103725699573614, + "test_acc": 0.8123076923076923, + "test_auc": 0.8719241256269566, + "val_auc": 0.8212974296205631, + "k": 2, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815 + ] + }, + { + "test_f1": 0.8237957965603699, + "test_acc": 0.8246153846153846, + "test_auc": 0.9017129801056989, + "val_auc": 0.8649530803753569, + "k": 5, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 10639, + 334, + 1524 + ] + }, + { + "test_f1": 0.880916660781388, + "test_acc": 0.8810256410256411, + "test_auc": 0.9459218366041675, + "val_auc": 0.8918808649530804, + "k": 10, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 10639, + 334, + 1524, + 6935, + 487, + 930, + 1866, + 1744 + ] + }, + { + "test_f1": 0.8840693937207817, + "test_acc": 0.884102564102564, + "test_auc": 0.9506176995320967, + "val_auc": 0.923296613627091, + "k": 16, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 10639, + 334, + 1524, + 6935, + 487, + 930, + 1866, + 1744, + 1221, + 508, + 821, + 2046, + 1158, + 913 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..29dd1ebac2a069d3d7634b0be27784a7b08c4c2d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8307803011803011, + "test_acc": 0.8317948717948718, + "test_auc": 0.87085956171946, + "val_auc": 0.8845369237046105, + "k": 1, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52 + ] + }, + { + "test_f1": 0.8296652024634188, + "test_acc": 0.8307692307692308, + "test_auc": 0.87085956171946, + "val_auc": 0.8845369237046105, + "k": 2, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 2131 + ] + }, + { + "test_f1": 0.8235756623112945, + "test_acc": 0.8256410256410256, + "test_auc": 0.87085956171946, + "val_auc": 0.8845369237046105, + "k": 5, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 2131, + 19230, + 1274, + 2454 + ] + }, + { + "test_f1": 0.8360660705970389, + "test_acc": 0.8369230769230769, + "test_auc": 0.8950794425556265, + "val_auc": 0.9102407180742554, + "k": 10, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 2131, + 19230, + 1274, + 2454, + 2484, + 17546, + 2046, + 1193, + 18545 + ] + }, + { + "test_f1": 0.8539826703286814, + "test_acc": 0.8543589743589743, + "test_auc": 0.9057208738681118, + "val_auc": 0.9481844145246837, + "k": 16, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 2131, + 19230, + 1274, + 2454, + 2484, + 17546, + 2046, + 1193, + 18545, + 238, + 821, + 1682, + 888, + 930, + 117 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..edb2575b1693f3ca17d1c9940fed65cdcc0f9f1f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8714510197981092, + "test_acc": 0.8716577540106952, + "test_auc": 0.9226285254122513, + "val_auc": 0.9252525252525253, + "k": 1, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839 + ] + }, + { + "test_f1": 0.8789982290679017, + "test_acc": 0.879144385026738, + "test_auc": 0.9368674390087666, + "val_auc": 0.932121212121212, + "k": 2, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682 + ] + }, + { + "test_f1": 0.8779978341687785, + "test_acc": 0.8780748663101604, + "test_auc": 0.9417311810245428, + "val_auc": 0.938989898989899, + "k": 5, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 3417 + ] + }, + { + "test_f1": 0.8951317475834774, + "test_acc": 0.895187165775401, + "test_auc": 0.9547209868409012, + "val_auc": 0.9551515151515151, + "k": 10, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 3417, + 343, + 930, + 340, + 1158, + 1895 + ] + }, + { + "test_f1": 0.8887395097835264, + "test_acc": 0.8887700534759359, + "test_auc": 0.9541879426783068, + "val_auc": 0.9535353535353535, + "k": 16, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 3417, + 343, + 930, + 340, + 1158, + 1895, + 888, + 874, + 238, + 1441, + 456, + 334 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..864c154f86e5d988f69ad326ebad28fd8765b586 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9650537333022758, + "test_acc": 0.9650574712643678, + "test_auc": 0.9742211598300773, + "val_auc": 0.9886363636363636, + "k": 1, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14949 + ] + }, + { + "test_f1": 0.9655138733963089, + "test_acc": 0.9655172413793104, + "test_auc": 0.9822264462362683, + "val_auc": 0.9991883116883117, + "k": 2, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14949, + 7554 + ] + }, + { + "test_f1": 0.9783906127429012, + "test_acc": 0.9783908045977011, + "test_auc": 0.9961594918556199, + "val_auc": 1.0, + "k": 5, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14949, + 7554, + 1072, + 4323, + 760 + ] + }, + { + "test_f1": 0.9788505210635835, + "test_acc": 0.9788505747126437, + "test_auc": 0.9953908828940959, + "val_auc": 1.0, + "k": 10, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14949, + 7554, + 1072, + 4323, + 760, + 192, + 1065, + 763, + 1101, + 1682 + ] + }, + { + "test_f1": 0.9793091551558699, + "test_acc": 0.9793103448275862, + "test_auc": 0.9947787014989988, + "val_auc": 1.0, + "k": 16, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14949, + 7554, + 1072, + 4323, + 760, + 192, + 1065, + 763, + 1101, + 1682, + 1622, + 7847, + 1876, + 1684, + 18979, + 652 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f1602fa4b379c33e26f61c96626bda72f3402132 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9708536348944701, + "test_acc": 0.9708552317247969, + "test_auc": 0.9853665485106313, + "val_auc": 1.0, + "k": 1, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286 + ] + }, + { + "test_f1": 0.9636797299911364, + "test_acc": 0.9636884854276159, + "test_auc": 0.9853665485106313, + "val_auc": 1.0, + "k": 2, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905 + ] + }, + { + "test_f1": 0.9641581520516956, + "test_acc": 0.9641662685140946, + "test_auc": 0.9853665485106313, + "val_auc": 1.0, + "k": 5, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1874, + 581, + 2874 + ] + }, + { + "test_f1": 0.988054681140885, + "test_acc": 0.9880554228380315, + "test_auc": 0.999443917886121, + "val_auc": 1.0, + "k": 10, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1874, + 581, + 2874, + 1233, + 9355, + 304, + 1682, + 1684 + ] + }, + { + "test_f1": 0.9923551530017528, + "test_acc": 0.9923554706163402, + "test_auc": 0.9992804717475587, + "val_auc": 1.0, + "k": 16, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1874, + 581, + 2874, + 1233, + 9355, + 304, + 1682, + 1684, + 1079, + 3877, + 37, + 88, + 239, + 1156 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ffff037ba010aaa33b1a53f02bc8bb5d08c63c9e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9944827096363509, + "test_acc": 0.9944827586206897, + "test_auc": 0.9989667325071703, + "val_auc": 1.0, + "k": 1, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559 + ] + }, + { + "test_f1": 0.9949424966626412, + "test_acc": 0.9949425287356322, + "test_auc": 0.9998545646409437, + "val_auc": 1.0, + "k": 2, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2236 + ] + }, + { + "test_f1": 0.9935632183908046, + "test_acc": 0.9935632183908046, + "test_auc": 0.9997497158937172, + "val_auc": 1.0, + "k": 5, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2236, + 1789, + 300, + 1934 + ] + }, + { + "test_f1": 0.9931034453601555, + "test_acc": 0.993103448275862, + "test_auc": 0.9998224335732453, + "val_auc": 1.0, + "k": 10, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2236, + 1789, + 300, + 1934, + 334, + 3825, + 1592, + 982, + 119 + ] + }, + { + "test_f1": 0.9926436470599098, + "test_acc": 0.9926436781609196, + "test_auc": 0.9989312192218193, + "val_auc": 1.0, + "k": 16, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2236, + 1789, + 300, + 1934, + 334, + 3825, + 1592, + 982, + 119, + 1158, + 1872, + 1268, + 913, + 6292, + 1769 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fe5cba3f72ea2316b1b035633cce015954734b1c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9558539502015315, + "test_acc": 0.9558620689655173, + "test_auc": 0.9830466340169922, + "val_auc": 0.9598214285714286, + "k": 1, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095 + ] + }, + { + "test_f1": 0.9687336097278813, + "test_acc": 0.968735632183908, + "test_auc": 0.9903699807890038, + "val_auc": 0.976055194805195, + "k": 2, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023 + ] + }, + { + "test_f1": 0.9701122890134933, + "test_acc": 0.9701149425287356, + "test_auc": 0.9904418529141187, + "val_auc": 0.9772727272727273, + "k": 5, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 494, + 1789 + ] + }, + { + "test_f1": 0.9802292330861537, + "test_acc": 0.9802298850574712, + "test_auc": 0.9959819254288652, + "val_auc": 0.984172077922078, + "k": 10, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 494, + 1789, + 1301, + 1682, + 211, + 1079, + 235 + ] + }, + { + "test_f1": 0.977928459044762, + "test_acc": 0.9779310344827586, + "test_auc": 0.9962660317116727, + "val_auc": 0.9898538961038961, + "k": 16, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 494, + 1789, + 1301, + 1682, + 211, + 1079, + 235, + 2011, + 1684, + 1193, + 292, + 902, + 1701 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5122bae049cba6661006af2697716ea9d7f29900 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9659723360307163, + "test_acc": 0.9659770114942529, + "test_auc": 0.993479084366037, + "val_auc": 0.9874188311688311, + "k": 1, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32 + ] + }, + { + "test_f1": 0.9848272333992235, + "test_acc": 0.9848275862068966, + "test_auc": 0.9981710658044267, + "val_auc": 0.994724025974026, + "k": 2, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683 + ] + }, + { + "test_f1": 0.9931033870456067, + "test_acc": 0.993103448275862, + "test_auc": 0.9997979124952648, + "val_auc": 0.9995941558441559, + "k": 5, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 816, + 1558 + ] + }, + { + "test_f1": 0.9967816010314028, + "test_acc": 0.9967816091954023, + "test_auc": 0.9999399656366685, + "val_auc": 1.0, + "k": 10, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 816, + 1558, + 1047, + 5094, + 232, + 1458, + 495 + ] + }, + { + "test_f1": 0.9958620689655172, + "test_acc": 0.9958620689655172, + "test_auc": 0.9999036067969046, + "val_auc": 1.0, + "k": 16, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 816, + 1558, + 1047, + 5094, + 232, + 1458, + 495, + 1023, + 868, + 1051, + 87, + 1682, + 769 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..19e2f1902cc74f2c511b12c75ae866dc784fd020 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8212138516148825, + "test_acc": 0.8213836477987422, + "test_auc": 0.868422158517284, + "val_auc": 0.9027611044417768, + "k": 1, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.891556843573966, + "test_acc": 0.8915723270440251, + "test_auc": 0.9470917097957648, + "val_auc": 0.9839935974389756, + "k": 2, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028 + ] + }, + { + "test_f1": 0.9144556311189818, + "test_acc": 0.9144654088050315, + "test_auc": 0.9647337978550721, + "val_auc": 0.9939975990396159, + "k": 5, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 1305, + 1839, + 778 + ] + }, + { + "test_f1": 0.9098839972868858, + "test_acc": 0.909937106918239, + "test_auc": 0.9624862916806323, + "val_auc": 0.993797519007603, + "k": 10, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 1305, + 1839, + 778, + 3417, + 5539, + 1478, + 654, + 4663 + ] + }, + { + "test_f1": 0.9101368105839569, + "test_acc": 0.910188679245283, + "test_auc": 0.9624928736991654, + "val_auc": 0.993797519007603, + "k": 16, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 1305, + 1839, + 778, + 3417, + 5539, + 1478, + 654, + 4663, + 1895, + 6065, + 1279, + 22848, + 1815, + 343 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6045632e91cc7599a223fcec309402a88ce6a8a6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9990804593813543, + "test_acc": 0.999080459770115, + "test_auc": 0.9999839344661507, + "val_auc": 1.0, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686 + ] + }, + { + "test_f1": 0.9977011465095805, + "test_acc": 0.9977011494252873, + "test_auc": 0.9999813978029114, + "val_auc": 1.0, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9991028667676822, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379, + 682, + 4787, + 23428 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9991967233075383, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379, + 682, + 4787, + 23428, + 1840, + 24163, + 1162, + 84, + 1931 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9996127360787921, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379, + 682, + 4787, + 23428, + 1840, + 24163, + 1162, + 84, + 1931, + 990, + 104, + 613, + 452, + 3818, + 1584 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..492b7ffde931dad1d847613540a40af5461a636c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9365517241379311, + "test_acc": 0.9365517241379311, + "test_auc": 0.9818129701282536, + "val_auc": 0.9866071428571428, + "k": 1, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586 + ] + }, + { + "test_f1": 0.9843670229826741, + "test_acc": 0.984367816091954, + "test_auc": 0.9951118499377671, + "val_auc": 1.0, + "k": 2, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 2211 + ] + }, + { + "test_f1": 0.9898849377317693, + "test_acc": 0.9898850574712643, + "test_auc": 0.997116659451269, + "val_auc": 1.0, + "k": 5, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 2211, + 10, + 21724, + 1934 + ] + }, + { + "test_f1": 0.9880456434463953, + "test_acc": 0.9880459770114942, + "test_auc": 0.9957967490123925, + "val_auc": 1.0, + "k": 10, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 2211, + 10, + 21724, + 1934, + 620, + 878, + 3223, + 1603, + 559 + ] + }, + { + "test_f1": 0.9885052368570504, + "test_acc": 0.9885057471264368, + "test_auc": 0.9964317603766437, + "val_auc": 1.0, + "k": 16, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 2211, + 10, + 21724, + 1934, + 620, + 878, + 3223, + 1603, + 559, + 610, + 495, + 1874, + 1682, + 760, + 1079 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..156203d8a171ba614eeaff4b8bdaba2b04adcc9a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8523277621426503, + "test_acc": 0.8524137931034482, + "test_auc": 0.9230714594945614, + "val_auc": 0.9159902597402598, + "k": 1, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1682 + ] + }, + { + "test_f1": 0.8896534929393846, + "test_acc": 0.8896551724137931, + "test_auc": 0.9556214994047297, + "val_auc": 0.974025974025974, + "k": 2, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1682, + 495 + ] + }, + { + "test_f1": 0.924137834816161, + "test_acc": 0.9241379310344827, + "test_auc": 0.97759449916121, + "val_auc": 0.994724025974026, + "k": 5, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1682, + 495, + 6979, + 1027, + 1708 + ] + }, + { + "test_f1": 0.94343564929575, + "test_acc": 0.9434482758620689, + "test_auc": 0.9876912644082473, + "val_auc": 0.992288961038961, + "k": 10, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1682, + 495, + 6979, + 1027, + 1708, + 1707, + 1631, + 713, + 425, + 1079 + ] + }, + { + "test_f1": 0.9618315769251731, + "test_acc": 0.9618390804597701, + "test_auc": 0.993952594837383, + "val_auc": 0.9979707792207793, + "k": 16, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1682, + 495, + 6979, + 1027, + 1708, + 1707, + 1631, + 713, + 425, + 1079, + 27304, + 2024, + 1841, + 1840, + 2616, + 117 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1d798ef0e65bd823acb61d202e725062f4eaceef --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9535478999310141, + "test_acc": 0.9535632183908046, + "test_auc": 0.9734719986200552, + "val_auc": 0.9772727272727273, + "k": 1, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734 + ] + }, + { + "test_f1": 0.9544645466763876, + "test_acc": 0.9544827586206897, + "test_auc": 0.9820640997889497, + "val_auc": 0.9959415584415585, + "k": 2, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734, + 1961 + ] + }, + { + "test_f1": 0.95492983825076, + "test_acc": 0.9549425287356322, + "test_auc": 0.9854412441149414, + "val_auc": 0.9967532467532468, + "k": 5, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734, + 1961, + 2689, + 7182, + 248 + ] + }, + { + "test_f1": 0.9622914373485374, + "test_acc": 0.9622988505747126, + "test_auc": 0.9870879613344878, + "val_auc": 0.9979707792207793, + "k": 10, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734, + 1961, + 2689, + 7182, + 248, + 732, + 1264, + 859, + 1536, + 653 + ] + }, + { + "test_f1": 0.968273809855754, + "test_acc": 0.9682758620689655, + "test_auc": 0.9915850424806537, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734, + 1961, + 2689, + 7182, + 248, + 732, + 1264, + 859, + 1536, + 653, + 1441, + 1482, + 2030, + 306, + 1682, + 1960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..60736851102ffcd754788fbf3ea0cbe5dab1cfd2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9535437743487909, + "test_acc": 0.9535632183908046, + "test_auc": 0.9881512460089832, + "val_auc": 0.9959415584415584, + "k": 1, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786 + ] + }, + { + "test_f1": 0.9553630466981429, + "test_acc": 0.9554022988505747, + "test_auc": 0.9881512460089832, + "val_auc": 0.9959415584415584, + "k": 2, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558 + ] + }, + { + "test_f1": 0.9609160702484142, + "test_acc": 0.960919540229885, + "test_auc": 0.9928694396341794, + "val_auc": 0.9959415584415584, + "k": 5, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 271 + ] + }, + { + "test_f1": 0.9553630466981429, + "test_acc": 0.9554022988505747, + "test_auc": 0.9881512460089832, + "val_auc": 0.9959415584415584, + "k": 10, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 271, + 653, + 292, + 1910, + 276, + 1051 + ] + }, + { + "test_f1": 0.9641261011344454, + "test_acc": 0.9641379310344828, + "test_auc": 0.994678926078251, + "val_auc": 0.9971590909090909, + "k": 16, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 271, + 653, + 292, + 1910, + 276, + 1051, + 760, + 67, + 1627, + 354, + 327, + 460 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..84dcf151334b6ecba81dd67b9bf9434db102faf8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9273549397715987, + "test_acc": 0.9273563218390805, + "test_auc": 0.9619314492126197, + "val_auc": 0.9446022727272727, + "k": 1, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856 + ] + }, + { + "test_f1": 0.9650571610302217, + "test_acc": 0.9650574712643678, + "test_auc": 0.9884226689755939, + "val_auc": 0.987012987012987, + "k": 2, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856, + 11831 + ] + }, + { + "test_f1": 0.9714934574490965, + "test_acc": 0.9714942528735632, + "test_auc": 0.9913854916391581, + "val_auc": 0.9878246753246753, + "k": 5, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856, + 11831, + 7739, + 1168, + 1167 + ] + }, + { + "test_f1": 0.9724080071708235, + "test_acc": 0.9724137931034482, + "test_auc": 0.9946383394664213, + "val_auc": 0.9955357142857143, + "k": 10, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856, + 11831, + 7739, + 1168, + 1167, + 1037, + 822, + 517, + 1252, + 1536 + ] + }, + { + "test_f1": 0.9710308085238832, + "test_acc": 0.9710344827586207, + "test_auc": 0.9948674847123762, + "val_auc": 0.997564935064935, + "k": 16, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856, + 11831, + 7739, + 1168, + 1167, + 1037, + 822, + 517, + 1252, + 1536, + 586, + 581, + 18934, + 1449, + 1366, + 1101 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c2e0b6aa465b1dd92f9e456cb101ccbedc045f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9627535028973594, + "test_acc": 0.9627586206896551, + "test_auc": 0.9740660005952704, + "val_auc": 0.9987824675324676, + "k": 1, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840 + ] + }, + { + "test_f1": 0.962755628924118, + "test_acc": 0.9627586206896551, + "test_auc": 0.9919528586503599, + "val_auc": 0.9991883116883117, + "k": 2, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 1765 + ] + }, + { + "test_f1": 0.9742523837165158, + "test_acc": 0.9742528735632184, + "test_auc": 0.9964698103252341, + "val_auc": 0.9995941558441559, + "k": 5, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 1765, + 5495, + 789, + 1763 + ] + }, + { + "test_f1": 0.9779304186702684, + "test_acc": 0.9779310344827586, + "test_auc": 0.9968477731478976, + "val_auc": 0.9995941558441559, + "k": 10, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 1765, + 5495, + 789, + 1763, + 1682, + 901, + 1797, + 434, + 27505 + ] + }, + { + "test_f1": 0.9866666497555677, + "test_acc": 0.9866666666666667, + "test_auc": 0.9975208344607391, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 1765, + 5495, + 789, + 1763, + 1682, + 901, + 1797, + 434, + 27505, + 1815, + 1310, + 10075, + 1801, + 677, + 29148 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e1885aec5f08a1182d2fb240591cf3c5443e5d75 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9586083876545313, + "test_acc": 0.9586206896551724, + "test_auc": 0.9734339486714649, + "val_auc": 0.9691558441558441, + "k": 1, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879 + ] + }, + { + "test_f1": 0.9714924088695389, + "test_acc": 0.9714942528735632, + "test_auc": 0.9874857946858596, + "val_auc": 0.9825487012987012, + "k": 2, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293 + ] + }, + { + "test_f1": 0.9894252202946138, + "test_acc": 0.9894252873563218, + "test_auc": 0.9962478522917908, + "val_auc": 1.0, + "k": 5, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 277, + 4455, + 30 + ] + }, + { + "test_f1": 0.9848270024587186, + "test_acc": 0.9848275862068966, + "test_auc": 0.9959438754802747, + "val_auc": 1.0, + "k": 10, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 277, + 4455, + 30, + 990, + 7313, + 1501, + 30210, + 7367 + ] + }, + { + "test_f1": 0.9885056110583972, + "test_acc": 0.9885057471264368, + "test_auc": 0.9934850032469289, + "val_auc": 1.0, + "k": 16, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 277, + 4455, + 30, + 990, + 7313, + 1501, + 30210, + 7367, + 715, + 1454, + 1484, + 419, + 948, + 613 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1c843041cafcab99afd04a3abb280a313b007140 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9802290073941264, + "test_acc": 0.9802298850574712, + "test_auc": 0.9903391180529248, + "val_auc": 1.0, + "k": 1, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439 + ] + }, + { + "test_f1": 0.9820676691063379, + "test_acc": 0.9820689655172414, + "test_auc": 0.9897594905027327, + "val_auc": 1.0, + "k": 2, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 822 + ] + }, + { + "test_f1": 0.9788495016824414, + "test_acc": 0.9788505747126437, + "test_auc": 0.9926631243573787, + "val_auc": 1.0, + "k": 5, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 822, + 5907, + 1795, + 26232 + ] + }, + { + "test_f1": 0.9770112221176575, + "test_acc": 0.9770114942528736, + "test_auc": 0.9941572190053574, + "val_auc": 1.0, + "k": 10, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 822, + 5907, + 1795, + 26232, + 5397, + 5535, + 1121, + 1790, + 802 + ] + }, + { + "test_f1": 0.9839079779436821, + "test_acc": 0.9839080459770115, + "test_auc": 0.9969483941230586, + "val_auc": 1.0, + "k": 16, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 822, + 5907, + 1795, + 26232, + 5397, + 5535, + 1121, + 1790, + 802, + 3676, + 5797, + 493, + 1441, + 1631, + 2020 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c8e50c1aaa57f3607aa06002c7648ae8c2877bf7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9590625464176658, + "test_acc": 0.9590804597701149, + "test_auc": 0.9776376224362789, + "val_auc": 0.9859983766233766, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272 + ] + }, + { + "test_f1": 0.9585983722470165, + "test_acc": 0.9586206896551724, + "test_auc": 0.9776376224362789, + "val_auc": 0.9859983766233766, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866 + ] + }, + { + "test_f1": 0.9622914373485374, + "test_acc": 0.9622988505747126, + "test_auc": 0.99350698766167, + "val_auc": 0.9963474025974026, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 1072, + 5703 + ] + }, + { + "test_f1": 0.9820689427747288, + "test_acc": 0.9820689655172414, + "test_auc": 0.9966110179122247, + "val_auc": 0.9979707792207793, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 1072, + 5703, + 165, + 918, + 1682, + 1787, + 1715 + ] + }, + { + "test_f1": 0.9880459618498032, + "test_acc": 0.9880459770114942, + "test_auc": 0.9975749499431787, + "val_auc": 1.0, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 1072, + 5703, + 165, + 918, + 1682, + 1787, + 1715, + 1726, + 508, + 1079, + 22196, + 1934, + 1023 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dc582856178db288b2e988e0e15374dcd6493444 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5680217127454736, + "test_acc": 0.5854088050314465, + "test_auc": 0.5925563446101876, + "val_auc": 0.6008403361344538, + "k": 1, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.5846277259606169, + "test_acc": 0.5866666666666667, + "test_auc": 0.6258862941109161, + "val_auc": 0.6496598639455782, + "k": 2, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556 + ] + }, + { + "test_f1": 0.5864262697958103, + "test_acc": 0.5891823899371069, + "test_auc": 0.642142487537201, + "val_auc": 0.6942777110844338, + "k": 5, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556, + 1279, + 1132, + 17275 + ] + }, + { + "test_f1": 0.6065077088432153, + "test_acc": 0.6065408805031447, + "test_auc": 0.6575308671353739, + "val_auc": 0.6798719487795118, + "k": 10, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556, + 1279, + 1132, + 17275, + 6475, + 1815, + 238, + 1158, + 581 + ] + }, + { + "test_f1": 0.6168550549272257, + "test_acc": 0.6168553459119497, + "test_auc": 0.6643774321824252, + "val_auc": 0.6870748299319728, + "k": 16, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556, + 1279, + 1132, + 17275, + 6475, + 1815, + 238, + 1158, + 581, + 1028, + 1895, + 7740, + 925, + 1101, + 1148 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2008dc1656c19f2c4a3a754522e771185f0ac09a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9701086743515656, + "test_acc": 0.9701149425287356, + "test_auc": 0.9799489454245359, + "val_auc": 0.9768668831168832, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963 + ] + }, + { + "test_f1": 0.9682674100425168, + "test_acc": 0.9682758620689655, + "test_auc": 0.9809116091238703, + "val_auc": 0.9837662337662338, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217 + ] + }, + { + "test_f1": 0.974711360698859, + "test_acc": 0.9747126436781609, + "test_auc": 0.9923519603333515, + "val_auc": 0.9910714285714286, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217, + 3772, + 1942, + 352 + ] + }, + { + "test_f1": 0.9793100999051152, + "test_acc": 0.9793103448275862, + "test_auc": 0.9941842767465772, + "val_auc": 0.9967532467532467, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217, + 3772, + 1942, + 352, + 1682, + 3243, + 1047, + 111, + 1898 + ] + }, + { + "test_f1": 0.978390201611706, + "test_acc": 0.9783908045977011, + "test_auc": 0.9942722144055414, + "val_auc": 0.9971590909090909, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217, + 3772, + 1942, + 352, + 1682, + 3243, + 1047, + 111, + 1898, + 285, + 581, + 1261, + 1726, + 760, + 1966 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8274b16b25af3bbae8aafeff91806d898466d23f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9825279600179616, + "test_acc": 0.9825287356321839, + "test_auc": 0.9960516836679474, + "val_auc": 1.0, + "k": 1, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833 + ] + }, + { + "test_f1": 0.9806867891766143, + "test_acc": 0.9806896551724138, + "test_auc": 0.9960516836679474, + "val_auc": 1.0, + "k": 2, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675 + ] + }, + { + "test_f1": 0.9871263224848184, + "test_acc": 0.9871264367816092, + "test_auc": 0.9989481303100818, + "val_auc": 1.0, + "k": 5, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682 + ] + }, + { + "test_f1": 0.9898850574712643, + "test_acc": 0.9898850574712643, + "test_auc": 0.9993015720547648, + "val_auc": 1.0, + "k": 10, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682, + 1588, + 913, + 491, + 1278, + 361 + ] + }, + { + "test_f1": 0.9889654705898647, + "test_acc": 0.9889655172413793, + "test_auc": 0.9981938957735808, + "val_auc": 1.0, + "k": 16, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682, + 1588, + 913, + 491, + 1278, + 361, + 859, + 1181, + 829, + 113, + 1515, + 1079 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b41e525b5e8c693890602f4a3802671e1627fb6b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8592844662673813, + "test_acc": 0.8593103448275862, + "test_auc": 0.9119566467341307, + "val_auc": 0.9468344155844155, + "k": 1, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020 + ] + }, + { + "test_f1": 0.9167808703503211, + "test_acc": 0.9167816091954023, + "test_auc": 0.9674968883597597, + "val_auc": 0.9866071428571428, + "k": 2, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760 + ] + }, + { + "test_f1": 0.9443401430772516, + "test_acc": 0.944367816091954, + "test_auc": 0.9857363426051193, + "val_auc": 0.9853896103896104, + "k": 5, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 913 + ] + }, + { + "test_f1": 0.9503249603761283, + "test_acc": 0.9503448275862069, + "test_auc": 0.9907648546999297, + "val_auc": 0.9914772727272727, + "k": 10, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 913, + 49, + 18571, + 170, + 1960, + 2401 + ] + }, + { + "test_f1": 0.962282346455246, + "test_acc": 0.9622988505747126, + "test_auc": 0.9946002895178311, + "val_auc": 0.9967532467532467, + "k": 16, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 913, + 49, + 18571, + 170, + 1960, + 2401, + 2817, + 2030, + 364, + 2023, + 901, + 2084 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ad15eb949b29b7e7f629a01758a19cb5d7b09b1a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9484942496072225, + "test_acc": 0.9485057471264368, + "test_auc": 0.9566306685967856, + "val_auc": 0.9545454545454547, + "k": 1, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123 + ] + }, + { + "test_f1": 0.9457463006951974, + "test_acc": 0.9457471264367816, + "test_auc": 0.9755740468910654, + "val_auc": 0.9604301948051949, + "k": 2, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1763 + ] + }, + { + "test_f1": 0.9696522086156946, + "test_acc": 0.9696551724137931, + "test_auc": 0.9809116091238703, + "val_auc": 0.9752435064935066, + "k": 5, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1763, + 1491, + 1682, + 1111 + ] + }, + { + "test_f1": 0.9664331912323958, + "test_acc": 0.9664367816091954, + "test_auc": 0.9913009361978462, + "val_auc": 0.984172077922078, + "k": 10, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1763, + 1491, + 1682, + 1111, + 167, + 866, + 679, + 204, + 523 + ] + }, + { + "test_f1": 0.9719504653961409, + "test_acc": 0.9719540229885058, + "test_auc": 0.9940117836463012, + "val_auc": 0.9890422077922079, + "k": 16, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1763, + 1491, + 1682, + 1111, + 167, + 866, + 679, + 204, + 523, + 144, + 1310, + 1486, + 7316, + 2014, + 687 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cf634a10ac994f62ba012c3e871d51e0b8a361ad --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.839775041151009, + "test_acc": 0.84, + "test_auc": 0.8898031211104498, + "val_auc": 0.8892045454545454, + "k": 1, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967 + ] + }, + { + "test_f1": 0.8722645281000476, + "test_acc": 0.8726436781609196, + "test_auc": 0.9188758184966719, + "val_auc": 0.9208603896103896, + "k": 2, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352 + ] + }, + { + "test_f1": 0.937436191321043, + "test_acc": 0.9374712643678161, + "test_auc": 0.9714904418529142, + "val_auc": 0.9646915584415584, + "k": 5, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 495, + 1682, + 19873 + ] + }, + { + "test_f1": 0.9475568074308742, + "test_acc": 0.9475862068965517, + "test_auc": 0.9789812083987228, + "val_auc": 0.984577922077922, + "k": 10, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 495, + 1682, + 19873, + 178, + 12346, + 108, + 859, + 507 + ] + }, + { + "test_f1": 0.9452735898618037, + "test_acc": 0.9452873563218391, + "test_auc": 0.9812067076140485, + "val_auc": 0.9959415584415584, + "k": 16, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 495, + 1682, + 19873, + 178, + 12346, + 108, + 859, + 507, + 960, + 182, + 1168, + 1769, + 1037, + 1797 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c91a348ee8d55cb9a3b2dec3eaf0ac491c4775c6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8954591097324247, + "test_acc": 0.895632183908046, + "test_auc": 0.9452144156880783, + "val_auc": 0.9614448051948052, + "k": 1, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335 + ] + }, + { + "test_f1": 0.9189492904593975, + "test_acc": 0.919080459770115, + "test_auc": 0.96543162170572, + "val_auc": 0.9715909090909091, + "k": 2, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443 + ] + }, + { + "test_f1": 0.960458666784231, + "test_acc": 0.9604597701149425, + "test_auc": 0.9922420382596461, + "val_auc": 0.9967532467532468, + "k": 5, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1079, + 1515 + ] + }, + { + "test_f1": 0.9632173644504474, + "test_acc": 0.9632183908045977, + "test_auc": 0.9936642607825099, + "val_auc": 0.9967532467532467, + "k": 10, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1079, + 1515, + 878, + 21, + 1281, + 165, + 581 + ] + }, + { + "test_f1": 0.9701132241140344, + "test_acc": 0.9701149425287356, + "test_auc": 0.9947474159857135, + "val_auc": 1.0, + "k": 16, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1079, + 1515, + 878, + 21, + 1281, + 165, + 581, + 760, + 235, + 1928, + 5792, + 1631, + 335 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a4613d8e6e0cc9c54ec46836b68d1a0721f5a842 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6998132868016239, + "test_acc": 0.7013836477987422, + "test_auc": 0.7256446327689338, + "val_auc": 0.7442977190876351, + "k": 1, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313 + ] + }, + { + "test_f1": 0.698689559125398, + "test_acc": 0.700377358490566, + "test_auc": 0.7256446327689338, + "val_auc": 0.7442977190876351, + "k": 2, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757 + ] + }, + { + "test_f1": 0.7037508700137302, + "test_acc": 0.7059119496855346, + "test_auc": 0.7641104553845468, + "val_auc": 0.807923169267707, + "k": 5, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757, + 6636, + 1067, + 4333 + ] + }, + { + "test_f1": 0.7188108104197828, + "test_acc": 0.720754716981132, + "test_auc": 0.7827211127864317, + "val_auc": 0.7795118047218887, + "k": 10, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757, + 6636, + 1067, + 4333, + 1532, + 1105, + 19000, + 121, + 433 + ] + }, + { + "test_f1": 0.7232561846903154, + "test_acc": 0.7245283018867924, + "test_auc": 0.7924799678797496, + "val_auc": 0.7783113245298119, + "k": 16, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757, + 6636, + 1067, + 4333, + 1532, + 1105, + 19000, + 121, + 433, + 24630, + 1928, + 26569, + 836, + 1019, + 6718 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..73ab3cdc2f93214629695467b81671e3908ce783 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6867451898305205, + "test_acc": 0.6881878209831255, + "test_auc": 0.7501571778607448, + "val_auc": 0.7644230769230769, + "k": 1, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394 + ] + }, + { + "test_f1": 0.6819161688032862, + "test_acc": 0.6837857666911226, + "test_auc": 0.7501571778607448, + "val_auc": 0.7644230769230769, + "k": 2, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041 + ] + }, + { + "test_f1": 0.6922232181406034, + "test_acc": 0.6933235509904622, + "test_auc": 0.746938261397548, + "val_auc": 0.7646233974358975, + "k": 5, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584 + ] + }, + { + "test_f1": 0.6913068803805701, + "test_acc": 0.6918561995597946, + "test_auc": 0.7522650836918281, + "val_auc": 0.7940705128205128, + "k": 10, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584, + 465, + 1551, + 1257, + 7649, + 28053 + ] + }, + { + "test_f1": 0.6966970852239309, + "test_acc": 0.6969919295671313, + "test_auc": 0.7657274751206825, + "val_auc": 0.8120993589743589, + "k": 16, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584, + 465, + 1551, + 1257, + 7649, + 28053, + 31, + 1737, + 24358, + 2367, + 1697, + 943 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5b15f817a3841015ef50367f249a112b679a1270 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6565941022400247, + "test_acc": 0.6578616352201258, + "test_auc": 0.7084818928670159, + "val_auc": 0.6574629851940776, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011 + ] + }, + { + "test_f1": 0.6784897114235606, + "test_acc": 0.6784905660377358, + "test_auc": 0.7543685616466793, + "val_auc": 0.7362945178071229, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 5653 + ] + }, + { + "test_f1": 0.7345876674789482, + "test_acc": 0.7345911949685534, + "test_auc": 0.8136154622754139, + "val_auc": 0.7551020408163265, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 5653, + 1927, + 24630, + 1848 + ] + }, + { + "test_f1": 0.7577344691167863, + "test_acc": 0.7577358490566037, + "test_auc": 0.8391810348755847, + "val_auc": 0.8055222088835534, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 5653, + 1927, + 24630, + 1848, + 1520, + 3801, + 7614, + 1500, + 1036 + ] + }, + { + "test_f1": 0.7582126429216806, + "test_acc": 0.7582389937106918, + "test_auc": 0.844001477410006, + "val_auc": 0.7955182072829131, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 5653, + 1927, + 24630, + 1848, + 1520, + 3801, + 7614, + 1500, + 1036, + 3089, + 1928, + 1470, + 6417, + 9605, + 17636 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3bf97f7c6232827ae65422b9b83ebfb9a0909aa7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7400821610298672, + "test_acc": 0.740125786163522, + "test_auc": 0.7999132186171888, + "val_auc": 0.7783113245298119, + "k": 1, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503 + ] + }, + { + "test_f1": 0.7494318689944057, + "test_acc": 0.7494339622641509, + "test_auc": 0.8171406395089207, + "val_auc": 0.7983193277310925, + "k": 2, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 1643 + ] + }, + { + "test_f1": 0.7679731315150438, + "test_acc": 0.7680503144654088, + "test_auc": 0.8391714150023442, + "val_auc": 0.8415366146458584, + "k": 5, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 1643, + 5891, + 17225, + 1928 + ] + }, + { + "test_f1": 0.7746840531110271, + "test_acc": 0.7748427672955975, + "test_auc": 0.8493639238551592, + "val_auc": 0.8751500600240096, + "k": 10, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 1643, + 5891, + 17225, + 1928, + 2472, + 1848, + 446, + 7303, + 18220 + ] + }, + { + "test_f1": 0.7789998250938827, + "test_acc": 0.7791194968553459, + "test_auc": 0.8557551904279223, + "val_auc": 0.8719487795118048, + "k": 16, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 1643, + 5891, + 17225, + 1928, + 2472, + 1848, + 446, + 7303, + 18220, + 7615, + 7032, + 6327, + 7708, + 2471, + 7841 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..695b6e7c6a7be7e170241e9a6099daa03497d912 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8300643937002975, + "test_acc": 0.8301886792452831, + "test_auc": 0.8643729513467318, + "val_auc": 0.8643457382953181, + "k": 1, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8557795431925598, + "test_acc": 0.8566037735849057, + "test_auc": 0.9236749130920398, + "val_auc": 0.9433773509403762, + "k": 2, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8633782320538388, + "test_acc": 0.8638993710691824, + "test_auc": 0.9328525253179873, + "val_auc": 0.9511804721888755, + "k": 5, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 100, + 7284 + ] + }, + { + "test_f1": 0.8701152721747937, + "test_acc": 0.870188679245283, + "test_auc": 0.9440308180233895, + "val_auc": 0.9639855942376951, + "k": 10, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 100, + 7284, + 4675, + 1132, + 794, + 1335, + 1874 + ] + }, + { + "test_f1": 0.8734439222788081, + "test_acc": 0.8737106918238994, + "test_auc": 0.945231023787415, + "val_auc": 0.9599839935974389, + "k": 16, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 100, + 7284, + 4675, + 1132, + 794, + 1335, + 1874, + 505, + 22141, + 6745, + 1928, + 32479, + 777 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..45859cbac1924ada078b255faca527a2af3abd03 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8622710939295734, + "test_acc": 0.8644025157232704, + "test_auc": 0.8681564728076562, + "val_auc": 0.8691476590636255, + "k": 1, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480 + ] + }, + { + "test_f1": 0.8634875477277815, + "test_acc": 0.8656603773584906, + "test_auc": 0.9407441123844223, + "val_auc": 0.9375750300120048, + "k": 2, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433 + ] + }, + { + "test_f1": 0.9188130739262544, + "test_acc": 0.9189937106918239, + "test_auc": 0.9668949783249067, + "val_auc": 0.9699879951980793, + "k": 5, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 7666 + ] + }, + { + "test_f1": 0.9189901115279374, + "test_acc": 0.9189937106918239, + "test_auc": 0.980031674698417, + "val_auc": 0.9911964785914367, + "k": 10, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 7666, + 3481, + 1335, + 1064, + 1928, + 6553 + ] + }, + { + "test_f1": 0.9330807445524426, + "test_acc": 0.9330817610062893, + "test_auc": 0.9833740743403552, + "val_auc": 0.994797919167667, + "k": 16, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 7666, + 3481, + 1335, + 1064, + 1928, + 6553, + 988, + 1500, + 993, + 4110, + 1927, + 2297 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..46d8fa503f00e3b830518ac07e1101f511172c39 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7688323917137476, + "test_acc": 0.7777777777777778, + "test_auc": 0.7751020408163266, + "val_auc": 0.7401842948717948, + "k": 1, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629 + ] + }, + { + "test_f1": 0.7373737373737373, + "test_acc": 0.7373737373737373, + "test_auc": 0.8210204081632653, + "val_auc": 0.803485576923077, + "k": 2, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625 + ] + }, + { + "test_f1": 0.7272727272727273, + "test_acc": 0.7272727272727273, + "test_auc": 0.8281632653061225, + "val_auc": 0.8485576923076922, + "k": 5, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 794, + 530 + ] + }, + { + "test_f1": 0.7978560110913052, + "test_acc": 0.797979797979798, + "test_auc": 0.8693877551020408, + "val_auc": 0.8798076923076923, + "k": 10, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 794, + 530, + 2046, + 777, + 1182, + 765, + 1345 + ] + }, + { + "test_f1": 0.8178098497247434, + "test_acc": 0.8181818181818182, + "test_auc": 0.8718367346938775, + "val_auc": 0.891826923076923, + "k": 16, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 794, + 530, + 2046, + 777, + 1182, + 765, + 1345, + 920, + 567, + 680, + 1335, + 613, + 25757 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..05fcab3e301c94183a6171dc99a6dd996f014ce5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9147043978213317, + "test_acc": 0.9147121535181236, + "test_auc": 0.9564193489725406, + "val_auc": 0.9467999999999999, + "k": 1, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655 + ] + }, + { + "test_f1": 0.9125720080921569, + "test_acc": 0.9125799573560768, + "test_auc": 0.9668303327877794, + "val_auc": 0.9740000000000001, + "k": 2, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188 + ] + }, + { + "test_f1": 0.9445613875359198, + "test_acc": 0.9445628997867804, + "test_auc": 0.9858883433351519, + "val_auc": 0.9952, + "k": 5, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188, + 1927, + 12485, + 1417 + ] + }, + { + "test_f1": 0.9466946112640567, + "test_acc": 0.9466950959488273, + "test_auc": 0.9899254409892708, + "val_auc": 0.9944, + "k": 10, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188, + 1927, + 12485, + 1417, + 6096, + 5664, + 5406, + 1625, + 1301 + ] + }, + { + "test_f1": 0.9594827467107534, + "test_acc": 0.9594882729211087, + "test_auc": 0.9927623204218948, + "val_auc": 0.9947999999999999, + "k": 16, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188, + 1927, + 12485, + 1417, + 6096, + 5664, + 5406, + 1625, + 1301, + 2187, + 3958, + 330, + 727, + 9092, + 567 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/scr/saebench_scr_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/scr/saebench_scr_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..4430300e2bac3432c9686d874fe7b88365117077 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/scr/saebench_scr_custom_sae_eval_results.json @@ -0,0 +1,326 @@ +{ + "eval_type_id": "scr", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": true, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "16178da4-05dc-4aec-b39d-88e5b454beb5", + "datetime_epoch_millis": 1776879732603, + "eval_result_metrics": { + "scr_metrics": { + "scr_dir1_threshold_2": 0.24522833486748902, + "scr_metric_threshold_2": 0.1599918274900149, + "scr_dir2_threshold_2": 0.15390957497722552, + "scr_dir1_threshold_5": 0.08218888910819386, + "scr_metric_threshold_5": 0.2446930633569621, + "scr_dir2_threshold_5": 0.23260741289072123, + "scr_dir1_threshold_10": 0.0021863691759744786, + "scr_metric_threshold_10": 0.2902708365028045, + "scr_dir2_threshold_10": 0.2671702454593483, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results", + "scr_dir1_threshold_2": 0.5256410746257707, + "scr_metric_threshold_2": 0.05440405706850797, + "scr_dir2_threshold_2": 0.05440405706850797, + "scr_dir1_threshold_5": 0.5769232238773122, + "scr_metric_threshold_5": 0.14766833137585111, + "scr_dir2_threshold_5": 0.14766833137585111, + "scr_dir1_threshold_10": 0.6153850268564743, + "scr_metric_threshold_10": 0.20725382280374663, + "scr_dir2_threshold_10": 0.20725382280374663, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results", + "scr_dir1_threshold_2": 0.23484857036936357, + "scr_metric_threshold_2": 0.17737010471824505, + "scr_dir2_threshold_2": 0.17737010471824505, + "scr_dir1_threshold_5": -0.6818178328929966, + "scr_metric_threshold_5": 0.34862389836223945, + "scr_dir2_threshold_5": 0.34862389836223945, + "scr_dir1_threshold_10": 0.09090925510917808, + "scr_metric_threshold_10": 0.4311927307857058, + "scr_dir2_threshold_10": 0.4311927307857058, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results", + "scr_dir1_threshold_2": 0.3906245634426147, + "scr_metric_threshold_2": 0.05276388042950681, + "scr_dir2_threshold_2": 0.05276388042950681, + "scr_dir1_threshold_5": 0.15624982537704588, + "scr_metric_threshold_5": 0.11055282101401862, + "scr_dir2_threshold_5": 0.11055282101401862, + "scr_dir1_threshold_10": -1.4218750873114772, + "scr_metric_threshold_10": 0.19095483181683465, + "scr_dir2_threshold_10": 0.19095483181683465, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results", + "scr_dir1_threshold_2": 0.20805342816739705, + "scr_metric_threshold_2": 0.00955420662613804, + "scr_dir2_threshold_2": 0.00955420662613804, + "scr_dir1_threshold_5": -0.2953022335742188, + "scr_metric_threshold_5": 0.015923551161094646, + "scr_dir2_threshold_5": 0.015923551161094646, + "scr_dir1_threshold_10": -0.3959732859163015, + "scr_metric_threshold_10": -0.12101906475380056, + "scr_dir2_threshold_10": -0.12101906475380056, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results", + "scr_dir1_threshold_2": 0.04210534901500186, + "scr_metric_threshold_2": 0.4920634770459469, + "scr_dir2_threshold_2": 0.4920634770459469, + "scr_dir1_threshold_5": 0.057894933322792726, + "scr_metric_threshold_5": 0.5793651112773629, + "scr_dir2_threshold_5": 0.5793651112773629, + "scr_dir1_threshold_10": 0.08421069803000372, + "scr_metric_threshold_10": 0.6190476077844602, + "scr_dir2_threshold_10": 0.6190476077844602, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results", + "scr_dir1_threshold_2": 0.23497262420079276, + "scr_metric_threshold_2": 0.11068695256721958, + "scr_dir2_threshold_2": 0.11068695256721958, + "scr_dir1_threshold_5": 0.3551913048859532, + "scr_metric_threshold_5": 0.17557251734733875, + "scr_dir2_threshold_5": 0.17557251734733875, + "scr_dir1_threshold_10": 0.4262296310048433, + "scr_metric_threshold_10": 0.23282436583052582, + "scr_dir2_threshold_10": 0.23282436583052582, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results", + "scr_dir1_threshold_2": 0.19999988910765484, + "scr_metric_threshold_2": 0.2575107614532381, + "scr_dir2_threshold_2": 0.2575107614532381, + "scr_dir1_threshold_5": 0.2604650492278843, + "scr_metric_threshold_5": 0.35193143367601315, + "scr_dir2_threshold_5": 0.35193143367601315, + "scr_dir1_threshold_10": 0.3116280049746306, + "scr_metric_threshold_10": 0.4549356870945189, + "scr_dir2_threshold_10": 0.4549356870945189, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results", + "scr_dir1_threshold_2": 0.12558118001131657, + "scr_metric_threshold_2": 0.12558118001131657, + "scr_dir2_threshold_2": 0.07692315990900178, + "scr_dir1_threshold_5": 0.227906842641778, + "scr_metric_threshold_5": 0.227906842641778, + "scr_dir2_threshold_5": 0.1312216389118512, + "scr_dir1_threshold_10": 0.3069767106604445, + "scr_metric_threshold_10": 0.3069767106604445, + "scr_dir2_threshold_10": 0.1221719823127949, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_scr", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": null +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/tpp/saebench_tpp_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/tpp/saebench_tpp_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..78be2eeff20073d444bbce2e6992293ab618c4fc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-109863-tokens-450000000/tpp/saebench_tpp_custom_sae_eval_results.json @@ -0,0 +1,297 @@ +{ + "eval_type_id": "tpp", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": false, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "8e658367-d1f8-4296-807c-f348d80fbe1f", + "datetime_epoch_millis": 1776879903944, + "eval_result_metrics": { + "tpp_metrics": { + "tpp_threshold_2_total_metric": 0.04852499067783356, + "tpp_threshold_2_intended_diff_only": 0.057199996709823606, + "tpp_threshold_2_unintended_diff_only": 0.008675006031990052, + "tpp_threshold_5_total_metric": 0.08755000531673432, + "tpp_threshold_5_intended_diff_only": 0.10810000896453857, + "tpp_threshold_5_unintended_diff_only": 0.020550003647804262, + "tpp_threshold_10_total_metric": 0.12970000505447388, + "tpp_threshold_10_intended_diff_only": 0.1782000124454498, + "tpp_threshold_10_unintended_diff_only": 0.04850000739097595, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results", + "tpp_threshold_2_total_metric": 0.08679999709129334, + "tpp_threshold_2_intended_diff_only": 0.10280001163482666, + "tpp_threshold_2_unintended_diff_only": 0.016000014543533326, + "tpp_threshold_5_total_metric": 0.14740001261234284, + "tpp_threshold_5_intended_diff_only": 0.18460001945495605, + "tpp_threshold_5_unintended_diff_only": 0.03720000684261322, + "tpp_threshold_10_total_metric": 0.2154000073671341, + "tpp_threshold_10_intended_diff_only": 0.3056000232696533, + "tpp_threshold_10_unintended_diff_only": 0.09020001590251922, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results", + "tpp_threshold_2_total_metric": 0.01024998426437378, + "tpp_threshold_2_intended_diff_only": 0.011599981784820556, + "tpp_threshold_2_unintended_diff_only": 0.0013499975204467774, + "tpp_threshold_5_total_metric": 0.027699998021125792, + "tpp_threshold_5_intended_diff_only": 0.031599998474121094, + "tpp_threshold_5_unintended_diff_only": 0.0039000004529953004, + "tpp_threshold_10_total_metric": 0.04400000274181366, + "tpp_threshold_10_intended_diff_only": 0.050800001621246337, + "tpp_threshold_10_unintended_diff_only": 0.006799998879432678, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_tpp", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1": { + "0": { + "tpp_threshold_2_total_metric": 0.16724996268749237, + "tpp_threshold_2_intended_diff_only": 0.20499998331069946, + "tpp_threshold_2_unintended_diff_only": 0.03775002062320709, + "tpp_threshold_5_total_metric": 0.19849996268749237, + "tpp_threshold_5_intended_diff_only": 0.24699997901916504, + "tpp_threshold_5_unintended_diff_only": 0.04850001633167267, + "tpp_threshold_10_total_metric": 0.2202499657869339, + "tpp_threshold_10_intended_diff_only": 0.296999990940094, + "tpp_threshold_10_unintended_diff_only": 0.0767500251531601 + }, + "1": { + "tpp_threshold_2_total_metric": 0.07999999821186066, + "tpp_threshold_2_intended_diff_only": 0.10600000619888306, + "tpp_threshold_2_unintended_diff_only": 0.0260000079870224, + "tpp_threshold_5_total_metric": 0.09274999797344208, + "tpp_threshold_5_intended_diff_only": 0.125, + "tpp_threshold_5_unintended_diff_only": 0.03225000202655792, + "tpp_threshold_10_total_metric": 0.1642499715089798, + "tpp_threshold_10_intended_diff_only": 0.22699999809265137, + "tpp_threshold_10_unintended_diff_only": 0.06275002658367157 + }, + "2": { + "tpp_threshold_2_total_metric": 0.12900003790855408, + "tpp_threshold_2_intended_diff_only": 0.14000004529953003, + "tpp_threshold_2_unintended_diff_only": 0.011000007390975952, + "tpp_threshold_5_total_metric": 0.1937500536441803, + "tpp_threshold_5_intended_diff_only": 0.2640000581741333, + "tpp_threshold_5_unintended_diff_only": 0.070250004529953, + "tpp_threshold_10_total_metric": 0.1977500468492508, + "tpp_threshold_10_intended_diff_only": 0.3340000510215759, + "tpp_threshold_10_unintended_diff_only": 0.13625000417232513 + }, + "6": { + "tpp_threshold_2_total_metric": 0.0495000034570694, + "tpp_threshold_2_intended_diff_only": 0.054000020027160645, + "tpp_threshold_2_unintended_diff_only": 0.0045000165700912476, + "tpp_threshold_5_total_metric": 0.14525005221366882, + "tpp_threshold_5_intended_diff_only": 0.16300004720687866, + "tpp_threshold_5_unintended_diff_only": 0.01774999499320984, + "tpp_threshold_10_total_metric": 0.2990000396966934, + "tpp_threshold_10_intended_diff_only": 0.362000048160553, + "tpp_threshold_10_unintended_diff_only": 0.06300000846385956 + }, + "9": { + "tpp_threshold_2_total_metric": 0.008249983191490173, + "tpp_threshold_2_intended_diff_only": 0.009000003337860107, + "tpp_threshold_2_unintended_diff_only": 0.0007500201463699341, + "tpp_threshold_5_total_metric": 0.1067499965429306, + "tpp_threshold_5_intended_diff_only": 0.12400001287460327, + "tpp_threshold_5_unintended_diff_only": 0.01725001633167267, + "tpp_threshold_10_total_metric": 0.19575001299381256, + "tpp_threshold_10_intended_diff_only": 0.30800002813339233, + "tpp_threshold_10_unintended_diff_only": 0.11225001513957977 + } + }, + "canrager/amazon_reviews_mcauley_1and5": { + "1": { + "tpp_threshold_2_total_metric": 0.01099996268749237, + "tpp_threshold_2_intended_diff_only": 0.006999969482421875, + "tpp_threshold_2_unintended_diff_only": -0.003999993205070496, + "tpp_threshold_5_total_metric": 0.010499998927116394, + "tpp_threshold_5_intended_diff_only": 0.009999990463256836, + "tpp_threshold_5_unintended_diff_only": -0.0005000084638595581, + "tpp_threshold_10_total_metric": 0.0040000081062316895, + "tpp_threshold_10_intended_diff_only": 0.013000011444091797, + "tpp_threshold_10_unintended_diff_only": 0.009000003337860107 + }, + "2": { + "tpp_threshold_2_total_metric": 0.005500033497810364, + "tpp_threshold_2_intended_diff_only": 0.012000024318695068, + "tpp_threshold_2_unintended_diff_only": 0.006499990820884705, + "tpp_threshold_5_total_metric": 0.04000002145767212, + "tpp_threshold_5_intended_diff_only": 0.04900002479553223, + "tpp_threshold_5_unintended_diff_only": 0.009000003337860107, + "tpp_threshold_10_total_metric": 0.08900001645088196, + "tpp_threshold_10_intended_diff_only": 0.09700000286102295, + "tpp_threshold_10_unintended_diff_only": 0.007999986410140991 + }, + "3": { + "tpp_threshold_2_total_metric": 0.018749982118606567, + "tpp_threshold_2_intended_diff_only": 0.014999985694885254, + "tpp_threshold_2_unintended_diff_only": -0.0037499964237213135, + "tpp_threshold_5_total_metric": 0.020250022411346436, + "tpp_threshold_5_intended_diff_only": 0.021000027656555176, + "tpp_threshold_5_unintended_diff_only": 0.0007500052452087402, + "tpp_threshold_10_total_metric": 0.016249969601631165, + "tpp_threshold_10_intended_diff_only": 0.02499997615814209, + "tpp_threshold_10_unintended_diff_only": 0.008750006556510925 + }, + "5": { + "tpp_threshold_2_total_metric": -0.005750030279159546, + "tpp_threshold_2_intended_diff_only": 0.0029999613761901855, + "tpp_threshold_2_unintended_diff_only": 0.008749991655349731, + "tpp_threshold_5_total_metric": 0.01024995744228363, + "tpp_threshold_5_intended_diff_only": 0.012999951839447021, + "tpp_threshold_5_unintended_diff_only": 0.002749994397163391, + "tpp_threshold_10_total_metric": 0.018249988555908203, + "tpp_threshold_10_intended_diff_only": 0.023000001907348633, + "tpp_threshold_10_unintended_diff_only": 0.00475001335144043 + }, + "6": { + "tpp_threshold_2_total_metric": 0.02174997329711914, + "tpp_threshold_2_intended_diff_only": 0.0209999680519104, + "tpp_threshold_2_unintended_diff_only": -0.0007500052452087402, + "tpp_threshold_5_total_metric": 0.05749998986721039, + "tpp_threshold_5_intended_diff_only": 0.06499999761581421, + "tpp_threshold_5_unintended_diff_only": 0.007500007748603821, + "tpp_threshold_10_total_metric": 0.09250003099441528, + "tpp_threshold_10_intended_diff_only": 0.09600001573562622, + "tpp_threshold_10_unintended_diff_only": 0.0034999847412109375 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ff2e4acefc07d456d6e10ea544473ed5ebc2f4ef --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9598954703123229, + "test_acc": 0.9599045346062053, + "test_auc": 0.9887555866634586, + "val_auc": 0.9841269841269842, + "k": 1, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431 + ] + }, + { + "test_f1": 0.9770882950488093, + "test_acc": 0.9770883054892602, + "test_auc": 0.9967168099331423, + "val_auc": 0.993993993993994, + "k": 2, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1758 + ] + }, + { + "test_f1": 0.9842474563203397, + "test_acc": 0.9842482100238663, + "test_auc": 0.9980022893472444, + "val_auc": 0.9987129987129987, + "k": 5, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1758, + 1472, + 542, + 461 + ] + }, + { + "test_f1": 0.9890213746557301, + "test_acc": 0.9890214797136038, + "test_auc": 0.9992809335287299, + "val_auc": 0.9978549978549979, + "k": 10, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1758, + 1472, + 542, + 461, + 1411, + 5951, + 786, + 1500, + 7560 + ] + }, + { + "test_f1": 0.9780401975957291, + "test_acc": 0.9780429594272076, + "test_auc": 0.9961613333624969, + "val_auc": 0.9969969969969971, + "k": 16, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1758, + 1472, + 542, + 461, + 1411, + 5951, + 786, + 1500, + 7560, + 583, + 1928, + 1608, + 126, + 1634, + 1872 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4c605256d09906f594622736e03c4e1388239398 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.933224147623617, + "test_acc": 0.933246921581335, + "test_auc": 0.9770770750589706, + "val_auc": 0.9706490285241836, + "k": 1, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1960 + ] + }, + { + "test_f1": 0.9397191439285256, + "test_acc": 0.9397278029812054, + "test_auc": 0.9797819936426013, + "val_auc": 0.9710624224886316, + "k": 2, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1960, + 152 + ] + }, + { + "test_f1": 0.9442613296893301, + "test_acc": 0.9442644199611148, + "test_auc": 0.9800810467530896, + "val_auc": 0.9710624224886316, + "k": 5, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1960, + 152, + 1441, + 4110, + 821 + ] + }, + { + "test_f1": 0.9630472421335851, + "test_acc": 0.9630589760207389, + "test_auc": 0.9833336693480642, + "val_auc": 0.9681686647374949, + "k": 10, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1960, + 152, + 1441, + 4110, + 821, + 1500, + 786, + 7856, + 1417, + 920 + ] + }, + { + "test_f1": 0.9643438301288979, + "test_acc": 0.9643551523007129, + "test_auc": 0.9851078271271413, + "val_auc": 0.9714758164530798, + "k": 16, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1960, + 152, + 1441, + 4110, + 821, + 1500, + 786, + 7856, + 1417, + 920, + 908, + 1118, + 654, + 302, + 1411, + 1703 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9f8bf13cab033417c8b5eb1cd7ae578bb954aaf5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.587125980000197, + "test_acc": 0.5939622641509434, + "test_auc": 0.6048029495543972, + "val_auc": 0.6196478591436575, + "k": 1, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1438 + ] + }, + { + "test_f1": 0.5856546960350886, + "test_acc": 0.5927044025157233, + "test_auc": 0.6048029495543972, + "val_auc": 0.6476590636254501, + "k": 2, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1438, + 7284 + ] + }, + { + "test_f1": 0.5892739901383439, + "test_acc": 0.5989937106918239, + "test_auc": 0.601629150848726, + "val_auc": 0.6202480992396958, + "k": 5, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1438, + 7284, + 152, + 4197, + 1335 + ] + }, + { + "test_f1": 0.5914589468181268, + "test_acc": 0.5984905660377359, + "test_auc": 0.6264869033020468, + "val_auc": 0.6086434573829531, + "k": 10, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1438, + 7284, + 152, + 4197, + 1335, + 1178, + 5792, + 1036, + 988, + 878 + ] + }, + { + "test_f1": 0.5978285558977487, + "test_acc": 0.6027672955974843, + "test_auc": 0.6401356047710521, + "val_auc": 0.6158463385354142, + "k": 16, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1438, + 7284, + 152, + 4197, + 1335, + 1178, + 5792, + 1036, + 988, + 878, + 10815, + 394, + 1255, + 969, + 26301, + 918 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e06191af9e2a13b791ae7d47ae562f1b91342b9e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6726758350957406, + "test_acc": 0.6807547169811321, + "test_auc": 0.6898230095216492, + "val_auc": 0.7006802721088435, + "k": 1, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567 + ] + }, + { + "test_f1": 0.678253188066235, + "test_acc": 0.6807547169811321, + "test_auc": 0.7277629035410247, + "val_auc": 0.7362945178071229, + "k": 2, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 1164 + ] + }, + { + "test_f1": 0.709170468311081, + "test_acc": 0.710691823899371, + "test_auc": 0.7660820230897211, + "val_auc": 0.7871148459383754, + "k": 5, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 1164, + 1629, + 4505, + 1335 + ] + }, + { + "test_f1": 0.7169344935946191, + "test_acc": 0.7177358490566038, + "test_auc": 0.7723183590724012, + "val_auc": 0.8035214085634255, + "k": 10, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 1164, + 1629, + 4505, + 1335, + 22873, + 6411, + 1731, + 1874, + 20445 + ] + }, + { + "test_f1": 0.7243159967586054, + "test_acc": 0.7250314465408805, + "test_auc": 0.77869608187626, + "val_auc": 0.8351340536214485, + "k": 16, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 1164, + 1629, + 4505, + 1335, + 22873, + 6411, + 1731, + 1874, + 20445, + 523, + 4196, + 20154, + 993, + 1551, + 534 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..82cdfc5151ccd8f9781bb079857286bd801456d3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7287846629204201, + "test_acc": 0.7295597484276729, + "test_auc": 0.8306298536057816, + "val_auc": 0.8557422969187675, + "k": 1, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 786 + ] + }, + { + "test_f1": 0.729388076903271, + "test_acc": 0.730062893081761, + "test_auc": 0.8306298536057816, + "val_auc": 0.8557422969187675, + "k": 2, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 786, + 1411 + ] + }, + { + "test_f1": 0.777084916386665, + "test_acc": 0.7788679245283019, + "test_auc": 0.8596322524983825, + "val_auc": 0.8907563025210085, + "k": 5, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 786, + 1411, + 800, + 1335, + 373 + ] + }, + { + "test_f1": 0.8027171888391386, + "test_acc": 0.8032704402515723, + "test_auc": 0.9000948316977863, + "val_auc": 0.9197679071628653, + "k": 10, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 786, + 1411, + 800, + 1335, + 373, + 6718, + 1758, + 181, + 1417, + 5054 + ] + }, + { + "test_f1": 0.7989322472023445, + "test_acc": 0.7989937106918239, + "test_auc": 0.9045465546170834, + "val_auc": 0.9205682272909164, + "k": 16, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 786, + 1411, + 800, + 1335, + 373, + 6718, + 1758, + 181, + 1417, + 5054, + 1360, + 1183, + 19471, + 993, + 1108, + 1703 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ab4704247e3d778f2262a598c6e09ccb0c190ff8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.762363759199011, + "test_acc": 0.769308176100629, + "test_auc": 0.773356292764134, + "val_auc": 0.7959183673469388, + "k": 1, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562 + ] + }, + { + "test_f1": 0.7621256750226106, + "test_acc": 0.7690566037735849, + "test_auc": 0.8390185096487329, + "val_auc": 0.8255302120848339, + "k": 2, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562, + 4675 + ] + }, + { + "test_f1": 0.8367257196920327, + "test_acc": 0.8372327044025157, + "test_auc": 0.9001784992795221, + "val_auc": 0.9125650260104041, + "k": 5, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562, + 4675, + 100, + 7867, + 10076 + ] + }, + { + "test_f1": 0.8771848081328371, + "test_acc": 0.8772327044025158, + "test_auc": 0.9394227721639348, + "val_auc": 0.9483793517406962, + "k": 10, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562, + 4675, + 100, + 7867, + 10076, + 7284, + 1438, + 253, + 186, + 3654 + ] + }, + { + "test_f1": 0.8774327537122911, + "test_acc": 0.8774842767295598, + "test_auc": 0.9416830626436019, + "val_auc": 0.9539815926370548, + "k": 16, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562, + 4675, + 100, + 7867, + 10076, + 7284, + 1438, + 253, + 186, + 3654, + 1431, + 7290, + 5704, + 373, + 26140, + 1108 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c04656a45e054262d614580b68656a37eec62fa9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5298129797092886, + "test_acc": 0.5364102564102564, + "test_auc": 0.5460581681085267, + "val_auc": 0.5703794369645043, + "k": 1, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264 + ] + }, + { + "test_f1": 0.556454068490933, + "test_acc": 0.558974358974359, + "test_auc": 0.5781886424075134, + "val_auc": 0.5895552835577316, + "k": 2, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945 + ] + }, + { + "test_f1": 0.5765352623625124, + "test_acc": 0.5805128205128205, + "test_auc": 0.6212698185612818, + "val_auc": 0.6319869441044471, + "k": 5, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 1065, + 551, + 1924 + ] + }, + { + "test_f1": 0.579099608246585, + "test_acc": 0.5805128205128205, + "test_auc": 0.6207922375197764, + "val_auc": 0.663810689514484, + "k": 10, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 1065, + 551, + 1924, + 6892, + 1815, + 1523, + 386, + 238 + ] + }, + { + "test_f1": 0.5797561489316889, + "test_acc": 0.5815384615384616, + "test_auc": 0.6287701551822802, + "val_auc": 0.7192982456140351, + "k": 16, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 1065, + 551, + 1924, + 6892, + 1815, + 1523, + 386, + 238, + 456, + 1072, + 495, + 4010, + 1551, + 1985 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..31b7c6b53ec097ccbd21fda5db37a2ce3c6e2d09 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5249383013874582, + "test_acc": 0.5251282051282051, + "test_auc": 0.5433020836839801, + "val_auc": 0.5324357405140758, + "k": 1, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 152 + ] + }, + { + "test_f1": 0.519250493096647, + "test_acc": 0.52, + "test_auc": 0.5233320429528394, + "val_auc": 0.5395756833945329, + "k": 2, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 152, + 239 + ] + }, + { + "test_f1": 0.5311784867135978, + "test_acc": 0.5312820512820513, + "test_auc": 0.5546967044804254, + "val_auc": 0.565279477764178, + "k": 5, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 152, + 239, + 6341, + 1264, + 4296 + ] + }, + { + "test_f1": 0.5343315426781452, + "test_acc": 0.5343589743589744, + "test_auc": 0.5563924327599555, + "val_auc": 0.5560995512035903, + "k": 10, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 152, + 239, + 6341, + 1264, + 4296, + 827, + 30, + 1409, + 601, + 400 + ] + }, + { + "test_f1": 0.5435897435897435, + "test_acc": 0.5435897435897435, + "test_auc": 0.5374911636987916, + "val_auc": 0.5059159526723787, + "k": 16, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 152, + 239, + 6341, + 1264, + 4296, + 827, + 30, + 1409, + 601, + 400, + 1847, + 13161, + 1689, + 6821, + 1815, + 2769 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..85acec21aaf8238c5bbfe717be3a8d71586914f5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5435695779547854, + "test_acc": 0.5435897435897435, + "test_auc": 0.5630175547850675, + "val_auc": 0.5964912280701754, + "k": 1, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815 + ] + }, + { + "test_f1": 0.5689643236620343, + "test_acc": 0.5712820512820512, + "test_auc": 0.573362338842697, + "val_auc": 0.598531211750306, + "k": 2, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156 + ] + }, + { + "test_f1": 0.5378628388682212, + "test_acc": 0.5456410256410257, + "test_auc": 0.5722430740229576, + "val_auc": 0.5738474092207262, + "k": 5, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 360, + 7616, + 1409 + ] + }, + { + "test_f1": 0.5288432393489865, + "test_acc": 0.5343589743589744, + "test_auc": 0.5533838825865958, + "val_auc": 0.6483068135454916, + "k": 10, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 360, + 7616, + 1409, + 24861, + 1684, + 698, + 15467, + 16809 + ] + }, + { + "test_f1": 0.5411828412871653, + "test_acc": 0.5425641025641026, + "test_auc": 0.5632637088901605, + "val_auc": 0.6315789473684211, + "k": 16, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 360, + 7616, + 1409, + 24861, + 1684, + 698, + 15467, + 16809, + 191, + 1847, + 400, + 1611, + 14220, + 1590 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dca43e6317f0e77b4bbdea6d403c1de731aa4be1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8066616664747505, + "test_acc": 0.8080808080808081, + "test_auc": 0.8404081632653061, + "val_auc": 0.8679887820512822, + "k": 1, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997 + ] + }, + { + "test_f1": 0.8066616664747505, + "test_acc": 0.8080808080808081, + "test_auc": 0.8359183673469388, + "val_auc": 0.8743990384615383, + "k": 2, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369 + ] + }, + { + "test_f1": 0.8374557718282082, + "test_acc": 0.8383838383838383, + "test_auc": 0.9175510204081633, + "val_auc": 0.9100560897435898, + "k": 5, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 12946, + 4531, + 2867 + ] + }, + { + "test_f1": 0.807255349190833, + "test_acc": 0.8080808080808081, + "test_auc": 0.8653061224489796, + "val_auc": 0.9419070512820513, + "k": 10, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 12946, + 4531, + 2867, + 1895, + 1212, + 851, + 2412, + 2900 + ] + }, + { + "test_f1": 0.7770958356064739, + "test_acc": 0.7777777777777778, + "test_auc": 0.8448979591836735, + "val_auc": 0.9431089743589743, + "k": 16, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 12946, + 4531, + 2867, + 1895, + 1212, + 851, + 2412, + 2900, + 314, + 1499, + 1284, + 1847, + 7481, + 2780 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..85588ab374f32bf4531425da1c7a5f864e6ad58a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8886617839096123, + "test_acc": 0.8888888888888888, + "test_auc": 0.9136734693877551, + "val_auc": 0.9507211538461539, + "k": 1, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8786393345216874, + "test_acc": 0.8787878787878788, + "test_auc": 0.9179591836734694, + "val_auc": 0.9627403846153846, + "k": 2, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889 + ] + }, + { + "test_f1": 0.9089050959260463, + "test_acc": 0.9090909090909091, + "test_auc": 0.9538775510204082, + "val_auc": 0.9805689102564101, + "k": 5, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 1592, + 789, + 1065 + ] + }, + { + "test_f1": 0.8787878787878788, + "test_acc": 0.8787878787878788, + "test_auc": 0.9597959183673469, + "val_auc": 0.9871794871794872, + "k": 10, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 1592, + 789, + 1065, + 581, + 901, + 2412, + 2335, + 2832 + ] + }, + { + "test_f1": 0.8686600618824769, + "test_acc": 0.8686868686868687, + "test_auc": 0.9583673469387755, + "val_auc": 0.9907852564102564, + "k": 16, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 1592, + 789, + 1065, + 581, + 901, + 2412, + 2335, + 2832, + 82, + 7481, + 2900, + 2314, + 925, + 6048 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1bac72563c1de4c6d25947af87b73212b07bf2d7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8460372598303634, + "test_acc": 0.8484848484848485, + "test_auc": 0.856734693877551, + "val_auc": 0.8036858974358975, + "k": 1, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388 + ] + }, + { + "test_f1": 0.8460372598303634, + "test_acc": 0.8484848484848485, + "test_auc": 0.9275510204081632, + "val_auc": 0.8918269230769231, + "k": 2, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 1847 + ] + }, + { + "test_f1": 0.7777777777777778, + "test_acc": 0.7777777777777778, + "test_auc": 0.9142857142857144, + "val_auc": 0.9479166666666667, + "k": 5, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 1847, + 1284, + 337, + 5463 + ] + }, + { + "test_f1": 0.8279318478603097, + "test_acc": 0.8282828282828283, + "test_auc": 0.9375510204081632, + "val_auc": 0.9631410256410258, + "k": 10, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 1847, + 1284, + 337, + 5463, + 7384, + 1596, + 1305, + 1815, + 4889 + ] + }, + { + "test_f1": 0.8281776631271375, + "test_acc": 0.8282828282828283, + "test_auc": 0.9500000000000001, + "val_auc": 0.9591346153846154, + "k": 16, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 1847, + 1284, + 337, + 5463, + 7384, + 1596, + 1305, + 1815, + 4889, + 2335, + 2832, + 6048, + 4201, + 2900, + 2867 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8ffd56f17651d000e581884c4172d59fd34f4f5d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.632534665603286, + "test_acc": 0.658864325964456, + "test_auc": 0.6683656981035082, + "val_auc": 0.7172000000000001, + "k": 1, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.6317909208983842, + "test_acc": 0.6584308625921109, + "test_auc": 0.7198995612380333, + "val_auc": 0.7727999999999999, + "k": 2, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889 + ] + }, + { + "test_f1": 0.683569954563861, + "test_acc": 0.6835717381881231, + "test_auc": 0.7745734509177324, + "val_auc": 0.864, + "k": 5, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 12946, + 7384 + ] + }, + { + "test_f1": 0.6956893738929083, + "test_acc": 0.6965756393584742, + "test_auc": 0.7922464342135128, + "val_auc": 0.858, + "k": 10, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 12946, + 7384, + 2314, + 1684, + 1851, + 3417, + 1478 + ] + }, + { + "test_f1": 0.7219890909633266, + "test_acc": 0.7221499783268314, + "test_auc": 0.8123623701864326, + "val_auc": 0.87, + "k": 16, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 12946, + 7384, + 2314, + 1684, + 1851, + 3417, + 1478, + 901, + 1596, + 2000, + 6396, + 1065, + 26171 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..371b7350fb0506fd97ef894c50387e22bada408b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.638137694844405, + "test_acc": 0.6640658864325965, + "test_auc": 0.6742481748313871, + "val_auc": 0.7108000000000001, + "k": 1, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.7120340407437366, + "test_acc": 0.7160814911140009, + "test_auc": 0.7497906899490592, + "val_auc": 0.7991999999999999, + "k": 2, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384 + ] + }, + { + "test_f1": 0.7336551021515249, + "test_acc": 0.7364542696142176, + "test_auc": 0.8223705471823184, + "val_auc": 0.8738, + "k": 5, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 7481, + 3526, + 1284 + ] + }, + { + "test_f1": 0.7498967607663519, + "test_acc": 0.752058951018639, + "test_auc": 0.8559157709298777, + "val_auc": 0.902, + "k": 10, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 7481, + 3526, + 1284, + 4201, + 4037, + 6661, + 4845, + 2185 + ] + }, + { + "test_f1": 0.7819366730156628, + "test_acc": 0.7824013870827915, + "test_auc": 0.8871176239814454, + "val_auc": 0.9064000000000001, + "k": 16, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 7481, + 3526, + 1284, + 4201, + 4037, + 6661, + 4845, + 2185, + 6244, + 789, + 1851, + 2900, + 5388, + 1091 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..429d8339aa76dae67692ebd6533b828edafcefd7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8288367582289091, + "test_acc": 0.8305158214130905, + "test_auc": 0.8500949222959923, + "val_auc": 0.8752, + "k": 1, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.834335551668128, + "test_acc": 0.8361508452535761, + "test_auc": 0.8819701750087557, + "val_auc": 0.8754, + "k": 2, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889 + ] + }, + { + "test_f1": 0.8217968161152149, + "test_acc": 0.8248807975726051, + "test_auc": 0.8935765488567989, + "val_auc": 0.9376, + "k": 5, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 1592, + 12946, + 4663 + ] + }, + { + "test_f1": 0.8326116372370318, + "test_acc": 0.835283918508886, + "test_auc": 0.9213843473660002, + "val_auc": 0.9508, + "k": 10, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 1592, + 12946, + 4663, + 1065, + 6048, + 789, + 3417, + 901 + ] + }, + { + "test_f1": 0.8321897881545541, + "test_acc": 0.834850455136541, + "test_auc": 0.9219059314785782, + "val_auc": 0.9508, + "k": 16, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 1592, + 12946, + 4663, + 1065, + 6048, + 789, + 3417, + 901, + 82, + 415, + 3526, + 1546, + 5843, + 1851 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6b62321a525259294ef19ca24dcf30785bc3a04e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8532913566555026, + "test_acc": 0.8533333333333334, + "test_auc": 0.9428396505874037, + "val_auc": 0.9341085271317828, + "k": 1, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3818 + ] + }, + { + "test_f1": 0.868713805585676, + "test_acc": 0.8687179487179487, + "test_auc": 0.9565927222540141, + "val_auc": 0.9594043247654018, + "k": 2, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3818, + 3958 + ] + }, + { + "test_f1": 0.949727407016708, + "test_acc": 0.9497435897435897, + "test_auc": 0.989118726226142, + "val_auc": 0.9906160750713995, + "k": 5, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3818, + 3958, + 1847, + 508, + 26171 + ] + }, + { + "test_f1": 0.9620490463944382, + "test_acc": 0.9620512820512821, + "test_auc": 0.9919715891877336, + "val_auc": 0.9942880456956344, + "k": 10, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3818, + 3958, + 1847, + 508, + 26171, + 26466, + 5463, + 386, + 1158, + 1279 + ] + }, + { + "test_f1": 0.9620476888952731, + "test_acc": 0.9620512820512821, + "test_auc": 0.9922240549365469, + "val_auc": 0.9914320685434517, + "k": 16, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3818, + 3958, + 1847, + 508, + 26171, + 26466, + 5463, + 386, + 1158, + 1279, + 1851, + 779, + 152, + 1482, + 5973, + 1253 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..524329baf34c6b153a18165fbb6f0b4fbccb7909 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8629804063922483, + "test_acc": 0.8646153846153846, + "test_auc": 0.8707017706264516, + "val_auc": 0.8859649122807017, + "k": 1, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.9117296857113426, + "test_acc": 0.9117948717948718, + "test_auc": 0.9360546335880432, + "val_auc": 0.9122807017543859, + "k": 2, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 3475 + ] + }, + { + "test_f1": 0.9333294058280671, + "test_acc": 0.9333333333333333, + "test_auc": 0.985525297068031, + "val_auc": 0.9979600163198694, + "k": 5, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 3475, + 1158, + 835, + 1279 + ] + }, + { + "test_f1": 0.9333240748631482, + "test_acc": 0.9333333333333333, + "test_auc": 0.9860133975157371, + "val_auc": 0.9979600163198694, + "k": 10, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 3475, + 1158, + 835, + 1279, + 12813, + 992, + 2590, + 1332, + 581 + ] + }, + { + "test_f1": 0.9784612665710571, + "test_acc": 0.9784615384615385, + "test_auc": 0.9968946712895951, + "val_auc": 1.0, + "k": 16, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 3475, + 1158, + 835, + 1279, + 12813, + 992, + 2590, + 1332, + 581, + 3818, + 4934, + 1847, + 1065, + 2708, + 1482 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9ee9c47a89c39365a38bb4ed932f1bed4032dea5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.956909206525174, + "test_acc": 0.9569230769230769, + "test_auc": 0.9808673040024236, + "val_auc": 0.9975520195838433, + "k": 1, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037 + ] + }, + { + "test_f1": 0.9579335842821622, + "test_acc": 0.9579487179487179, + "test_auc": 0.9836654660517723, + "val_auc": 0.9979600163198694, + "k": 2, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508 + ] + }, + { + "test_f1": 0.9671732024768461, + "test_acc": 0.9671794871794872, + "test_auc": 0.9901580435587571, + "val_auc": 1.0, + "k": 5, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1253, + 1028, + 5472 + ] + }, + { + "test_f1": 0.9723075175231607, + "test_acc": 0.9723076923076923, + "test_auc": 0.9951821119601441, + "val_auc": 1.0, + "k": 10, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1253, + 1028, + 5472, + 1158, + 1039, + 3818, + 1279, + 778 + ] + }, + { + "test_f1": 0.9794862731541124, + "test_acc": 0.9794871794871794, + "test_auc": 0.9966001279159794, + "val_auc": 1.0, + "k": 16, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1253, + 1028, + 5472, + 1158, + 1039, + 3818, + 1279, + 778, + 1079, + 1065, + 29801, + 1478, + 1602, + 238 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7e5c99d2917c5b838499918d35b88ee307dacb17 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5918832497751685, + "test_acc": 0.5977459904638058, + "test_auc": 0.6032668902313458, + "val_auc": 0.6572, + "k": 1, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2046 + ] + }, + { + "test_f1": 0.6044508482860358, + "test_acc": 0.6142175986129172, + "test_auc": 0.6719581650460482, + "val_auc": 0.742, + "k": 2, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2046, + 1351 + ] + }, + { + "test_f1": 0.6445539486395127, + "test_acc": 0.6545296922410057, + "test_auc": 0.716528429340384, + "val_auc": 0.7871999999999999, + "k": 5, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2046, + 1351, + 3078, + 30138, + 1441 + ] + }, + { + "test_f1": 0.673925217693897, + "test_acc": 0.6757693974859125, + "test_auc": 0.7479080268337739, + "val_auc": 0.8028, + "k": 10, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2046, + 1351, + 3078, + 30138, + 1441, + 848, + 2038, + 654, + 7856, + 343 + ] + }, + { + "test_f1": 0.6844286529513368, + "test_acc": 0.6861725184221933, + "test_auc": 0.7551688684931631, + "val_auc": 0.7984, + "k": 16, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2046, + 1351, + 3078, + 30138, + 1441, + 848, + 2038, + 654, + 7856, + 343, + 7595, + 268, + 817, + 2691, + 52, + 4110 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1e6bfbf253bfd5492a4685dcd04b811cfc6b02a7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6267988039593516, + "test_acc": 0.6276549631556133, + "test_auc": 0.6550544055820023, + "val_auc": 0.6252000000000001, + "k": 1, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3078 + ] + }, + { + "test_f1": 0.691017734774832, + "test_acc": 0.693541395752059, + "test_auc": 0.727295308298298, + "val_auc": 0.7424, + "k": 2, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3078, + 52 + ] + }, + { + "test_f1": 0.7157953205233482, + "test_acc": 0.717815344603381, + "test_auc": 0.7517413694363735, + "val_auc": 0.8202, + "k": 5, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3078, + 52, + 3274, + 817, + 2046 + ] + }, + { + "test_f1": 0.7186714809117053, + "test_acc": 0.7199826614651061, + "test_auc": 0.7964961422316286, + "val_auc": 0.8548, + "k": 10, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3078, + 52, + 3274, + 817, + 2046, + 567, + 121, + 1524, + 2955, + 649 + ] + }, + { + "test_f1": 0.7316409862945544, + "test_acc": 0.7325530992631123, + "test_auc": 0.8327958411558424, + "val_auc": 0.8424, + "k": 16, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3078, + 52, + 3274, + 817, + 2046, + 567, + 121, + 1524, + 2955, + 649, + 9923, + 698, + 848, + 1602, + 1158, + 136 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3c7475f447743d13c2d425f235d5751ed7d35ed8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5846507328420428, + "test_acc": 0.6055483311660165, + "test_auc": 0.6124156559408731, + "val_auc": 0.6452, + "k": 1, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.5915972488473428, + "test_acc": 0.6085825747724317, + "test_auc": 0.627384518722164, + "val_auc": 0.6759999999999999, + "k": 2, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 778 + ] + }, + { + "test_f1": 0.594812494069537, + "test_acc": 0.6146510619852622, + "test_auc": 0.6873110760716148, + "val_auc": 0.7817999999999998, + "k": 5, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 778, + 581, + 121, + 52 + ] + }, + { + "test_f1": 0.6359375483184297, + "test_acc": 0.6367576939748592, + "test_auc": 0.7182664167472091, + "val_auc": 0.8011999999999999, + "k": 10, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 778, + 581, + 121, + 52, + 835, + 878, + 1253, + 925, + 51 + ] + }, + { + "test_f1": 0.647934735800552, + "test_acc": 0.6480277416558301, + "test_auc": 0.7335982840333634, + "val_auc": 0.8056, + "k": 16, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 778, + 581, + 121, + 52, + 835, + 878, + 1253, + 925, + 51, + 3078, + 4654, + 55, + 1731, + 628, + 7595 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..152117516a121beed2155478e411b0dfaa59a2bb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6225292059005385, + "test_acc": 0.6226012793176973, + "test_auc": 0.6516275686488453, + "val_auc": 0.6586, + "k": 1, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048 + ] + }, + { + "test_f1": 0.6246993504138595, + "test_acc": 0.6247334754797441, + "test_auc": 0.6918076013820695, + "val_auc": 0.6902, + "k": 2, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 1504 + ] + }, + { + "test_f1": 0.657399889008967, + "test_acc": 0.6588486140724946, + "test_auc": 0.7419985451900346, + "val_auc": 0.768, + "k": 5, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 1504, + 499, + 1818, + 7129 + ] + }, + { + "test_f1": 0.6698554080595486, + "test_acc": 0.6716417910447762, + "test_auc": 0.759338061465721, + "val_auc": 0.7708, + "k": 10, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 1504, + 499, + 1818, + 7129, + 1927, + 437, + 1974, + 6905, + 927 + ] + }, + { + "test_f1": 0.736698306881818, + "test_acc": 0.7377398720682303, + "test_auc": 0.8095653755228223, + "val_auc": 0.8024, + "k": 16, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 1504, + 499, + 1818, + 7129, + 1927, + 437, + 1974, + 6905, + 927, + 3636, + 3370, + 1258, + 1371, + 1761, + 4672 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f6f0f9a80996a83ddf5c85dd4b8d2c67a0110ee2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9090352334202494, + "test_acc": 0.9090909090909091, + "test_auc": 0.9616326530612245, + "val_auc": 0.9545088566827699, + "k": 1, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299 + ] + }, + { + "test_f1": 0.9090909090909091, + "test_acc": 0.9090909090909091, + "test_auc": 0.9893877551020408, + "val_auc": 0.9667874396135266, + "k": 2, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 25 + ] + }, + { + "test_f1": 0.9898969278371137, + "test_acc": 0.98989898989899, + "test_auc": 1.0, + "val_auc": 0.9907407407407408, + "k": 5, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 25, + 114, + 23633, + 4580 + ] + }, + { + "test_f1": 1.0, + "test_acc": 1.0, + "test_auc": 1.0, + "val_auc": 0.9955716586151369, + "k": 10, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 25, + 114, + 23633, + 4580, + 1616, + 5997, + 1849, + 6877, + 5027 + ] + }, + { + "test_f1": 1.0, + "test_acc": 1.0, + "test_auc": 1.0, + "val_auc": 1.0, + "k": 16, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 25, + 114, + 23633, + 4580, + 1616, + 5997, + 1849, + 6877, + 5027, + 3172, + 3959, + 2424, + 6662, + 561, + 1585 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2822c7d4c85c647efa125137f07e5a72b4f22069 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7844521637625086, + "test_acc": 0.7878787878787878, + "test_auc": 0.8038775510204081, + "val_auc": 0.7975040257648953, + "k": 1, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429 + ] + }, + { + "test_f1": 0.8179590017825311, + "test_acc": 0.8181818181818182, + "test_auc": 0.8648979591836733, + "val_auc": 0.8743961352657004, + "k": 2, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429, + 1585 + ] + }, + { + "test_f1": 0.9292929292929293, + "test_acc": 0.9292929292929293, + "test_auc": 0.9840816326530613, + "val_auc": 0.9690016103059582, + "k": 5, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429, + 1585, + 1411, + 114, + 385 + ] + }, + { + "test_f1": 0.9696969696969697, + "test_acc": 0.9696969696969697, + "test_auc": 0.9922448979591836, + "val_auc": 0.9887278582930759, + "k": 10, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429, + 1585, + 1411, + 114, + 385, + 1360, + 23791, + 278, + 16, + 1865 + ] + }, + { + "test_f1": 0.9595877138734282, + "test_acc": 0.9595959595959596, + "test_auc": 0.9967346938775511, + "val_auc": 0.9847020933977455, + "k": 16, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 429, + 1585, + 1411, + 114, + 385, + 1360, + 23791, + 278, + 16, + 1865, + 1299, + 5387, + 6662, + 1396, + 1960, + 7164 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a203ef6a79a74d39c5a335debfb2a8afad68d616 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9191424044365222, + "test_acc": 0.9191919191919192, + "test_auc": 0.9555102040816327, + "val_auc": 0.9657809983896941, + "k": 1, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9595132999388319, + "test_acc": 0.9595959595959596, + "test_auc": 0.9959183673469387, + "val_auc": 0.9883252818035426, + "k": 2, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 6662 + ] + }, + { + "test_f1": 0.9595959595959596, + "test_acc": 0.9595959595959596, + "test_auc": 0.9975510204081632, + "val_auc": 0.9895330112721418, + "k": 5, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 6662, + 1347, + 1360, + 1724 + ] + }, + { + "test_f1": 0.9696907835113407, + "test_acc": 0.9696969696969697, + "test_auc": 0.9971428571428571, + "val_auc": 0.9911433172302738, + "k": 10, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 6662, + 1347, + 1360, + 1724, + 2592, + 177, + 1616, + 114, + 322 + ] + }, + { + "test_f1": 0.9797979797979798, + "test_acc": 0.9797979797979798, + "test_auc": 0.9987755102040816, + "val_auc": 0.9935587761674718, + "k": 16, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 6662, + 1347, + 1360, + 1724, + 2592, + 177, + 1616, + 114, + 322, + 23882, + 419, + 1610, + 5387, + 5948, + 17948 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..93c8f4e2ec40fc8059a428335c874d11925ea03d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9288860795055486, + "test_acc": 0.9288888888888889, + "test_auc": 0.9465470922882427, + "val_auc": 0.9431818181818182, + "k": 1, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1369 + ] + }, + { + "test_f1": 0.937775319567355, + "test_acc": 0.9377777777777778, + "test_auc": 0.9587547408343869, + "val_auc": 0.9435876623376624, + "k": 2, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1369, + 1309 + ] + }, + { + "test_f1": 0.937775319567355, + "test_acc": 0.9377777777777778, + "test_auc": 0.9550410872313526, + "val_auc": 0.9618506493506492, + "k": 5, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1369, + 1309, + 15752, + 22580, + 4580 + ] + }, + { + "test_f1": 0.9422222222222222, + "test_acc": 0.9422222222222222, + "test_auc": 0.9636536030341339, + "val_auc": 0.9778814935064936, + "k": 10, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1369, + 1309, + 15752, + 22580, + 4580, + 1295, + 10260, + 1488, + 17948, + 1327 + ] + }, + { + "test_f1": 0.9422222222222222, + "test_acc": 0.9422222222222222, + "test_auc": 0.9636536030341338, + "val_auc": 0.9758522727272727, + "k": 16, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1369, + 1309, + 15752, + 22580, + 4580, + 1295, + 10260, + 1488, + 17948, + 1327, + 1360, + 4850, + 7164, + 505, + 4332, + 31863 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6d2ebf1882bbb16597cde8c8ed87dea5e2fdf8ba --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7769834756222191, + "test_acc": 0.7822222222222223, + "test_auc": 0.7911662452591656, + "val_auc": 0.749594155844156, + "k": 1, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1570 + ] + }, + { + "test_f1": 0.8348707643970291, + "test_acc": 0.8355555555555556, + "test_auc": 0.8910398230088494, + "val_auc": 0.8782467532467533, + "k": 2, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1570, + 406 + ] + }, + { + "test_f1": 0.8436044215908816, + "test_acc": 0.8444444444444444, + "test_auc": 0.9196823640960808, + "val_auc": 0.8904220779220778, + "k": 5, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1570, + 406, + 4580, + 854, + 322 + ] + }, + { + "test_f1": 0.8437966690242167, + "test_acc": 0.8444444444444444, + "test_auc": 0.9195638432364097, + "val_auc": 0.8904220779220778, + "k": 10, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1570, + 406, + 4580, + 854, + 322, + 1182, + 1804, + 1585, + 6662, + 1209 + ] + }, + { + "test_f1": 0.8484939366738828, + "test_acc": 0.8488888888888889, + "test_auc": 0.9197218710493046, + "val_auc": 0.8871753246753248, + "k": 16, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1570, + 406, + 4580, + 854, + 322, + 1182, + 1804, + 1585, + 6662, + 1209, + 1232, + 5997, + 920, + 800, + 786, + 5826 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..16bab531700167e6088c9c44b25635d3c7937f6c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9422153739444982, + "test_acc": 0.9422222222222222, + "test_auc": 0.9650758533501895, + "val_auc": 0.952922077922078, + "k": 1, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9377088825617772, + "test_acc": 0.9377777777777778, + "test_auc": 0.9731352718078381, + "val_auc": 0.9598214285714286, + "k": 2, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16 + ] + }, + { + "test_f1": 0.9733333333333334, + "test_acc": 0.9733333333333334, + "test_auc": 0.9953381795195955, + "val_auc": 0.992086038961039, + "k": 5, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 114, + 1658, + 4580 + ] + }, + { + "test_f1": 0.9644402297033876, + "test_acc": 0.9644444444444444, + "test_auc": 0.9941529709228825, + "val_auc": 0.9941152597402598, + "k": 10, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 114, + 1658, + 4580, + 26719, + 1835, + 7095, + 23882, + 567 + ] + }, + { + "test_f1": 0.9644444444444444, + "test_acc": 0.9644444444444444, + "test_auc": 0.9954962073324904, + "val_auc": 0.9939123376623378, + "k": 16, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 114, + 1658, + 4580, + 26719, + 1835, + 7095, + 23882, + 567, + 1417, + 969, + 6662, + 1232, + 4114, + 1458 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9fe37c08baa7b01e738fdcc91fe8e9bfb46e4b34 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6537524905568688, + "test_acc": 0.6688339835283919, + "test_auc": 0.6753657477066081, + "val_auc": 0.7420000000000001, + "k": 1, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2041 + ] + }, + { + "test_f1": 0.6506511140975828, + "test_acc": 0.6653662765496315, + "test_auc": 0.6944073256263142, + "val_auc": 0.7512000000000001, + "k": 2, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2041, + 7246 + ] + }, + { + "test_f1": 0.6959216584678246, + "test_acc": 0.7009102730819246, + "test_auc": 0.751419325067152, + "val_auc": 0.8064, + "k": 5, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2041, + 7246, + 11960, + 32305, + 5394 + ] + }, + { + "test_f1": 0.7406410809093305, + "test_acc": 0.7433896835717382, + "test_auc": 0.818757036500366, + "val_auc": 0.8504, + "k": 10, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2041, + 7246, + 11960, + 32305, + 5394, + 821, + 8604, + 1560, + 1737, + 253 + ] + }, + { + "test_f1": 0.754773805359343, + "test_acc": 0.7563935847420893, + "test_auc": 0.833894249196956, + "val_auc": 0.8738, + "k": 16, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2041, + 7246, + 11960, + 32305, + 5394, + 821, + 8604, + 1560, + 1737, + 253, + 500, + 3792, + 411, + 24597, + 1257, + 1503 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..86f0e1ffa2c9e71b02bef86dd671b753d87b17c9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5205743197158866, + "test_acc": 0.5400953619419159, + "test_auc": 0.5523109783685389, + "val_auc": 0.5864, + "k": 1, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253 + ] + }, + { + "test_f1": 0.6054001597054444, + "test_acc": 0.6103164282618119, + "test_auc": 0.6267697409064742, + "val_auc": 0.6502, + "k": 2, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 2041 + ] + }, + { + "test_f1": 0.5995887048208072, + "test_acc": 0.6059817945383615, + "test_auc": 0.6472573243486587, + "val_auc": 0.6842, + "k": 5, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 2041, + 26450, + 11960, + 1551 + ] + }, + { + "test_f1": 0.6417737536581886, + "test_acc": 0.6419592544429996, + "test_auc": 0.7084295959151095, + "val_auc": 0.7284, + "k": 10, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 2041, + 26450, + 11960, + 1551, + 1411, + 3802, + 3245, + 411, + 5394 + ] + }, + { + "test_f1": 0.6662328270240331, + "test_acc": 0.6662332032943217, + "test_auc": 0.7395258544885546, + "val_auc": 0.7627999999999999, + "k": 16, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 2041, + 26450, + 11960, + 1551, + 1411, + 3802, + 3245, + 411, + 5394, + 8604, + 1099, + 1758, + 6718, + 1417, + 821 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3098245057e7363b0027e6c018dcbd423fc8c25e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6221357612966897, + "test_acc": 0.6341569137407889, + "test_auc": 0.657605583204691, + "val_auc": 0.6951999999999999, + "k": 1, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253 + ] + }, + { + "test_f1": 0.6932581420405262, + "test_acc": 0.6987429562201994, + "test_auc": 0.7363144295418027, + "val_auc": 0.7192, + "k": 2, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 1099 + ] + }, + { + "test_f1": 0.7081383050907799, + "test_acc": 0.7121803207628955, + "test_auc": 0.792813863615525, + "val_auc": 0.7832, + "k": 5, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 1099, + 5394, + 46, + 2041 + ] + }, + { + "test_f1": 0.7329946700097439, + "test_acc": 0.7342869527524925, + "test_auc": 0.8264090662441885, + "val_auc": 0.8328, + "k": 10, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 1099, + 5394, + 46, + 2041, + 844, + 821, + 4481, + 5653, + 1503 + ] + }, + { + "test_f1": 0.7546139946666425, + "test_acc": 0.7546597312527091, + "test_auc": 0.8422388434360819, + "val_auc": 0.8744, + "k": 16, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 1099, + 5394, + 46, + 2041, + 844, + 821, + 4481, + 5653, + 1503, + 7246, + 9390, + 1737, + 4675, + 11960, + 32305 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c28f752976ecc8f83c0f508065455f9a632ec02c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7847279221749137, + "test_acc": 0.7911864406779661, + "test_auc": 0.7974502947200436, + "val_auc": 0.7429549114331724, + "k": 1, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7873035952587122, + "test_acc": 0.792542372881356, + "test_auc": 0.8206031924634037, + "val_auc": 0.8631239935587762, + "k": 2, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 20312 + ] + }, + { + "test_f1": 0.8124623863351316, + "test_acc": 0.8142372881355933, + "test_auc": 0.8695730512257632, + "val_auc": 0.9074074074074074, + "k": 5, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 20312, + 776, + 508, + 431 + ] + }, + { + "test_f1": 0.8478970571215695, + "test_acc": 0.848813559322034, + "test_auc": 0.9246046191805202, + "val_auc": 0.9331723027375202, + "k": 10, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 20312, + 776, + 508, + 431, + 1697, + 1433, + 1528, + 5406, + 737 + ] + }, + { + "test_f1": 0.8676289314815075, + "test_acc": 0.8677966101694915, + "test_auc": 0.9363382643324397, + "val_auc": 0.9444444444444444, + "k": 16, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 20312, + 776, + 508, + 431, + 1697, + 1433, + 1528, + 5406, + 737, + 27673, + 4413, + 4674, + 1628, + 1417, + 6362 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8cea93d710b60e1e370988765036bf5b8a586cc7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.782485339938211, + "test_acc": 0.7877966101694915, + "test_auc": 0.797759171621567, + "val_auc": 0.7105475040257649, + "k": 1, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528 + ] + }, + { + "test_f1": 0.8156567482138543, + "test_acc": 0.8176271186440678, + "test_auc": 0.8338757064639846, + "val_auc": 0.7870370370370371, + "k": 2, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 6362 + ] + }, + { + "test_f1": 0.83435801636796, + "test_acc": 0.8352542372881356, + "test_auc": 0.8757827639334739, + "val_auc": 0.8268921095008052, + "k": 5, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 6362, + 5020, + 1789, + 737 + ] + }, + { + "test_f1": 0.8261071301944475, + "test_acc": 0.8271186440677966, + "test_auc": 0.9007163002430567, + "val_auc": 0.8470209339774557, + "k": 10, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 6362, + 5020, + 1789, + 737, + 203, + 2023, + 1628, + 19225, + 1815 + ] + }, + { + "test_f1": 0.8259350529060596, + "test_acc": 0.8264406779661017, + "test_auc": 0.914285740550757, + "val_auc": 0.8502415458937197, + "k": 16, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 6362, + 5020, + 1789, + 737, + 203, + 2023, + 1628, + 19225, + 1815, + 152, + 1866, + 32278, + 1544, + 25949, + 461 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d8bde2d9ead9b308caf2064f2f1c5fffff4075f9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6197638979929457, + "test_acc": 0.6393220338983051, + "test_auc": 0.6436470640147379, + "val_auc": 0.6326489533011271, + "k": 1, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20312 + ] + }, + { + "test_f1": 0.6080652533158046, + "test_acc": 0.6501694915254237, + "test_auc": 0.7376945648696651, + "val_auc": 0.7546296296296295, + "k": 2, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20312, + 895 + ] + }, + { + "test_f1": 0.6982880387409202, + "test_acc": 0.6996610169491525, + "test_auc": 0.7712279327677944, + "val_auc": 0.8202495974235104, + "k": 5, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20312, + 895, + 431, + 776, + 5406 + ] + }, + { + "test_f1": 0.7247404488648804, + "test_acc": 0.7247457627118644, + "test_auc": 0.8250350244343692, + "val_auc": 0.8864734299516909, + "k": 10, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20312, + 895, + 431, + 776, + 5406, + 4961, + 27673, + 1866, + 2893, + 27 + ] + }, + { + "test_f1": 0.8041810407637319, + "test_acc": 0.8054237288135593, + "test_auc": 0.8913286487003269, + "val_auc": 0.9359903381642513, + "k": 16, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20312, + 895, + 431, + 776, + 5406, + 4961, + 27673, + 1866, + 2893, + 27, + 19225, + 1628, + 3927, + 508, + 373, + 737 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ee87806cfaf5385c3e200a697412d0c159598b4e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5557279105666203, + "test_acc": 0.5714285714285714, + "test_auc": 0.5870393036855706, + "val_auc": 0.5242096838735494, + "k": 1, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639 + ] + }, + { + "test_f1": 0.6135844299109606, + "test_acc": 0.6209912536443148, + "test_auc": 0.6273289813681492, + "val_auc": 0.5422168867547019, + "k": 2, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639, + 219 + ] + }, + { + "test_f1": 0.6056099036021034, + "test_acc": 0.6122448979591837, + "test_auc": 0.6328199374405005, + "val_auc": 0.5722288915566226, + "k": 5, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639, + 219, + 4381, + 1508, + 993 + ] + }, + { + "test_f1": 0.6010243747644831, + "test_acc": 0.60932944606414, + "test_auc": 0.6737216102271182, + "val_auc": 0.6906762705082032, + "k": 10, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639, + 219, + 4381, + 1508, + 993, + 275, + 1799, + 799, + 6907, + 1560 + ] + }, + { + "test_f1": 0.6451132140608362, + "test_acc": 0.6472303206997084, + "test_auc": 0.7083503331973344, + "val_auc": 0.7178871548619448, + "k": 16, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639, + 219, + 4381, + 1508, + 993, + 275, + 1799, + 799, + 6907, + 1560, + 30192, + 801, + 1090, + 786, + 1635, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1fe440c9330bbbb303d177f8a1b41b7caf3acec6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7249878473556911, + "test_acc": 0.7405247813411079, + "test_auc": 0.744390044879641, + "val_auc": 0.7078831532613046, + "k": 1, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7782368651756407, + "test_acc": 0.7842565597667639, + "test_auc": 0.7970386236910104, + "val_auc": 0.8037214885954382, + "k": 2, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275 + ] + }, + { + "test_f1": 0.8147642233742373, + "test_acc": 0.8163265306122449, + "test_auc": 0.8586971304229566, + "val_auc": 0.8815526210484194, + "k": 5, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 191, + 534, + 765 + ] + }, + { + "test_f1": 0.8035792765612676, + "test_acc": 0.8046647230320699, + "test_auc": 0.8652930776553788, + "val_auc": 0.9103641456582633, + "k": 10, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 191, + 534, + 765, + 1928, + 152, + 799, + 786, + 1508 + ] + }, + { + "test_f1": 0.88022209617646, + "test_acc": 0.880466472303207, + "test_auc": 0.9417584659322725, + "val_auc": 0.9281712685074031, + "k": 16, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 191, + 534, + 765, + 1928, + 152, + 799, + 786, + 1508, + 26639, + 801, + 3233, + 506, + 1799, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a3e8a3cbeb9ee178c8922855623d30f461ad9035 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.687098401875405, + "test_acc": 0.6938775510204082, + "test_auc": 0.6923364613083096, + "val_auc": 0.6968787515006002, + "k": 1, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7149484452542242, + "test_acc": 0.717201166180758, + "test_auc": 0.7469230246158031, + "val_auc": 0.7681072428971589, + "k": 2, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191 + ] + }, + { + "test_f1": 0.7276920154092327, + "test_acc": 0.7288629737609329, + "test_auc": 0.8033795729634162, + "val_auc": 0.792717086834734, + "k": 5, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 534, + 4003, + 275 + ] + }, + { + "test_f1": 0.7929677217418736, + "test_acc": 0.793002915451895, + "test_auc": 0.8400652794777642, + "val_auc": 0.8539415766306523, + "k": 10, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 534, + 4003, + 275, + 1239, + 142, + 152, + 5897, + 1195 + ] + }, + { + "test_f1": 0.7813373906055017, + "test_acc": 0.7813411078717201, + "test_auc": 0.8594281245750035, + "val_auc": 0.8763505402160865, + "k": 16, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 534, + 4003, + 275, + 1239, + 142, + 152, + 5897, + 1195, + 1596, + 8006, + 743, + 1046, + 501, + 1837 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e0cbfc0baf12a5c71c2208a589dff7a3f5b130de --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320 + ] + }, + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 2, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 113 + ] + }, + { + "test_f1": 0.49828511897477407, + "test_acc": 0.5050505050505051, + "test_auc": 0.47632653061224484, + "val_auc": 0.547070707070707, + "k": 5, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 113, + 220, + 987, + 1408 + ] + }, + { + "test_f1": 0.4842167200879516, + "test_acc": 0.48484848484848486, + "test_auc": 0.4848979591836734, + "val_auc": 0.5688888888888889, + "k": 10, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 113, + 220, + 987, + 1408, + 13612, + 2779, + 100, + 5883, + 152 + ] + }, + { + "test_f1": 0.47085671530115975, + "test_acc": 0.47474747474747475, + "test_auc": 0.5155102040816326, + "val_auc": 0.5531313131313131, + "k": 16, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 113, + 220, + 987, + 1408, + 13612, + 2779, + 100, + 5883, + 152, + 4712, + 1928, + 346, + 859, + 1149, + 7926 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..93772db59ea46acc79c44f8ca430a43e5056b2d0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5653905596745241, + "test_acc": 0.5656565656565656, + "test_auc": 0.5508163265306123, + "val_auc": 0.5822222222222222, + "k": 1, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1302 + ] + }, + { + "test_f1": 0.5653905596745241, + "test_acc": 0.5656565656565656, + "test_auc": 0.5661224489795919, + "val_auc": 0.581010101010101, + "k": 2, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1302, + 2779 + ] + }, + { + "test_f1": 0.5651239012506258, + "test_acc": 0.5656565656565656, + "test_auc": 0.5795918367346939, + "val_auc": 0.6290909090909091, + "k": 5, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1302, + 2779, + 1116, + 4712, + 9291 + ] + }, + { + "test_f1": 0.5757575757575758, + "test_acc": 0.5757575757575758, + "test_auc": 0.6065306122448979, + "val_auc": 0.6222222222222221, + "k": 10, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1302, + 2779, + 1116, + 4712, + 9291, + 1030, + 5551, + 859, + 7026, + 7666 + ] + }, + { + "test_f1": 0.5555555555555556, + "test_acc": 0.5555555555555556, + "test_auc": 0.5546938775510204, + "val_auc": 0.6266666666666667, + "k": 16, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1302, + 2779, + 1116, + 4712, + 9291, + 1030, + 5551, + 859, + 7026, + 7666, + 3057, + 13612, + 1417, + 1152, + 1366, + 1387 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cf0b62aceb742c6199611230c825bc581d8d9ae8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5514403292181069, + "test_acc": 0.5555555555555556, + "test_auc": 0.5942857142857143, + "val_auc": 0.5923232323232324, + "k": 1, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666 + ] + }, + { + "test_f1": 0.5310512532734755, + "test_acc": 0.5353535353535354, + "test_auc": 0.5738775510204082, + "val_auc": 0.6032323232323232, + "k": 2, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 3802 + ] + }, + { + "test_f1": 0.60581934482108, + "test_acc": 0.6060606060606061, + "test_auc": 0.5851020408163264, + "val_auc": 0.616969696969697, + "k": 5, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 3802, + 1030, + 4985, + 111 + ] + }, + { + "test_f1": 0.5656565656565656, + "test_acc": 0.5656565656565656, + "test_auc": 0.6244897959183673, + "val_auc": 0.6492929292929293, + "k": 10, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 3802, + 1030, + 4985, + 111, + 2682, + 1635, + 7026, + 1960, + 6804 + ] + }, + { + "test_f1": 0.49433056050703106, + "test_acc": 0.494949494949495, + "test_auc": 0.5661224489795919, + "val_auc": 0.6351515151515151, + "k": 16, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 3802, + 1030, + 4985, + 111, + 2682, + 1635, + 7026, + 1960, + 6804, + 1116, + 2595, + 5883, + 25437, + 1928, + 1417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7a5a96fb45593bc5fbd97de356021094435d65ab --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.49356750779793046, + "test_acc": 0.551864406779661, + "test_auc": 0.5534770714057209, + "val_auc": 0.590780998389694, + "k": 1, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217 + ] + }, + { + "test_f1": 0.49200685252783755, + "test_acc": 0.5511864406779661, + "test_auc": 0.605676348486687, + "val_auc": 0.5885668276972625, + "k": 2, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 20 + ] + }, + { + "test_f1": 0.586075104078657, + "test_acc": 0.5898305084745763, + "test_auc": 0.6271745485433144, + "val_auc": 0.610708534621578, + "k": 5, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 20, + 311, + 1541, + 186 + ] + }, + { + "test_f1": 0.5920943201103296, + "test_acc": 0.5932203389830508, + "test_auc": 0.6238467676399967, + "val_auc": 0.607085346215781, + "k": 10, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 20, + 311, + 1541, + 186, + 1812, + 681, + 392, + 26270, + 100 + ] + }, + { + "test_f1": 0.5829987136957268, + "test_acc": 0.5830508474576271, + "test_auc": 0.6123429416112343, + "val_auc": 0.6386876006441223, + "k": 16, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 20, + 311, + 1541, + 186, + 1812, + 681, + 392, + 26270, + 100, + 152, + 1715, + 373, + 1398, + 253, + 443 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..442967f2ac462b754e62508d260fe088ca49f15e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5801154707434467, + "test_acc": 0.5871186440677966, + "test_auc": 0.6089949366250785, + "val_auc": 0.6380837359098229, + "k": 1, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20 + ] + }, + { + "test_f1": 0.576322127558349, + "test_acc": 0.6176271186440678, + "test_auc": 0.6833239567131086, + "val_auc": 0.6948470209339775, + "k": 2, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20, + 1541 + ] + }, + { + "test_f1": 0.655550156350676, + "test_acc": 0.655593220338983, + "test_auc": 0.7158001566447143, + "val_auc": 0.6688808373590982, + "k": 5, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20, + 1541, + 100, + 3063, + 152 + ] + }, + { + "test_f1": 0.6589501315661012, + "test_acc": 0.6589830508474577, + "test_auc": 0.7189091497427864, + "val_auc": 0.7556360708534622, + "k": 10, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20, + 1541, + 100, + 3063, + 152, + 186, + 1874, + 1096, + 908, + 253 + ] + }, + { + "test_f1": 0.6622121035179097, + "test_acc": 0.6630508474576271, + "test_auc": 0.7228169941129533, + "val_auc": 0.7801932367149759, + "k": 16, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20, + 1541, + 100, + 3063, + 152, + 186, + 1874, + 1096, + 908, + 253, + 207, + 373, + 311, + 1319, + 1335, + 17442 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..71511b201d73cd2b74f370fb66ae74003b530a35 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5718240839127793, + "test_acc": 0.5898305084745763, + "test_auc": 0.5949998713012911, + "val_auc": 0.5418679549114331, + "k": 1, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.5417469758980404, + "test_acc": 0.5491525423728814, + "test_auc": 0.5991715480248425, + "val_auc": 0.6082930756843801, + "k": 2, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 521 + ] + }, + { + "test_f1": 0.5874040535735451, + "test_acc": 0.5877966101694915, + "test_auc": 0.6093516159042188, + "val_auc": 0.6481481481481481, + "k": 5, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 521, + 20, + 1255, + 4511 + ] + }, + { + "test_f1": 0.5884734413526771, + "test_acc": 0.5884745762711865, + "test_auc": 0.6303432946134074, + "val_auc": 0.6763285024154589, + "k": 10, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 521, + 20, + 1255, + 4511, + 12485, + 186, + 1596, + 17442, + 1562 + ] + }, + { + "test_f1": 0.5979583403835055, + "test_acc": 0.5979661016949153, + "test_auc": 0.6408515442006524, + "val_auc": 0.676731078904992, + "k": 16, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 521, + 20, + 1255, + 4511, + 12485, + 186, + 1596, + 17442, + 1562, + 1942, + 253, + 4012, + 1335, + 2277, + 10 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fffb374ec0502b2fc3a6ab7c246f9c85eb080fe3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5276536090735036, + "test_acc": 0.5359801488833746, + "test_auc": 0.5533717550859564, + "val_auc": 0.5570228091236493, + "k": 1, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1101 + ] + }, + { + "test_f1": 0.5285238017796158, + "test_acc": 0.5384615384615384, + "test_auc": 0.5586670607359243, + "val_auc": 0.556422569027611, + "k": 2, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1101, + 635 + ] + }, + { + "test_f1": 0.563194750573771, + "test_acc": 0.5632754342431762, + "test_auc": 0.6101177281907295, + "val_auc": 0.6018407362945178, + "k": 5, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1101, + 635, + 456, + 124, + 1268 + ] + }, + { + "test_f1": 0.5630817266342948, + "test_acc": 0.5632754342431762, + "test_auc": 0.6197601103393922, + "val_auc": 0.6730692276910765, + "k": 10, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1101, + 635, + 456, + 124, + 1268, + 1120, + 1067, + 1441, + 349, + 1770 + ] + }, + { + "test_f1": 0.5527715419021669, + "test_acc": 0.5533498759305211, + "test_auc": 0.6166937589281316, + "val_auc": 0.6434573829531813, + "k": 16, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1101, + 635, + 456, + 124, + 1268, + 1120, + 1067, + 1441, + 349, + 1770, + 1070, + 1090, + 2536, + 142, + 1619, + 1551 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..70b8e9fbb436b8c799cd3a6ae16990475c7e41aa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5635911372884078, + "test_acc": 0.5707196029776674, + "test_auc": 0.5779764543618541, + "val_auc": 0.5950380152060823, + "k": 1, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5645 + ] + }, + { + "test_f1": 0.6062381593761779, + "test_acc": 0.607940446650124, + "test_auc": 0.626964188956209, + "val_auc": 0.615046018407363, + "k": 2, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5645, + 1625 + ] + }, + { + "test_f1": 0.6076791107889332, + "test_acc": 0.6129032258064516, + "test_auc": 0.6640929018274961, + "val_auc": 0.6586634653861544, + "k": 5, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5645, + 1625, + 1653, + 2655, + 913 + ] + }, + { + "test_f1": 0.6109577750943974, + "test_acc": 0.6129032258064516, + "test_auc": 0.6597458253288015, + "val_auc": 0.6734693877551021, + "k": 10, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5645, + 1625, + 1653, + 2655, + 913, + 1619, + 1799, + 266, + 680, + 326 + ] + }, + { + "test_f1": 0.6062381593761779, + "test_acc": 0.607940446650124, + "test_auc": 0.6705334712575737, + "val_auc": 0.7102841136454582, + "k": 16, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5645, + 1625, + 1653, + 2655, + 913, + 1619, + 1799, + 266, + 680, + 326, + 575, + 1551, + 821, + 124, + 1962, + 119 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a82400a4088d94635e55120361794c10df7818 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5526004999513273, + "test_acc": 0.5533498759305211, + "test_auc": 0.5731491059553716, + "val_auc": 0.5418167266906763, + "k": 1, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619 + ] + }, + { + "test_f1": 0.5794486694855743, + "test_acc": 0.5806451612903226, + "test_auc": 0.6301660016747943, + "val_auc": 0.70328131252501, + "k": 2, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 25162 + ] + }, + { + "test_f1": 0.6441488741801011, + "test_acc": 0.6451612903225806, + "test_auc": 0.698069060637407, + "val_auc": 0.6962785114045619, + "k": 5, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 25162, + 1068, + 2536, + 635 + ] + }, + { + "test_f1": 0.6600454413395179, + "test_acc": 0.6600496277915633, + "test_auc": 0.7554061376286884, + "val_auc": 0.6782713085234093, + "k": 10, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 25162, + 1068, + 2536, + 635, + 772, + 1397, + 1094, + 1662, + 549 + ] + }, + { + "test_f1": 0.717086747413905, + "test_acc": 0.71712158808933, + "test_auc": 0.7878183340722132, + "val_auc": 0.7066826730692277, + "k": 16, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 25162, + 1068, + 2536, + 635, + 772, + 1397, + 1094, + 1662, + 549, + 1198, + 296, + 2655, + 349, + 394, + 119 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..964054e72eee4743af605b86ec57856a7e2f4f2b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8315209192760887, + "test_acc": 0.8353413654618473, + "test_auc": 0.8391994380008889, + "val_auc": 0.8517171717171718, + "k": 1, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731 + ] + }, + { + "test_f1": 0.8315209192760887, + "test_acc": 0.8353413654618473, + "test_auc": 0.8867794010121718, + "val_auc": 0.9086868686868688, + "k": 2, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226 + ] + }, + { + "test_f1": 0.854858709488814, + "test_acc": 0.856760374832664, + "test_auc": 0.9529110693753495, + "val_auc": 0.9624242424242424, + "k": 5, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4529, + 22601, + 1193 + ] + }, + { + "test_f1": 0.8495329847404812, + "test_acc": 0.8500669344042838, + "test_auc": 0.9477570214047111, + "val_auc": 0.9668686868686869, + "k": 10, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4529, + 22601, + 1193, + 805, + 23003, + 5964, + 830, + 21301 + ] + }, + { + "test_f1": 0.8589594976540055, + "test_acc": 0.8594377510040161, + "test_auc": 0.954083095582859, + "val_auc": 0.974949494949495, + "k": 16, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4529, + 22601, + 1193, + 805, + 23003, + 5964, + 830, + 21301, + 4148, + 3734, + 27163, + 1629, + 13701, + 5973 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..27d974377c4c890febeb93a15f149f7d0e906063 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8685441324813237, + "test_acc": 0.8701472556894244, + "test_auc": 0.8762813436366503, + "val_auc": 0.8777777777777778, + "k": 1, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226 + ] + }, + { + "test_f1": 0.8685441324813237, + "test_acc": 0.8701472556894244, + "test_auc": 0.9036214534558644, + "val_auc": 0.9030303030303031, + "k": 2, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4505 + ] + }, + { + "test_f1": 0.8749750750782979, + "test_acc": 0.8755020080321285, + "test_auc": 0.964903728978796, + "val_auc": 0.9696969696969697, + "k": 5, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4505, + 4731, + 1425, + 22601 + ] + }, + { + "test_f1": 0.8712852731893603, + "test_acc": 0.8714859437751004, + "test_auc": 0.9652083841091884, + "val_auc": 0.9701010101010101, + "k": 10, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4505, + 4731, + 1425, + 22601, + 1529, + 1447, + 353, + 1682, + 6352 + ] + }, + { + "test_f1": 0.8793987746926573, + "test_acc": 0.8795180722891566, + "test_auc": 0.9671259193416584, + "val_auc": 0.9741414141414142, + "k": 16, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4505, + 4731, + 1425, + 22601, + 1529, + 1447, + 353, + 1682, + 6352, + 1193, + 5402, + 1065, + 152, + 1156, + 27163 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..22ef2f2cb9e6efbdac92f8ba6decdf6b03c9d8eb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8709305616326836, + "test_acc": 0.8728246318607764, + "test_auc": 0.87493010852891, + "val_auc": 0.863030303030303, + "k": 1, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22601 + ] + }, + { + "test_f1": 0.8709305616326836, + "test_acc": 0.8728246318607764, + "test_auc": 0.9148936932803832, + "val_auc": 0.8674747474747475, + "k": 2, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22601, + 4505 + ] + }, + { + "test_f1": 0.8900892977595135, + "test_acc": 0.8902275769745649, + "test_auc": 0.9710685151467362, + "val_auc": 0.9656565656565657, + "k": 5, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22601, + 4505, + 23226, + 4731, + 4529 + ] + }, + { + "test_f1": 0.8847278000892459, + "test_acc": 0.8848728246318608, + "test_auc": 0.9712799816490086, + "val_auc": 0.9638383838383838, + "k": 10, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22601, + 4505, + 23226, + 4731, + 4529, + 1529, + 1425, + 3417, + 31914, + 23 + ] + }, + { + "test_f1": 0.896864458290409, + "test_acc": 0.8969210174029452, + "test_auc": 0.9697710427090651, + "val_auc": 0.964040404040404, + "k": 16, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22601, + 4505, + 23226, + 4731, + 4529, + 1529, + 1425, + 3417, + 31914, + 23, + 2832, + 4648, + 152, + 1305, + 293, + 6369 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..190a855e4a5f711d34d338140e336ffa99098b45 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6546739134133612, + "test_acc": 0.6851282051282052, + "test_auc": 0.6857727135018683, + "val_auc": 0.6491228070175439, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 476 + ] + }, + { + "test_f1": 0.6555681767305613, + "test_acc": 0.6861538461538461, + "test_auc": 0.7909204901201736, + "val_auc": 0.7931456548347613, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 476, + 186 + ] + }, + { + "test_f1": 0.7648210920239902, + "test_acc": 0.7692307692307693, + "test_auc": 0.875904668933248, + "val_auc": 0.8608731130150958, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 476, + 186, + 1438, + 152, + 1335 + ] + }, + { + "test_f1": 0.7950726474613629, + "test_acc": 0.7979487179487179, + "test_auc": 0.8936151412125087, + "val_auc": 0.8637290901672786, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 476, + 186, + 1438, + 152, + 1335, + 253, + 2632, + 1562, + 8013, + 374 + ] + }, + { + "test_f1": 0.8137077475761264, + "test_acc": 0.8153846153846154, + "test_auc": 0.9054978624566601, + "val_auc": 0.8725010199918402, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 476, + 186, + 1438, + 152, + 1335, + 253, + 2632, + 1562, + 8013, + 374, + 21369, + 1431, + 1361, + 373, + 1030, + 549 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6a990a07f4713d33a7d54235bc255414db01f67e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8437433401635632, + "test_acc": 0.8460504201680672, + "test_auc": 0.854245305189781, + "val_auc": 0.8542753914090726, + "k": 1, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852 + ] + }, + { + "test_f1": 0.8720438748542193, + "test_acc": 0.8736134453781512, + "test_auc": 0.9040716225206268, + "val_auc": 0.9413890004014451, + "k": 2, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782 + ] + }, + { + "test_f1": 0.8963545660478147, + "test_acc": 0.8968067226890756, + "test_auc": 0.9434557382620705, + "val_auc": 0.9658771577679648, + "k": 5, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 60, + 173, + 15231 + ] + }, + { + "test_f1": 0.8950759572582099, + "test_acc": 0.8954621848739496, + "test_auc": 0.9647599988430194, + "val_auc": 0.9827378562826175, + "k": 10, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 60, + 173, + 15231, + 11077, + 6919, + 784, + 1493, + 1883 + ] + }, + { + "test_f1": 0.8983618942703705, + "test_acc": 0.8988235294117647, + "test_auc": 0.9677071356776652, + "val_auc": 0.9793255720594138, + "k": 16, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 60, + 173, + 15231, + 11077, + 6919, + 784, + 1493, + 1883, + 319, + 27312, + 325, + 1152, + 1021, + 23720 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ca8a86bf538d67c73562e91e427e066142ed678f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.701230323055188, + "test_acc": 0.706890756302521, + "test_auc": 0.6986607949902741, + "val_auc": 0.7450822962665595, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021 + ] + }, + { + "test_f1": 0.8010288261058609, + "test_acc": 0.8050420168067227, + "test_auc": 0.9054819185630301, + "val_auc": 0.9460056202328382, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 576 + ] + }, + { + "test_f1": 0.8115722060350141, + "test_acc": 0.8127731092436975, + "test_auc": 0.9150066707160986, + "val_auc": 0.9714973906061822, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 576, + 6404, + 1914, + 1960 + ] + }, + { + "test_f1": 0.8425580544204068, + "test_acc": 0.8433613445378151, + "test_auc": 0.9298228011945825, + "val_auc": 0.9735046166198315, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 576, + 6404, + 1914, + 1960, + 1794, + 1493, + 4790, + 1152, + 6919 + ] + }, + { + "test_f1": 0.8372732609612961, + "test_acc": 0.837983193277311, + "test_auc": 0.9253135146177265, + "val_auc": 0.9706945002007226, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 576, + 6404, + 1914, + 1960, + 1794, + 1493, + 4790, + 1152, + 6919, + 1067, + 1153, + 430, + 4235, + 32440, + 27312 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..51fc2384dc8443aab212fe3b6a6845a7fd7f81c0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8244259287705219, + "test_acc": 0.826890756302521, + "test_auc": 0.8172585345394857, + "val_auc": 0.8438378161380972, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 877 + ] + }, + { + "test_f1": 0.8215226007107449, + "test_acc": 0.824201680672269, + "test_auc": 0.7924977945057885, + "val_auc": 0.8113207547169812, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 877, + 4235 + ] + }, + { + "test_f1": 0.8193910801550013, + "test_acc": 0.8208403361344537, + "test_auc": 0.898705899154681, + "val_auc": 0.9026495383380169, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 877, + 4235, + 576, + 21169, + 430 + ] + }, + { + "test_f1": 0.8240470447309701, + "test_acc": 0.826218487394958, + "test_auc": 0.9255049135518582, + "val_auc": 0.9317543155359294, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 877, + 4235, + 576, + 21169, + 430, + 25737, + 42, + 5189, + 2046, + 964 + ] + }, + { + "test_f1": 0.8528913416091332, + "test_acc": 0.853109243697479, + "test_auc": 0.9380269684939729, + "val_auc": 0.9526294660778805, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 877, + 4235, + 576, + 21169, + 430, + 25737, + 42, + 5189, + 2046, + 964, + 0, + 2089, + 197, + 7046, + 1717, + 27312 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cc0a764d4ff4fb6eeecc3648b23524784e7ca319 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9387343589959853, + "test_acc": 0.9388235294117647, + "test_auc": 0.9599684722794687, + "val_auc": 0.9799277398635086, + "k": 1, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703 + ] + }, + { + "test_f1": 0.9731049712023268, + "test_acc": 0.973109243697479, + "test_auc": 0.9812786081523742, + "val_auc": 1.0, + "k": 2, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431 + ] + }, + { + "test_f1": 0.9909239390509728, + "test_acc": 0.9909243697478992, + "test_auc": 0.9945879522167025, + "val_auc": 1.0, + "k": 5, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431, + 786, + 1411, + 4148 + ] + }, + { + "test_f1": 0.9986554603618987, + "test_acc": 0.998655462184874, + "test_auc": 0.999976950777708, + "val_auc": 1.0, + "k": 10, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431, + 786, + 1411, + 4148, + 279, + 461, + 27673, + 1360, + 4674 + ] + }, + { + "test_f1": 0.9986554612733863, + "test_acc": 0.998655462184874, + "test_auc": 0.9999950285991135, + "val_auc": 1.0, + "k": 16, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431, + 786, + 1411, + 4148, + 279, + 461, + 27673, + 1360, + 4674, + 1634, + 1493, + 727, + 1914, + 4790, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..20445b035abd0796c1a989cea4f308c7ba0d0741 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9231234393342531, + "test_acc": 0.9233613445378152, + "test_auc": 0.944774063388073, + "val_auc": 0.9373745483741469, + "k": 1, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703 + ] + }, + { + "test_f1": 0.9528926358817142, + "test_acc": 0.9529411764705882, + "test_auc": 0.9643643205270046, + "val_auc": 0.9598554797270173, + "k": 2, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431 + ] + }, + { + "test_f1": 0.9717541551967765, + "test_acc": 0.971764705882353, + "test_auc": 0.978963743121389, + "val_auc": 0.9729024488157366, + "k": 5, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431, + 786, + 1411, + 279 + ] + }, + { + "test_f1": 0.9946218147104708, + "test_acc": 0.9946218487394958, + "test_auc": 0.999103340058283, + "val_auc": 1.0, + "k": 10, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431, + 786, + 1411, + 279, + 461, + 1360, + 737, + 1758, + 1493 + ] + }, + { + "test_f1": 0.9952941016959969, + "test_acc": 0.9952941176470588, + "test_auc": 0.999260843077279, + "val_auc": 1.0, + "k": 16, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431, + 786, + 1411, + 279, + 461, + 1360, + 737, + 1758, + 1493, + 727, + 1914, + 126, + 27673, + 4790, + 1634 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9f706d9f52f352e1ae9a0058cc132221b3b78756 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9190128709711732, + "test_acc": 0.9193277310924369, + "test_auc": 0.9391767179353682, + "val_auc": 0.969891609795263, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703 + ] + }, + { + "test_f1": 0.9528701414709411, + "test_acc": 0.9529411764705882, + "test_auc": 0.9590455994967133, + "val_auc": 0.969891609795263, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431 + ] + }, + { + "test_f1": 0.9942855786990665, + "test_acc": 0.9942857142857143, + "test_auc": 0.9970112841761214, + "val_auc": 1.0, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431, + 786, + 4148, + 461 + ] + }, + { + "test_f1": 0.9932771286223971, + "test_acc": 0.9932773109243698, + "test_auc": 0.9980611536542509, + "val_auc": 1.0, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431, + 786, + 4148, + 461, + 1411, + 279, + 27673, + 1360, + 1758 + ] + }, + { + "test_f1": 0.9952940878716621, + "test_acc": 0.9952941176470588, + "test_auc": 0.999517774114006, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1703, + 431, + 786, + 4148, + 461, + 1411, + 279, + 27673, + 1360, + 1758, + 1493, + 727, + 1914, + 1634, + 4790, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..094caa03890c102a122987bf18fb303d68e82052 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7582137446632694, + "test_acc": 0.7584905660377359, + "test_auc": 0.7914826654947299, + "val_auc": 0.8717486994797918, + "k": 1, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.7821098196186002, + "test_acc": 0.7821383647798742, + "test_auc": 0.8605174074137831, + "val_auc": 0.9095638255302121, + "k": 2, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 266 + ] + }, + { + "test_f1": 0.8168323928392031, + "test_acc": 0.8168553459119496, + "test_auc": 0.8964161162242708, + "val_auc": 0.9327731092436975, + "k": 5, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 266, + 3604, + 508, + 32700 + ] + }, + { + "test_f1": 0.8324213307777346, + "test_acc": 0.8324528301886792, + "test_auc": 0.9115938712293895, + "val_auc": 0.9315726290516206, + "k": 10, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 266, + 3604, + 508, + 32700, + 1327, + 1866, + 1411, + 1608, + 19666 + ] + }, + { + "test_f1": 0.8392092590397887, + "test_acc": 0.8392452830188679, + "test_auc": 0.9149177905885237, + "val_auc": 0.933173269307723, + "k": 16, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 266, + 3604, + 508, + 32700, + 1327, + 1866, + 1411, + 1608, + 19666, + 3739, + 1697, + 1847, + 1804, + 28635, + 27673 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4ba61ab469720e79d09781a67a4a5de8a248d988 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.716052286508916, + "test_acc": 0.716323296354992, + "test_auc": 0.7665712276471771, + "val_auc": 0.7861144457783112, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.7399932128381145, + "test_acc": 0.7400950871632329, + "test_auc": 0.8233875828812537, + "val_auc": 0.8767507002801121, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 468 + ] + }, + { + "test_f1": 0.7840883958372497, + "test_acc": 0.7844690966719493, + "test_auc": 0.8763612618042997, + "val_auc": 0.9179671868747499, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 468, + 508, + 1608, + 913 + ] + }, + { + "test_f1": 0.7951736619060152, + "test_acc": 0.7955625990491284, + "test_auc": 0.8824291742013262, + "val_auc": 0.9295718287314926, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 468, + 508, + 1608, + 913, + 542, + 272, + 32053, + 1866, + 3739 + ] + }, + { + "test_f1": 0.803279007298814, + "test_acc": 0.803486529318542, + "test_auc": 0.8899537874221418, + "val_auc": 0.9391756702681073, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 468, + 508, + 1608, + 913, + 542, + 272, + 32053, + 1866, + 3739, + 322, + 31567, + 19666, + 279, + 8946, + 4148 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9cc402ae2d06be34bb50281e2c83698c099324ee --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.700336723311322, + "test_acc": 0.7010804321728692, + "test_auc": 0.698262543811105, + "val_auc": 0.7322866344605475, + "k": 1, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6782713085234093, + "test_acc": 0.6782713085234093, + "test_auc": 0.7251285510053496, + "val_auc": 0.7330917874396136, + "k": 2, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447 + ] + }, + { + "test_f1": 0.7093556246027822, + "test_acc": 0.709483793517407, + "test_auc": 0.7798174921601181, + "val_auc": 0.7888486312399355, + "k": 5, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 1417, + 152 + ] + }, + { + "test_f1": 0.7020438912844473, + "test_acc": 0.7022809123649459, + "test_auc": 0.7867552112156428, + "val_auc": 0.7894524959742351, + "k": 10, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 1417, + 152, + 2837, + 31567, + 1866, + 776, + 6718 + ] + }, + { + "test_f1": 0.7189637031283101, + "test_acc": 0.7190876350540216, + "test_auc": 0.796042012543811, + "val_auc": 0.7697262479871175, + "k": 16, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 1417, + 152, + 2837, + 31567, + 1866, + 776, + 6718, + 800, + 1960, + 3184, + 4605, + 16922, + 32700 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1b4e56ab11ca5a5921c4c5bb21057f0a01987269 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.751420847618118, + "test_acc": 0.7520661157024794, + "test_auc": 0.7411814704632385, + "val_auc": 0.6866746698679471, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7327580310924489, + "test_acc": 0.7327823691460055, + "test_auc": 0.7624612956104668, + "val_auc": 0.7454981992797118, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 18582 + ] + }, + { + "test_f1": 0.7546869030598163, + "test_acc": 0.7548209366391184, + "test_auc": 0.7899490012749681, + "val_auc": 0.7543017206882754, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 18582, + 6501, + 29310, + 1447 + ] + }, + { + "test_f1": 0.7520435352029986, + "test_acc": 0.7520661157024794, + "test_auc": 0.8184384676097383, + "val_auc": 0.7899159663865546, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 18582, + 6501, + 29310, + 1447, + 508, + 31567, + 19688, + 3507, + 4413 + ] + }, + { + "test_f1": 0.7876273849607184, + "test_acc": 0.7878787878787878, + "test_auc": 0.8279400157853197, + "val_auc": 0.7983193277310925, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 18582, + 6501, + 29310, + 1447, + 508, + 31567, + 19688, + 3507, + 4413, + 1417, + 1377, + 6718, + 468, + 7229, + 6853 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5da067ccf885b00a40c3f408ef215f1f1d920731 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6209049476188817, + "test_acc": 0.6281761006289308, + "test_auc": 0.637645956260968, + "val_auc": 0.6612645058023209, + "k": 1, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6043894264180981, + "test_acc": 0.6153459119496856, + "test_auc": 0.6719359944265493, + "val_auc": 0.7290916366546618, + "k": 2, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 776 + ] + }, + { + "test_f1": 0.6580203568658082, + "test_acc": 0.6581132075471698, + "test_auc": 0.7130910779219858, + "val_auc": 0.7899159663865547, + "k": 5, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 776, + 27673, + 1411, + 1804 + ] + }, + { + "test_f1": 0.6600986811194132, + "test_acc": 0.660125786163522, + "test_auc": 0.7131250006328864, + "val_auc": 0.7871148459383754, + "k": 10, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 776, + 27673, + 1411, + 1804, + 266, + 6718, + 373, + 3739, + 31863 + ] + }, + { + "test_f1": 0.6638904368527105, + "test_acc": 0.6638993710691824, + "test_auc": 0.7203831949928052, + "val_auc": 0.7871148459383754, + "k": 16, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 776, + 27673, + 1411, + 1804, + 266, + 6718, + 373, + 3739, + 31863, + 7246, + 27502, + 1866, + 226, + 2033, + 3832 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fc0efe5eed483ac0afd427c04cba1e2dfc01b23f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7638371204793973, + "test_acc": 0.7642767295597485, + "test_auc": 0.8250886800420033, + "val_auc": 0.9119647859143658, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 409 + ] + }, + { + "test_f1": 0.7836644482878995, + "test_acc": 0.7841509433962264, + "test_auc": 0.8669619630212072, + "val_auc": 0.9563825530212084, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 409, + 343 + ] + }, + { + "test_f1": 0.8238506024189984, + "test_acc": 0.8244025157232704, + "test_auc": 0.9063105355839112, + "val_auc": 0.9603841536614646, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 409, + 343, + 1062, + 293, + 1441 + ] + }, + { + "test_f1": 0.8287335102602233, + "test_acc": 0.8291823899371069, + "test_auc": 0.9079028777597644, + "val_auc": 0.9623849539815927, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 409, + 343, + 1062, + 293, + 1441, + 1928, + 200, + 1313, + 61, + 334 + ] + }, + { + "test_f1": 0.8374108053893701, + "test_acc": 0.8374842767295597, + "test_auc": 0.9170723890398252, + "val_auc": 0.9623849539815927, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 409, + 343, + 1062, + 293, + 1441, + 1928, + 200, + 1313, + 61, + 334, + 765, + 1934, + 14, + 138, + 2022, + 3975 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7d3f69f271699296a4544600e8b8b7cbc99d44c4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.65188246097337, + "test_acc": 0.6565656565656566, + "test_auc": 0.6763265306122449, + "val_auc": 0.6234443998394219, + "k": 1, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377 + ] + }, + { + "test_f1": 0.6612819716267991, + "test_acc": 0.6666666666666666, + "test_auc": 0.7089795918367346, + "val_auc": 0.6473303894018467, + "k": 2, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027 + ] + }, + { + "test_f1": 0.6662578777039687, + "test_acc": 0.6666666666666666, + "test_auc": 0.7187755102040816, + "val_auc": 0.6874749096748294, + "k": 5, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 681, + 1339 + ] + }, + { + "test_f1": 0.6850677878715262, + "test_acc": 0.6868686868686869, + "test_auc": 0.7212244897959184, + "val_auc": 0.7101565636290647, + "k": 10, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 681, + 1339, + 253, + 409, + 473, + 2138, + 1915 + ] + }, + { + "test_f1": 0.6283891547049442, + "test_acc": 0.6363636363636364, + "test_auc": 0.7183673469387755, + "val_auc": 0.7189883580891208, + "k": 16, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 681, + 1339, + 253, + 409, + 473, + 2138, + 1915, + 1551, + 979, + 2046, + 1790, + 505, + 1042 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..705be5cc16f2651292e47aabd14333f125959baf --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.626603287900493, + "test_acc": 0.6346955796497081, + "test_auc": 0.6539162493043962, + "val_auc": 0.7588566827697263, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6985719160395866, + "test_acc": 0.7014178482068391, + "test_auc": 0.7386380077907624, + "val_auc": 0.7628824476650564, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1062 + ] + }, + { + "test_f1": 0.7242631017305421, + "test_acc": 0.7264386989157632, + "test_auc": 0.7749485253199777, + "val_auc": 0.8599033816425121, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1062, + 3876, + 6719, + 1114 + ] + }, + { + "test_f1": 0.7587657722817304, + "test_acc": 0.7597998331943286, + "test_auc": 0.8375180856983861, + "val_auc": 0.9206924315619968, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1062, + 3876, + 6719, + 1114, + 817, + 409, + 363, + 2744, + 238 + ] + }, + { + "test_f1": 0.7647050523001004, + "test_acc": 0.7656380316930775, + "test_auc": 0.8479855314412911, + "val_auc": 0.9319645732689211, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1062, + 3876, + 6719, + 1114, + 817, + 409, + 363, + 2744, + 238, + 1396, + 756, + 410, + 1585, + 4534, + 1662 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d222650494bd70d273c27f59d03fa52ce637cd63 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5604600487279183, + "test_acc": 0.5859119496855346, + "test_auc": 0.5924093883887116, + "val_auc": 0.5214085634253701, + "k": 1, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 486 + ] + }, + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 2, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 486, + 3876 + ] + }, + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 5, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 486, + 3876, + 588, + 1928, + 6400 + ] + }, + { + "test_f1": 0.6196932024599767, + "test_acc": 0.6211320754716981, + "test_auc": 0.6674279446178835, + "val_auc": 0.5418167266906763, + "k": 10, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 486, + 3876, + 588, + 1928, + 6400, + 794, + 138, + 152, + 409, + 1960 + ] + }, + { + "test_f1": 0.6391453612822403, + "test_acc": 0.6397484276729559, + "test_auc": 0.6942048617826739, + "val_auc": 0.53781512605042, + "k": 16, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 486, + 3876, + 588, + 1928, + 6400, + 794, + 138, + 152, + 409, + 1960, + 1099, + 2034, + 1825, + 7918, + 7160, + 1739 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..51fc7189923abae03f666cee6e54c99036ea064e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5180284668691187, + "test_acc": 0.5469778575703171, + "test_auc": 0.55163023235825, + "val_auc": 0.5142513046969088, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498 + ] + }, + { + "test_f1": 0.5383278634620238, + "test_acc": 0.5415918611609815, + "test_auc": 0.5643268486949546, + "val_auc": 0.5096346848655159, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 6611 + ] + }, + { + "test_f1": 0.33366585529398596, + "test_acc": 0.5002992220227409, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 6611, + 7163, + 1342, + 19642 + ] + }, + { + "test_f1": 0.33366585529398596, + "test_acc": 0.5002992220227409, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 6611, + 7163, + 1342, + 19642, + 588, + 486, + 4072, + 150, + 3727 + ] + }, + { + "test_f1": 0.5665625870698315, + "test_acc": 0.5667265110712149, + "test_auc": 0.5795633613156462, + "val_auc": 0.518466479325572, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 6611, + 7163, + 1342, + 19642, + 588, + 486, + 4072, + 150, + 3727, + 1336, + 4218, + 2363, + 23174, + 27404, + 462 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8c1387b6d98d6347dd2008d5a3aea3a9980054bd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5275402559124975, + "test_acc": 0.5562264150943397, + "test_auc": 0.5668383729655233, + "val_auc": 0.6286514605842337, + "k": 1, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.5569767583673452, + "test_acc": 0.5841509433962264, + "test_auc": 0.5889423101264861, + "val_auc": 0.640656262505002, + "k": 2, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 27965 + ] + }, + { + "test_f1": 0.5980906948125729, + "test_acc": 0.6120754716981132, + "test_auc": 0.6282702505926349, + "val_auc": 0.7024809923969588, + "k": 5, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 27965, + 777, + 1377, + 566 + ] + }, + { + "test_f1": 0.5987638451056093, + "test_acc": 0.6145911949685534, + "test_auc": 0.6379573363684876, + "val_auc": 0.752701080432173, + "k": 10, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 27965, + 777, + 1377, + 566, + 6400, + 7143, + 1345, + 3876, + 4675 + ] + }, + { + "test_f1": 0.6029543010313995, + "test_acc": 0.6186163522012579, + "test_auc": 0.6958591002481926, + "val_auc": 0.7541016406562625, + "k": 16, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 27965, + 777, + 1377, + 566, + 6400, + 7143, + 1345, + 3876, + 4675, + 1011, + 3245, + 3053, + 121, + 6659, + 1946 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..556d1950b85aa3976a6214c5d0337d85d3b95ec1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9989937094180884, + "test_acc": 0.9989937106918239, + "test_auc": 0.9999989873817642, + "val_auc": 1.0, + "k": 1, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625 + ] + }, + { + "test_f1": 0.9987421359765241, + "test_acc": 0.9987421383647799, + "test_auc": 0.9999929116723493, + "val_auc": 1.0, + "k": 2, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625, + 1703 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.9999997468454409, + "val_auc": 1.0, + "k": 5, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625, + 1703, + 1551, + 680, + 1500 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.9999992405363232, + "val_auc": 1.0, + "k": 10, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625, + 1703, + 1551, + 680, + 1500, + 1472, + 1960, + 1255, + 4252, + 1525 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.999993417981467, + "val_auc": 1.0, + "k": 16, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625, + 1703, + 1551, + 680, + 1500, + 1472, + 1960, + 1255, + 4252, + 1525, + 6718, + 490, + 1758, + 1042, + 374, + 567 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..31f1bfc77f29e7d696d54701a958c3d2e25cfc94 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.622889358540223, + "test_acc": 0.6228930817610063, + "test_auc": 0.6724144565429822, + "val_auc": 0.7555022008803521, + "k": 1, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6365334838923017, + "test_acc": 0.6369811320754717, + "test_auc": 0.6874943673110632, + "val_auc": 0.7789115646258504, + "k": 2, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1255 + ] + }, + { + "test_f1": 0.650198104995565, + "test_acc": 0.650314465408805, + "test_auc": 0.7028962906781404, + "val_auc": 0.7667066826730692, + "k": 5, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1255, + 7143, + 1928, + 6400 + ] + }, + { + "test_f1": 0.6769788832816692, + "test_acc": 0.6769811320754717, + "test_auc": 0.7498212728813748, + "val_auc": 0.789515806322529, + "k": 10, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1255, + 7143, + 1928, + 6400, + 2046, + 3876, + 3245, + 1585, + 1149 + ] + }, + { + "test_f1": 0.6945180758244079, + "test_acc": 0.6945911949685535, + "test_auc": 0.7692326581532476, + "val_auc": 0.8211284513805521, + "k": 16, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1255, + 7143, + 1928, + 6400, + 2046, + 3876, + 3245, + 1585, + 1149, + 121, + 447, + 1541, + 1327, + 765, + 18474 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..67e678e3a7edf37cfe2762211a6871b2a3c6f6b7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6560772225539482, + "test_acc": 0.6571069182389937, + "test_auc": 0.698847083507588, + "val_auc": 0.6966786714685875, + "k": 1, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6557166439691634, + "test_acc": 0.6573584905660378, + "test_auc": 0.698847083507588, + "val_auc": 0.6966786714685875, + "k": 2, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3876 + ] + }, + { + "test_f1": 0.6764611773027432, + "test_acc": 0.6777358490566038, + "test_auc": 0.7447672952663136, + "val_auc": 0.7120848339335735, + "k": 5, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3876, + 7143, + 1360, + 2592 + ] + }, + { + "test_f1": 0.6936992626119104, + "test_acc": 0.6945911949685535, + "test_auc": 0.7658930432114581, + "val_auc": 0.7669067627050821, + "k": 10, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3876, + 7143, + 1360, + 2592, + 3245, + 1019, + 9390, + 121, + 6400 + ] + }, + { + "test_f1": 0.7126695627245748, + "test_acc": 0.7129559748427673, + "test_auc": 0.7929598223462567, + "val_auc": 0.8099239695878352, + "k": 16, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3876, + 7143, + 1360, + 2592, + 3245, + 1019, + 9390, + 121, + 6400, + 253, + 1928, + 409, + 6993, + 1149, + 1011 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6ab0646a38764b052f22d222c3a2fedaa61d7494 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7707545212377447, + "test_acc": 0.7718791064388962, + "test_auc": 0.7992759232889792, + "val_auc": 0.7448, + "k": 1, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062 + ] + }, + { + "test_f1": 0.7852446990701472, + "test_acc": 0.7876478318002628, + "test_auc": 0.8789315448019701, + "val_auc": 0.8188000000000001, + "k": 2, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448 + ] + }, + { + "test_f1": 0.8233650215917669, + "test_acc": 0.8233902759526939, + "test_auc": 0.9078048219839945, + "val_auc": 0.8371999999999999, + "k": 5, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 1503, + 32507, + 2073 + ] + }, + { + "test_f1": 0.8493821447190094, + "test_acc": 0.8494086727989487, + "test_auc": 0.9239274641346082, + "val_auc": 0.8592, + "k": 10, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 1503, + 32507, + 2073, + 817, + 903, + 29329, + 811, + 5636 + ] + }, + { + "test_f1": 0.8599200920858178, + "test_acc": 0.8599211563731932, + "test_auc": 0.9327082204035579, + "val_auc": 0.8887999999999999, + "k": 16, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 1503, + 32507, + 2073, + 817, + 903, + 29329, + 811, + 5636, + 1942, + 296, + 1313, + 1222, + 765, + 5669 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4e2fb17fe8cf8b52d01a2c60d80f19542421f506 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5230525962574157, + "test_acc": 0.5240251572327044, + "test_auc": 0.5355927715259853, + "val_auc": 0.6390556222488996, + "k": 1, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411 + ] + }, + { + "test_f1": 0.5232937993852397, + "test_acc": 0.5242767295597485, + "test_auc": 0.5355927715259853, + "val_auc": 0.6390556222488996, + "k": 2, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 1377 + ] + }, + { + "test_f1": 0.5536886171248395, + "test_acc": 0.5559748427672956, + "test_auc": 0.5957253333792386, + "val_auc": 0.6404561824729892, + "k": 5, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 1377, + 800, + 786, + 1703 + ] + }, + { + "test_f1": 0.5557171206878182, + "test_acc": 0.5564779874213837, + "test_auc": 0.5988145784622176, + "val_auc": 0.6442577030812324, + "k": 10, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 1377, + 800, + 786, + 1703, + 3725, + 5689, + 3802, + 1417, + 1928 + ] + }, + { + "test_f1": 0.5749233516893819, + "test_acc": 0.5758490566037736, + "test_auc": 0.6096852883784842, + "val_auc": 0.6466586634653861, + "k": 16, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 1377, + 800, + 786, + 1703, + 3725, + 5689, + 3802, + 1417, + 1928, + 373, + 4880, + 334, + 152, + 11455, + 583 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..895a5ed79c8ec9b82cd8e7fae21a6b41665e3cd3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9937105405642113, + "test_acc": 0.9937106918238994, + "test_auc": 0.9973851665605105, + "val_auc": 1.0, + "k": 1, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.9909427909065129, + "test_acc": 0.9909433962264151, + "test_auc": 0.9988704243579241, + "val_auc": 1.0, + "k": 2, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 152 + ] + }, + { + "test_f1": 0.9924527556744696, + "test_acc": 0.9924528301886792, + "test_auc": 0.9993716703846633, + "val_auc": 1.0, + "k": 5, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 152, + 26988, + 487, + 292 + ] + }, + { + "test_f1": 0.9929559052961715, + "test_acc": 0.9929559748427673, + "test_auc": 0.9996562161089335, + "val_auc": 1.0, + "k": 10, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 152, + 26988, + 487, + 292, + 32198, + 1839, + 1934, + 878, + 138 + ] + }, + { + "test_f1": 0.9947169008855618, + "test_acc": 0.9947169811320755, + "test_auc": 0.9997772239881159, + "val_auc": 1.0, + "k": 16, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 152, + 26988, + 487, + 292, + 32198, + 1839, + 1934, + 878, + 138, + 6935, + 1288, + 7368, + 3173, + 1744, + 334 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d928db0a596f0fc57b6fb4d96cf684ae74c44007 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8402295203938994, + "test_acc": 0.840503144654088, + "test_auc": 0.8863442355187997, + "val_auc": 0.9035614245698279, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487 + ] + }, + { + "test_f1": 0.8564804961113728, + "test_acc": 0.8566037735849057, + "test_auc": 0.9282628331640574, + "val_auc": 0.942376950780312, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 878 + ] + }, + { + "test_f1": 0.9340794142495072, + "test_acc": 0.9340880503144654, + "test_auc": 0.9781398506793149, + "val_auc": 0.9531812725090036, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 878, + 1744, + 6631, + 3173 + ] + }, + { + "test_f1": 0.9527010438309886, + "test_acc": 0.9527044025157233, + "test_auc": 0.98603865265068, + "val_auc": 0.9695878351340536, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 878, + 1744, + 6631, + 3173, + 26988, + 27655, + 1524, + 1028, + 1288 + ] + }, + { + "test_f1": 0.957484163717293, + "test_acc": 0.9574842767295597, + "test_auc": 0.9891596686308086, + "val_auc": 0.9771908763505401, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 878, + 1744, + 6631, + 3173, + 26988, + 27655, + 1524, + 1028, + 1288, + 7368, + 771, + 848, + 930, + 5539, + 3816 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5c5bc77370e61c908bf2bd62f2219614df5dbafc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8702227902275183, + "test_acc": 0.8713924050632912, + "test_auc": 0.8796561780884289, + "val_auc": 0.8419367747098838, + "k": 1, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104 + ] + }, + { + "test_f1": 0.8712940662357316, + "test_acc": 0.8724050632911392, + "test_auc": 0.9024217663635357, + "val_auc": 0.8689475790316126, + "k": 2, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292 + ] + }, + { + "test_f1": 0.8694868710488868, + "test_acc": 0.8708860759493671, + "test_auc": 0.9335367879600803, + "val_auc": 0.9119647859143658, + "k": 5, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 4201, + 1708, + 918 + ] + }, + { + "test_f1": 0.8695685954582622, + "test_acc": 0.8708860759493671, + "test_auc": 0.9344853541382097, + "val_auc": 0.9155662264905962, + "k": 10, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 4201, + 1708, + 918, + 990, + 1438, + 3385, + 1627, + 1284 + ] + }, + { + "test_f1": 0.8688708428592916, + "test_acc": 0.8693670886075949, + "test_auc": 0.9417698296477692, + "val_auc": 0.9263705482192878, + "k": 16, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 4201, + 1708, + 918, + 990, + 1438, + 3385, + 1627, + 1284, + 1745, + 4555, + 10639, + 67, + 902, + 6747 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..75a3083ba7c00a6c5c786375057298111d558698 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7448513750051456, + "test_acc": 0.7507692307692307, + "test_auc": 0.7659095499377251, + "val_auc": 0.7768257853937168, + "k": 1, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468 + ] + }, + { + "test_f1": 0.8274218550076278, + "test_acc": 0.8287179487179487, + "test_auc": 0.8751725182616892, + "val_auc": 0.8620971032231742, + "k": 2, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351 + ] + }, + { + "test_f1": 0.8419999959266553, + "test_acc": 0.8430769230769231, + "test_auc": 0.8918478809708149, + "val_auc": 0.8655650754793962, + "k": 5, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 1474, + 26988 + ] + }, + { + "test_f1": 0.856728804346957, + "test_acc": 0.8574358974358974, + "test_auc": 0.9120788029757296, + "val_auc": 0.8886168910648715, + "k": 10, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 1474, + 26988, + 343, + 21810, + 1305, + 7595, + 805 + ] + }, + { + "test_f1": 0.8683787450850086, + "test_acc": 0.8687179487179487, + "test_auc": 0.9164380449052413, + "val_auc": 0.8873929008567931, + "k": 16, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 1474, + 26988, + 343, + 21810, + 1305, + 7595, + 805, + 1839, + 3173, + 52, + 559, + 391, + 918 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8f5551f9306743e5e594fd12c1996f99bb9e48d9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.888830722087599, + "test_acc": 0.889308176100629, + "test_auc": 0.9032485046160204, + "val_auc": 0.9503801520608243, + "k": 1, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.8892910282062017, + "test_acc": 0.889811320754717, + "test_auc": 0.9149933572243729, + "val_auc": 0.9519807923169268, + "k": 2, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 32198 + ] + }, + { + "test_f1": 0.9229332386883256, + "test_acc": 0.9230188679245283, + "test_auc": 0.9677372235425638, + "val_auc": 0.994797919167667, + "k": 5, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 32198, + 1279, + 1447, + 1012 + ] + }, + { + "test_f1": 0.9383401167716435, + "test_acc": 0.9383647798742139, + "test_auc": 0.9805710204862795, + "val_auc": 0.9923969587835133, + "k": 10, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 32198, + 1279, + 1447, + 1012, + 1426, + 7740, + 20705, + 1556, + 238 + ] + }, + { + "test_f1": 0.9370871292513657, + "test_acc": 0.9371069182389937, + "test_auc": 0.982690683608445, + "val_auc": 0.9919967987194878, + "k": 16, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 32198, + 1279, + 1447, + 1012, + 1426, + 7740, + 20705, + 1556, + 238, + 26750, + 3475, + 25127, + 11381, + 1839, + 1101 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ec50e6a3576d29e7942d3e1bdf3b11f7ef52bfb7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8569721730018661, + "test_acc": 0.8584615384615385, + "test_auc": 0.8655241188945366, + "val_auc": 0.8890248878008976, + "k": 1, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628 + ] + }, + { + "test_f1": 0.8590450400598102, + "test_acc": 0.8605128205128205, + "test_auc": 0.8847851516477598, + "val_auc": 0.8914728682170543, + "k": 2, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 9381 + ] + }, + { + "test_f1": 0.8444785007020233, + "test_acc": 0.8471794871794872, + "test_auc": 0.8655241188945366, + "val_auc": 0.8890248878008976, + "k": 5, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 9381, + 805, + 5964, + 334 + ] + }, + { + "test_f1": 0.8445633476103046, + "test_acc": 0.8471794871794872, + "test_auc": 0.8867017874575016, + "val_auc": 0.9277845777233782, + "k": 10, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 9381, + 805, + 5964, + 334, + 238, + 4211, + 3734, + 2062, + 737 + ] + }, + { + "test_f1": 0.8978326516479597, + "test_acc": 0.8984615384615384, + "test_auc": 0.9320530514693507, + "val_auc": 0.9481844145246838, + "k": 16, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 9381, + 805, + 5964, + 334, + 238, + 4211, + 3734, + 2062, + 737, + 1524, + 416, + 52, + 1288, + 878, + 930 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..68a1ee83755a647a6a9fb82efc29b567af5fe58a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8594561327784747, + "test_acc": 0.8594871794871795, + "test_auc": 0.903579122765678, + "val_auc": 0.9320685434516524, + "k": 1, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524 + ] + }, + { + "test_f1": 0.8573458615940215, + "test_acc": 0.8574358974358974, + "test_auc": 0.9114749890598175, + "val_auc": 0.9318645450836394, + "k": 2, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 3816 + ] + }, + { + "test_f1": 0.8542255675082507, + "test_acc": 0.8543589743589743, + "test_auc": 0.9106586831386542, + "val_auc": 0.9314565483476133, + "k": 5, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 3816, + 24947, + 930, + 969 + ] + }, + { + "test_f1": 0.8696860163348626, + "test_acc": 0.8697435897435898, + "test_auc": 0.9261958460968795, + "val_auc": 0.9424724602203183, + "k": 10, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 3816, + 24947, + 930, + 969, + 6631, + 27655, + 862, + 7015, + 776 + ] + }, + { + "test_f1": 0.8697222110942533, + "test_acc": 0.8697435897435898, + "test_auc": 0.9280135994883362, + "val_auc": 0.944920440636475, + "k": 16, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 3816, + 24947, + 930, + 969, + 6631, + 27655, + 862, + 7015, + 776, + 1474, + 888, + 1178, + 1874, + 1682, + 483 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2e906c52ef096a19368f51fdfcf55f8a981493f2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8083795508810565, + "test_acc": 0.8102564102564103, + "test_auc": 0.8474244285858552, + "val_auc": 0.7904936760505916, + "k": 1, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.8291645923956346, + "test_acc": 0.8307692307692308, + "test_auc": 0.9087504628538727, + "val_auc": 0.8204814361485108, + "k": 2, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 152 + ] + }, + { + "test_f1": 0.8498375179076934, + "test_acc": 0.8502564102564103, + "test_auc": 0.9061668966910155, + "val_auc": 0.8527131782945736, + "k": 5, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 152, + 1148, + 1934, + 334 + ] + }, + { + "test_f1": 0.8703491898561547, + "test_acc": 0.8707692307692307, + "test_auc": 0.9318384555828593, + "val_auc": 0.8967768257853936, + "k": 10, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 152, + 1148, + 1934, + 334, + 3159, + 1524, + 1866, + 5539, + 913 + ] + }, + { + "test_f1": 0.8676769963158012, + "test_acc": 0.8676923076923077, + "test_auc": 0.9454505672063823, + "val_auc": 0.9179926560587516, + "k": 16, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 152, + 1148, + 1934, + 334, + 3159, + 1524, + 1866, + 5539, + 913, + 6631, + 487, + 508, + 901, + 1874, + 1556 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d0f4c464aec386982b8e67a8065867f2a50b9fac --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8443600680315864, + "test_acc": 0.8451282051282051, + "test_auc": 0.8816924462247956, + "val_auc": 0.882700938392493, + "k": 1, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52 + ] + }, + { + "test_f1": 0.8547088147076733, + "test_acc": 0.8553846153846154, + "test_auc": 0.8972779984515433, + "val_auc": 0.8918808649530804, + "k": 2, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 29567 + ] + }, + { + "test_f1": 0.8589441347702217, + "test_acc": 0.8594871794871795, + "test_auc": 0.9039409903389772, + "val_auc": 0.8967768257853937, + "k": 5, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 29567, + 1132, + 1274, + 2131 + ] + }, + { + "test_f1": 0.8591939164846409, + "test_acc": 0.8594871794871795, + "test_auc": 0.9092848486888613, + "val_auc": 0.9045287637698899, + "k": 10, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 29567, + 1132, + 1274, + 2131, + 18545, + 14503, + 238, + 25681, + 1874 + ] + }, + { + "test_f1": 0.8714500616939641, + "test_acc": 0.8717948717948718, + "test_auc": 0.919730619046016, + "val_auc": 0.937984496124031, + "k": 16, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 29567, + 1132, + 1274, + 2131, + 18545, + 14503, + 238, + 25681, + 1874, + 559, + 776, + 862, + 6631, + 1934, + 508 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..110b5df2655ab02586f612f3938e996b806673e2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8448299762003776, + "test_acc": 0.8449197860962567, + "test_auc": 0.907071871739966, + "val_auc": 0.9147474747474748, + "k": 1, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839 + ] + }, + { + "test_f1": 0.8501727788037631, + "test_acc": 0.8502673796791443, + "test_auc": 0.9249391460312233, + "val_auc": 0.9434343434343435, + "k": 2, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682 + ] + }, + { + "test_f1": 0.867281003310415, + "test_acc": 0.8673796791443851, + "test_auc": 0.9341587510752393, + "val_auc": 0.9422222222222222, + "k": 5, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1305, + 152, + 1274 + ] + }, + { + "test_f1": 0.8919160523878492, + "test_acc": 0.8919786096256684, + "test_auc": 0.950040264280093, + "val_auc": 0.9587878787878789, + "k": 10, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1305, + 152, + 1274, + 1158, + 334, + 340, + 343, + 1960 + ] + }, + { + "test_f1": 0.9037232724940983, + "test_acc": 0.9037433155080213, + "test_auc": 0.9568668899504016, + "val_auc": 0.9498989898989899, + "k": 16, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1305, + 152, + 1274, + 1158, + 334, + 340, + 343, + 1960, + 456, + 3173, + 238, + 1441, + 6631, + 1556 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..61509d6aab66d6930ae91173d4a9de907e9f2212 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9383734523020597, + "test_acc": 0.9383908045977012, + "test_auc": 0.9826272390280859, + "val_auc": 0.9931006493506493, + "k": 1, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7847 + ] + }, + { + "test_f1": 0.9563183302196446, + "test_acc": 0.9563218390804598, + "test_auc": 0.9912485118242329, + "val_auc": 0.9979707792207793, + "k": 2, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7847, + 1072 + ] + }, + { + "test_f1": 0.9770114359386776, + "test_acc": 0.9770114942528736, + "test_auc": 0.9965636668650901, + "val_auc": 1.0, + "k": 5, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7847, + 1072, + 495, + 760, + 1876 + ] + }, + { + "test_f1": 0.9779307732329513, + "test_acc": 0.9779310344827586, + "test_auc": 0.9966194734563559, + "val_auc": 1.0, + "k": 10, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7847, + 1072, + 495, + 760, + 1876, + 1065, + 1101, + 1622, + 192, + 835 + ] + }, + { + "test_f1": 0.9811493455911703, + "test_acc": 0.9811494252873563, + "test_auc": 0.9963590426971157, + "val_auc": 1.0, + "k": 16, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7847, + 1072, + 495, + 760, + 1876, + 1065, + 1101, + 1622, + 192, + 835, + 763, + 1682, + 1684, + 3045, + 1787, + 652 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..27d4797d4899704204375d36792f592205cb1430 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9741993835086503, + "test_acc": 0.9741997133301481, + "test_auc": 0.9934731117405462, + "val_auc": 0.9965679965679966, + "k": 1, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9355 + ] + }, + { + "test_f1": 0.981843488379163, + "test_acc": 0.9818442427138079, + "test_auc": 0.9945624482953206, + "val_auc": 0.9982839982839983, + "k": 2, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9355, + 6080 + ] + }, + { + "test_f1": 0.9890103066496144, + "test_acc": 0.989010989010989, + "test_auc": 0.9980167317711899, + "val_auc": 1.0, + "k": 5, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9355, + 6080, + 213, + 835, + 1905 + ] + }, + { + "test_f1": 0.9928329984782678, + "test_acc": 0.9928332537028189, + "test_auc": 0.9991078945398032, + "val_auc": 1.0, + "k": 10, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9355, + 6080, + 213, + 835, + 1905, + 88, + 1715, + 1233, + 4095, + 1877 + ] + }, + { + "test_f1": 0.993788663906784, + "test_acc": 0.9937888198757764, + "test_auc": 0.999374521760251, + "val_auc": 1.0, + "k": 16, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9355, + 6080, + 213, + 835, + 1905, + 88, + 1715, + 1233, + 4095, + 1877, + 1682, + 760, + 495, + 3877, + 337, + 1645 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8f51e48ac02297dc43b1742e1bf2ed2a597d9864 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9908042944570734, + "test_acc": 0.9908045977011494, + "test_auc": 0.9984885714865522, + "val_auc": 1.0, + "k": 1, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559 + ] + }, + { + "test_f1": 0.9898849035194436, + "test_acc": 0.9898850574712643, + "test_auc": 0.9988246793657665, + "val_auc": 1.0, + "k": 2, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2481 + ] + }, + { + "test_f1": 0.9926436470599098, + "test_acc": 0.9926436781609196, + "test_auc": 0.9998055224849829, + "val_auc": 1.0, + "k": 5, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2481, + 1789, + 1832, + 2236 + ] + }, + { + "test_f1": 0.9931033666350384, + "test_acc": 0.993103448275862, + "test_auc": 0.9996846082039071, + "val_auc": 1.0, + "k": 10, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2481, + 1789, + 1832, + 2236, + 1934, + 334, + 1762, + 3825, + 982 + ] + }, + { + "test_f1": 0.9935632102268172, + "test_acc": 0.9935632183908046, + "test_auc": 0.9998089047026354, + "val_auc": 1.0, + "k": 16, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2481, + 1789, + 1832, + 2236, + 1934, + 334, + 1762, + 3825, + 982, + 6292, + 1268, + 3203, + 913, + 2418, + 2020 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..38f61e4389015fc4beb2cfe48a607954cf616a59 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9370100295664736, + "test_acc": 0.9370114942528736, + "test_auc": 0.9712858176849397, + "val_auc": 0.963474025974026, + "k": 1, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095 + ] + }, + { + "test_f1": 0.9687282152574507, + "test_acc": 0.968735632183908, + "test_auc": 0.9892682233887115, + "val_auc": 0.9857954545454546, + "k": 2, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023 + ] + }, + { + "test_f1": 0.9760855844754918, + "test_acc": 0.9760919540229885, + "test_auc": 0.9929582228475566, + "val_auc": 0.9857954545454546, + "k": 5, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 494, + 1682, + 1789 + ] + }, + { + "test_f1": 0.9788488756993098, + "test_acc": 0.9788505747126437, + "test_auc": 0.9958111234374154, + "val_auc": 0.9914772727272727, + "k": 10, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 494, + 1682, + 1789, + 1684, + 1301, + 292, + 1079, + 586 + ] + }, + { + "test_f1": 0.9737897791406565, + "test_acc": 0.9737931034482759, + "test_auc": 0.9966938822447102, + "val_auc": 0.9939123376623377, + "k": 16, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 494, + 1682, + 1789, + 1684, + 1301, + 292, + 1079, + 586, + 2039, + 7315, + 6679, + 862, + 129, + 1260 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..eb1a5f76151516d083093b21f443f4a236fa8bf7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9885053680733451, + "test_acc": 0.9885057471264368, + "test_auc": 0.9990216935440229, + "val_auc": 0.9939123376623377, + "k": 1, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2683 + ] + }, + { + "test_f1": 0.9885053680733451, + "test_acc": 0.9885057471264368, + "test_auc": 0.9990216935440229, + "val_auc": 0.9939123376623377, + "k": 2, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2683, + 32 + ] + }, + { + "test_f1": 0.9935631421929442, + "test_acc": 0.9935632183908046, + "test_auc": 0.9996211916229232, + "val_auc": 0.9979707792207791, + "k": 5, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2683, + 32, + 1023, + 816, + 12798 + ] + }, + { + "test_f1": 0.9940229809249097, + "test_acc": 0.9940229885057471, + "test_auc": 0.999813132474701, + "val_auc": 0.9979707792207791, + "k": 10, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2683, + 32, + 1023, + 816, + 12798, + 1233, + 1558, + 232, + 2022, + 769 + ] + }, + { + "test_f1": 0.994482756288126, + "test_acc": 0.9944827586206897, + "test_auc": 0.9997877658423074, + "val_auc": 0.9955357142857143, + "k": 16, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2683, + 32, + 1023, + 816, + 12798, + 1233, + 1558, + 232, + 2022, + 769, + 1529, + 21, + 1458, + 868, + 495, + 575 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4aea9c17e5e44017f4dccb9262298c801c365868 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8343677938960081, + "test_acc": 0.8344654088050314, + "test_auc": 0.8781589891639723, + "val_auc": 0.9197679071628653, + "k": 1, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.8915148455835722, + "test_acc": 0.8915723270440251, + "test_auc": 0.9506146086382411, + "val_auc": 0.9813925570228091, + "k": 2, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028 + ] + }, + { + "test_f1": 0.9063919538318146, + "test_acc": 0.9064150943396226, + "test_auc": 0.9617769526064286, + "val_auc": 0.991796718687475, + "k": 5, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 1305, + 1478, + 848 + ] + }, + { + "test_f1": 0.9068966727902773, + "test_acc": 0.9069182389937107, + "test_auc": 0.9617764462973106, + "val_auc": 0.991796718687475, + "k": 10, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 1305, + 1478, + 848, + 778, + 4663, + 1839, + 343, + 3173 + ] + }, + { + "test_f1": 0.9204781313737264, + "test_acc": 0.920503144654088, + "test_auc": 0.9659762804304438, + "val_auc": 0.9931972789115646, + "k": 16, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 1305, + 1478, + 848, + 778, + 4663, + 1839, + 343, + 3173, + 1279, + 5539, + 654, + 334, + 17275, + 1447 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8405cc1792755cfaa9b8234aa62cb18bf4908c37 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9977011465095805, + "test_acc": 0.9977011494252873, + "test_auc": 0.9990538246117215, + "val_auc": 1.0, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379 + ] + }, + { + "test_f1": 0.9986206879057483, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999898533470426, + "val_auc": 1.0, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686 + ] + }, + { + "test_f1": 0.9977011465095805, + "test_acc": 0.9977011494252873, + "test_auc": 0.9990800367985281, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 1840, + 1931, + 4278 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9991079400941609, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 1840, + 1931, + 4278, + 14067, + 23428, + 4787, + 3516, + 298 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9990851101250068, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 1840, + 1931, + 4278, + 14067, + 23428, + 4787, + 3516, + 298, + 1162, + 931, + 642, + 5034, + 598, + 3818 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5277f2c1c42f0e73eecbac2329087473b42835b9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.977468406629687, + "test_acc": 0.9774712643678161, + "test_auc": 0.9884869311109908, + "val_auc": 0.9983766233766234, + "k": 1, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586 + ] + }, + { + "test_f1": 0.9820666909909755, + "test_acc": 0.9820689655172414, + "test_auc": 0.9896664795172899, + "val_auc": 0.9991883116883117, + "k": 2, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 21724 + ] + }, + { + "test_f1": 0.9866662269650803, + "test_acc": 0.9866666666666667, + "test_auc": 0.9956411670003789, + "val_auc": 0.9983766233766234, + "k": 5, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 21724, + 1934, + 10, + 878 + ] + }, + { + "test_f1": 0.98666635662356, + "test_acc": 0.9866666666666667, + "test_auc": 0.9971293427674658, + "val_auc": 0.9991883116883117, + "k": 10, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 21724, + 1934, + 10, + 878, + 266, + 6336, + 835, + 1874, + 1121 + ] + }, + { + "test_f1": 0.9903445091816099, + "test_acc": 0.9903448275862069, + "test_auc": 0.9972401103955841, + "val_auc": 1.0, + "k": 16, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 21724, + 1934, + 10, + 878, + 266, + 6336, + 835, + 1874, + 1121, + 6147, + 1023, + 2211, + 1682, + 1991, + 760 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..655ef97fd3b83a6b630406f7065a202002c8ea74 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8978383897929167, + "test_acc": 0.8979310344827586, + "test_auc": 0.9489428878727204, + "val_auc": 0.9606331168831168, + "k": 1, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979 + ] + }, + { + "test_f1": 0.9337926835846441, + "test_acc": 0.9337931034482758, + "test_auc": 0.9787825031116402, + "val_auc": 0.9979707792207793, + "k": 2, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 1682 + ] + }, + { + "test_f1": 0.9370109350165414, + "test_acc": 0.9370114942528736, + "test_auc": 0.9800449158504249, + "val_auc": 0.9963474025974026, + "k": 5, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 1682, + 495, + 17284, + 653 + ] + }, + { + "test_f1": 0.954482566183558, + "test_acc": 0.9544827586206897, + "test_auc": 0.987890815249743, + "val_auc": 0.9967532467532468, + "k": 10, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 1682, + 495, + 17284, + 653, + 1707, + 1966, + 713, + 361, + 1079 + ] + }, + { + "test_f1": 0.9531004145918218, + "test_acc": 0.953103448275862, + "test_auc": 0.9895506385626929, + "val_auc": 0.9979707792207793, + "k": 16, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 1682, + 495, + 17284, + 653, + 1707, + 1966, + 713, + 361, + 1079, + 532, + 1840, + 2024, + 27304, + 7925, + 1583 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2e7552fcc24c389a82f385a7de3db315f6e27b42 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9273530969167877, + "test_acc": 0.9273563218390805, + "test_auc": 0.9596898844634449, + "val_auc": 0.9565746753246753, + "k": 1, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2689 + ] + }, + { + "test_f1": 0.9452873563218391, + "test_acc": 0.9452873563218391, + "test_auc": 0.9808211348016667, + "val_auc": 0.9728084415584416, + "k": 2, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2689, + 7182 + ] + }, + { + "test_f1": 0.9452870093503921, + "test_acc": 0.9452873563218391, + "test_auc": 0.9826534512148926, + "val_auc": 0.974025974025974, + "k": 5, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2689, + 7182, + 732, + 23259, + 595 + ] + }, + { + "test_f1": 0.9659770114942529, + "test_acc": 0.9659770114942529, + "test_auc": 0.9875745778992369, + "val_auc": 0.989448051948052, + "k": 10, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2689, + 7182, + 732, + 23259, + 595, + 1961, + 1536, + 3589, + 760, + 1264 + ] + }, + { + "test_f1": 0.9724135481817537, + "test_acc": 0.9724137931034482, + "test_auc": 0.9924627279614697, + "val_auc": 0.9926948051948052, + "k": 16, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2689, + 7182, + 732, + 23259, + 595, + 1961, + 1536, + 3589, + 760, + 1264, + 560, + 2030, + 506, + 1682, + 248, + 1441 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fbdf5d11605c7d3134a0cb6cb108670af6befa37 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9526303404082088, + "test_acc": 0.9526436781609195, + "test_auc": 0.9846235929974565, + "val_auc": 0.9926948051948052, + "k": 1, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786 + ] + }, + { + "test_f1": 0.9618294786431068, + "test_acc": 0.9618390804597701, + "test_auc": 0.9906845270306835, + "val_auc": 0.9926948051948052, + "k": 2, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558 + ] + }, + { + "test_f1": 0.9682678798358357, + "test_acc": 0.9682758620689655, + "test_auc": 0.9924906312571027, + "val_auc": 0.9955357142857143, + "k": 5, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3930, + 760, + 271 + ] + }, + { + "test_f1": 0.9687273027756429, + "test_acc": 0.968735632183908, + "test_auc": 0.9925033145732993, + "val_auc": 0.9955357142857143, + "k": 10, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3930, + 760, + 271, + 292, + 2024, + 67, + 1910, + 276 + ] + }, + { + "test_f1": 0.9678075175631621, + "test_acc": 0.967816091954023, + "test_auc": 0.9927028654147951, + "val_auc": 0.9955357142857143, + "k": 16, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3930, + 760, + 271, + 292, + 2024, + 67, + 1910, + 276, + 5479, + 2042, + 968, + 1596, + 1142, + 2022 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b3804e1ae5478615acd23e872b3e506f67fbf6da --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9705743643106016, + "test_acc": 0.9705747126436781, + "test_auc": 0.9905352866767683, + "val_auc": 0.9951298701298701, + "k": 1, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20819 + ] + }, + { + "test_f1": 0.9705735805324612, + "test_acc": 0.9705747126436781, + "test_auc": 0.9921016762270685, + "val_auc": 0.9939123376623377, + "k": 2, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20819, + 1856 + ] + }, + { + "test_f1": 0.9742521551153132, + "test_acc": 0.9742528735632184, + "test_auc": 0.9916214013204179, + "val_auc": 0.9951298701298701, + "k": 5, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20819, + 1856, + 7739, + 1168, + 1167 + ] + }, + { + "test_f1": 0.9756300202725408, + "test_acc": 0.975632183908046, + "test_auc": 0.9951380621245738, + "val_auc": 0.9971590909090909, + "k": 10, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20819, + 1856, + 7739, + 1168, + 1167, + 1536, + 7499, + 517, + 1815, + 1037 + ] + }, + { + "test_f1": 0.977928459044762, + "test_acc": 0.9779310344827586, + "test_auc": 0.9950070011905406, + "val_auc": 0.9983766233766235, + "k": 16, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 20819, + 1856, + 7739, + 1168, + 1167, + 1536, + 7499, + 517, + 1815, + 1037, + 2011, + 548, + 760, + 835, + 765, + 1121 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..399bb3f0364edfac9505fd0f501f39070606f44d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9152248897284851, + "test_acc": 0.9154022988505747, + "test_auc": 0.9357594262405975, + "val_auc": 0.9699675324675325, + "k": 1, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765 + ] + }, + { + "test_f1": 0.9401329986347822, + "test_acc": 0.9402298850574713, + "test_auc": 0.952844275934845, + "val_auc": 0.9890422077922076, + "k": 2, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765, + 5495 + ] + }, + { + "test_f1": 0.964124872203006, + "test_acc": 0.9641379310344828, + "test_auc": 0.9943702987174631, + "val_auc": 0.9987824675324676, + "k": 5, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765, + 5495, + 10075, + 789, + 1763 + ] + }, + { + "test_f1": 0.9719469662316808, + "test_acc": 0.9719540229885058, + "test_auc": 0.9953587518263975, + "val_auc": 0.9987824675324676, + "k": 10, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765, + 5495, + 10075, + 789, + 1763, + 1175, + 170, + 1682, + 1797, + 1051 + ] + }, + { + "test_f1": 0.9756245476802405, + "test_acc": 0.975632183908046, + "test_auc": 0.9961180596893771, + "val_auc": 1.0, + "k": 16, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765, + 5495, + 10075, + 789, + 1763, + 1175, + 170, + 1682, + 1797, + 1051, + 27505, + 1271, + 1310, + 434, + 3197, + 2013 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9a26d9dc4ca54b3d4066561970afaf2fcffd78e2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9636719253329035, + "test_acc": 0.9636781609195403, + "test_auc": 0.9835793332972562, + "val_auc": 0.9797077922077921, + "k": 1, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879 + ] + }, + { + "test_f1": 0.9779309785011305, + "test_acc": 0.9779310344827586, + "test_auc": 0.9936583419016181, + "val_auc": 0.9857954545454545, + "k": 2, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293 + ] + }, + { + "test_f1": 0.9908045860383102, + "test_acc": 0.9908045977011494, + "test_auc": 0.9973745535472699, + "val_auc": 1.0, + "k": 5, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 37, + 277, + 7367 + ] + }, + { + "test_f1": 0.9940229506013034, + "test_acc": 0.9940229885057471, + "test_auc": 0.9996930637480383, + "val_auc": 1.0, + "k": 10, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 37, + 277, + 7367, + 304, + 30210, + 1404, + 160, + 5847 + ] + }, + { + "test_f1": 0.9885057471264368, + "test_acc": 0.9885057471264368, + "test_auc": 0.9981118769955084, + "val_auc": 1.0, + "k": 16, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 37, + 277, + 7367, + 304, + 30210, + 1404, + 160, + 5847, + 1867, + 613, + 419, + 1484, + 990, + 315 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..15f885875d319b84aa17fc964a4910064413fdb4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9793091551558699, + "test_acc": 0.9793103448275862, + "test_auc": 0.9898702581308513, + "val_auc": 1.0, + "k": 1, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439 + ] + }, + { + "test_f1": 0.9742512080111965, + "test_acc": 0.9742528735632184, + "test_auc": 0.989565858542129, + "val_auc": 1.0, + "k": 2, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 5907 + ] + }, + { + "test_f1": 0.9719530980989619, + "test_acc": 0.9719540229885058, + "test_auc": 0.9900554345473239, + "val_auc": 1.0, + "k": 5, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 5907, + 5397, + 1121, + 5672 + ] + }, + { + "test_f1": 0.974252710282917, + "test_acc": 0.9742528735632184, + "test_auc": 0.9896588695275719, + "val_auc": 1.0, + "k": 10, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 5907, + 5397, + 1121, + 5672, + 5535, + 1795, + 26232, + 879, + 493 + ] + }, + { + "test_f1": 0.9829879447485509, + "test_acc": 0.9829885057471265, + "test_auc": 0.9978336895935928, + "val_auc": 1.0, + "k": 16, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 5907, + 5397, + 1121, + 5672, + 5535, + 1795, + 26232, + 879, + 493, + 27328, + 2020, + 345, + 1913, + 1441, + 1426 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3ec175f49a375f9ad721bff075f7c778886c26e0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9016054099054172, + "test_acc": 0.9016091954022989, + "test_auc": 0.9377976351534175, + "val_auc": 0.913961038961039, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1934 + ] + }, + { + "test_f1": 0.9144765756895569, + "test_acc": 0.9144827586206896, + "test_auc": 0.951471095567942, + "val_auc": 0.9466314935064934, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1934, + 1866 + ] + }, + { + "test_f1": 0.9393072656974405, + "test_acc": 0.9393103448275862, + "test_auc": 0.9853664125493802, + "val_auc": 0.9681412337662337, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1934, + 1866, + 7889, + 1345, + 7037 + ] + }, + { + "test_f1": 0.9747103983960325, + "test_acc": 0.9747126436781609, + "test_auc": 0.9957747645976514, + "val_auc": 0.9939123376623377, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1934, + 1866, + 7889, + 1345, + 7037, + 1072, + 1682, + 508, + 1079, + 918 + ] + }, + { + "test_f1": 0.9811485884455626, + "test_acc": 0.9811494252873563, + "test_auc": 0.9976721887006872, + "val_auc": 0.997564935064935, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1934, + 1866, + 7889, + 1345, + 7037, + 1072, + 1682, + 508, + 1079, + 918, + 1023, + 22196, + 415, + 1715, + 7981, + 165 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3564c4bcfd1f5755423cad92b58bb09f9244ba9b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5624346157850625, + "test_acc": 0.5761006289308176, + "test_auc": 0.587181620168925, + "val_auc": 0.6082432973189276, + "k": 1, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.5749410106506435, + "test_acc": 0.5778616352201258, + "test_auc": 0.6106158845372183, + "val_auc": 0.6442577030812324, + "k": 2, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556 + ] + }, + { + "test_f1": 0.5776259816595296, + "test_acc": 0.5844025157232704, + "test_auc": 0.6312119065677406, + "val_auc": 0.6802721088435374, + "k": 5, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556, + 1279, + 31914, + 1148 + ] + }, + { + "test_f1": 0.607020957029741, + "test_acc": 0.6095597484276729, + "test_auc": 0.6445994791091795, + "val_auc": 0.6546618647458984, + "k": 10, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556, + 1279, + 31914, + 1148, + 835, + 1866, + 925, + 1928, + 1028 + ] + }, + { + "test_f1": 0.6368970245263221, + "test_acc": 0.6369811320754717, + "test_auc": 0.6772632523880069, + "val_auc": 0.7450980392156863, + "k": 16, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556, + 1279, + 31914, + 1148, + 835, + 1866, + 925, + 1928, + 1028, + 1698, + 238, + 581, + 7740, + 1351, + 28838 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a86bfd928942c8c001c527ae7395fed576213e2d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9374375402843803, + "test_acc": 0.9374712643678161, + "test_auc": 0.9630974687483087, + "val_auc": 0.9756493506493507, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772 + ] + }, + { + "test_f1": 0.960902431971928, + "test_acc": 0.960919540229885, + "test_auc": 0.9760065479733753, + "val_auc": 1.0, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963 + ] + }, + { + "test_f1": 0.9475246414887794, + "test_acc": 0.9475862068965517, + "test_auc": 0.9746883286433249, + "val_auc": 0.9987824675324676, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963, + 1942, + 217, + 352 + ] + }, + { + "test_f1": 0.9512187684161769, + "test_acc": 0.9512643678160919, + "test_auc": 0.989386601006548, + "val_auc": 0.9991883116883117, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963, + 1942, + 217, + 352, + 1877, + 684, + 1682, + 111, + 1047 + ] + }, + { + "test_f1": 0.9512187684161769, + "test_acc": 0.9512643678160919, + "test_auc": 0.9893857554521349, + "val_auc": 0.9991883116883117, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963, + 1942, + 217, + 352, + 1877, + 684, + 1682, + 111, + 1047, + 1898, + 847, + 1726, + 760, + 835, + 3243 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..90f037cca016fa49d4124bd016267eb128ed05bb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9862068615632066, + "test_acc": 0.9862068965517241, + "test_auc": 0.9986957323177661, + "val_auc": 1.0, + "k": 1, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10353 + ] + }, + { + "test_f1": 0.9871264041256593, + "test_acc": 0.9871264367816092, + "test_auc": 0.9984238865739488, + "val_auc": 1.0, + "k": 2, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10353, + 1675 + ] + }, + { + "test_f1": 0.9889654472641072, + "test_acc": 0.9889655172413793, + "test_auc": 0.9992618309973483, + "val_auc": 1.0, + "k": 5, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10353, + 1675, + 280, + 1682, + 165 + ] + }, + { + "test_f1": 0.9912643308837131, + "test_acc": 0.991264367816092, + "test_auc": 0.9993649886357487, + "val_auc": 1.0, + "k": 10, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10353, + 1675, + 280, + 1682, + 165, + 596, + 1588, + 913, + 491, + 5260 + ] + }, + { + "test_f1": 0.9912639800113583, + "test_acc": 0.991264367816092, + "test_auc": 0.9995958249905298, + "val_auc": 1.0, + "k": 16, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10353, + 1675, + 280, + 1682, + 165, + 596, + 1588, + 913, + 491, + 5260, + 1181, + 1278, + 1674, + 1331, + 361, + 1079 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9237ec8c9c2eefd3c2f34cf8f258fc1314885b92 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8809089669922265, + "test_acc": 0.8809195402298851, + "test_auc": 0.922040728664971, + "val_auc": 0.9610389610389611, + "k": 1, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 760 + ] + }, + { + "test_f1": 0.9342415863620912, + "test_acc": 0.9342528735632184, + "test_auc": 0.9752159545971102, + "val_auc": 0.9837662337662337, + "k": 2, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 760, + 2020 + ] + }, + { + "test_f1": 0.9484920281010593, + "test_acc": 0.9485057471264368, + "test_auc": 0.9897231316629689, + "val_auc": 0.989448051948052, + "k": 5, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 760, + 2020, + 1682, + 913, + 170 + ] + }, + { + "test_f1": 0.9563066917331932, + "test_acc": 0.9563218390804598, + "test_auc": 0.9913635072244169, + "val_auc": 0.9902597402597403, + "k": 10, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 760, + 2020, + 1682, + 913, + 170, + 2084, + 1482, + 1928, + 2023, + 6935 + ] + }, + { + "test_f1": 0.9668896086049882, + "test_acc": 0.9668965517241379, + "test_auc": 0.9950061556361276, + "val_auc": 0.9931006493506493, + "k": 16, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 760, + 2020, + 1682, + 913, + 170, + 2084, + 1482, + 1928, + 2023, + 6935, + 2817, + 1613, + 49, + 930, + 3842, + 13042 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fe57ed9c030cedca47f507e5555109519d1e29f4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9439078088318119, + "test_acc": 0.9439080459770115, + "test_auc": 0.9629875466746036, + "val_auc": 0.966314935064935, + "k": 1, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1584 + ] + }, + { + "test_f1": 0.9645961295197151, + "test_acc": 0.9645977011494253, + "test_auc": 0.9762914998105958, + "val_auc": 0.9569805194805194, + "k": 2, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1584, + 17123 + ] + }, + { + "test_f1": 0.9678123361069447, + "test_acc": 0.967816091954023, + "test_auc": 0.9773332228475566, + "val_auc": 0.9557629870129871, + "k": 5, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1584, + 17123, + 7469, + 1763, + 1491 + ] + }, + { + "test_f1": 0.9678109478589979, + "test_acc": 0.967816091954023, + "test_auc": 0.986107540992478, + "val_auc": 0.992288961038961, + "k": 10, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1584, + 17123, + 7469, + 1763, + 1491, + 24647, + 866, + 1682, + 1457, + 104 + ] + }, + { + "test_f1": 0.9664348516972999, + "test_acc": 0.9664367816091954, + "test_auc": 0.9936921640781428, + "val_auc": 0.9983766233766234, + "k": 16, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1584, + 17123, + 7469, + 1763, + 1491, + 24647, + 866, + 1682, + 1457, + 104, + 204, + 523, + 2014, + 687, + 3573, + 1267 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..80b3f430c15eb40cb184fb47f80625ea92067474 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8865541093263206, + "test_acc": 0.886896551724138, + "test_auc": 0.9245909207749337, + "val_auc": 0.9078733766233767, + "k": 1, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967 + ] + }, + { + "test_f1": 0.888929108135109, + "test_acc": 0.8891954022988505, + "test_auc": 0.9398616334758374, + "val_auc": 0.9383116883116883, + "k": 2, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352 + ] + }, + { + "test_f1": 0.9300869812701064, + "test_acc": 0.9301149425287356, + "test_auc": 0.9738258631419447, + "val_auc": 0.9748376623376622, + "k": 5, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 1682, + 1070, + 182 + ] + }, + { + "test_f1": 0.9319158552611351, + "test_acc": 0.9319540229885057, + "test_auc": 0.9759279114129551, + "val_auc": 0.984172077922078, + "k": 10, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 1682, + 1070, + 182, + 1635, + 960, + 1168, + 178, + 859 + ] + }, + { + "test_f1": 0.9494086041639616, + "test_acc": 0.9494252873563218, + "test_auc": 0.9854015030575247, + "val_auc": 0.9983766233766234, + "k": 16, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 1682, + 1070, + 182, + 1635, + 960, + 1168, + 178, + 859, + 1684, + 1797, + 662, + 1264, + 5437, + 61 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c0b2bad12ea101d67ac66a0b11d46c468af7f131 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8995513701272617, + "test_acc": 0.8997701149425288, + "test_auc": 0.9426680285729747, + "val_auc": 0.948051948051948, + "k": 1, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335 + ] + }, + { + "test_f1": 0.9172390700135065, + "test_acc": 0.9172413793103448, + "test_auc": 0.9807501082309649, + "val_auc": 0.984577922077922, + "k": 2, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 31626 + ] + }, + { + "test_f1": 0.9388491527908102, + "test_acc": 0.9388505747126437, + "test_auc": 0.9869353387629201, + "val_auc": 0.9922889610389611, + "k": 5, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 31626, + 1443, + 1682, + 835 + ] + }, + { + "test_f1": 0.9641372487480045, + "test_acc": 0.9641379310344828, + "test_auc": 0.9940286947345636, + "val_auc": 0.9963474025974026, + "k": 10, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 31626, + 1443, + 1682, + 835, + 878, + 1079, + 1515, + 760, + 315 + ] + }, + { + "test_f1": 0.9632179553882518, + "test_acc": 0.9632183908045977, + "test_auc": 0.9951042399480492, + "val_auc": 0.9991883116883117, + "k": 16, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 31626, + 1443, + 1682, + 835, + 878, + 1079, + 1515, + 760, + 315, + 21, + 1233, + 31176, + 5792, + 335, + 1928 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3ed43f31ec3a16603a2eb1b42b0c2a99f72e4863 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6915873810727228, + "test_acc": 0.6968553459119496, + "test_auc": 0.7112854530302095, + "val_auc": 0.7148859543817526, + "k": 1, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313 + ] + }, + { + "test_f1": 0.6894692432846873, + "test_acc": 0.6945911949685535, + "test_auc": 0.7242819017780564, + "val_auc": 0.7170868347338936, + "k": 2, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 1067 + ] + }, + { + "test_f1": 0.691635422833255, + "test_acc": 0.6940880503144654, + "test_auc": 0.7585924454629133, + "val_auc": 0.7274909963985594, + "k": 5, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 1067, + 1105, + 765, + 1019 + ] + }, + { + "test_f1": 0.7029564626457283, + "test_acc": 0.7046540880503145, + "test_auc": 0.7697422582804325, + "val_auc": 0.7458983593437375, + "k": 10, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 1067, + 1105, + 765, + 1019, + 4333, + 6718, + 121, + 821, + 24630 + ] + }, + { + "test_f1": 0.693487669494448, + "test_acc": 0.6953459119496855, + "test_auc": 0.7746818100348442, + "val_auc": 0.7627050820328131, + "k": 16, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 1067, + 1105, + 765, + 1019, + 4333, + 6718, + 121, + 821, + 24630, + 800, + 2424, + 10708, + 20445, + 142, + 4252 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f5771ac35924d6d0b23bcde7ba5ece7239170abe --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6863442499800385, + "test_acc": 0.6881878209831255, + "test_auc": 0.7452265299003967, + "val_auc": 0.7632211538461537, + "k": 1, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394 + ] + }, + { + "test_f1": 0.6958229327411024, + "test_acc": 0.6969919295671313, + "test_auc": 0.7474969963956749, + "val_auc": 0.7712339743589745, + "k": 2, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041 + ] + }, + { + "test_f1": 0.7005930249676962, + "test_acc": 0.7013939838591343, + "test_auc": 0.7512929493887289, + "val_auc": 0.7740384615384616, + "k": 5, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 11960, + 1602, + 5133 + ] + }, + { + "test_f1": 0.68677436056271, + "test_acc": 0.6874541452677916, + "test_auc": 0.7561708458752654, + "val_auc": 0.8008814102564102, + "k": 10, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 11960, + 1602, + 5133, + 1551, + 8604, + 13568, + 3632, + 821 + ] + }, + { + "test_f1": 0.6982185454878203, + "test_acc": 0.698459280997799, + "test_auc": 0.7731256001825846, + "val_auc": 0.8325320512820513, + "k": 16, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 11960, + 1602, + 5133, + 1551, + 8604, + 13568, + 3632, + 821, + 1257, + 126, + 31, + 1737, + 1608, + 5653 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..481c14fc815e2bff5dd9b01c87581f072d60ce77 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6852194279640934, + "test_acc": 0.6852830188679245, + "test_auc": 0.7336795053157394, + "val_auc": 0.7126850740296119, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1500 + ] + }, + { + "test_f1": 0.7169655478539049, + "test_acc": 0.7169811320754716, + "test_auc": 0.7777444738891324, + "val_auc": 0.7609043617446979, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1500, + 5653 + ] + }, + { + "test_f1": 0.7237228005066428, + "test_acc": 0.7240251572327044, + "test_auc": 0.8083675682681899, + "val_auc": 0.8073229291716686, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1500, + 5653, + 3801, + 1927, + 24630 + ] + }, + { + "test_f1": 0.7407277016913196, + "test_acc": 0.740880503144654, + "test_auc": 0.8279493265582423, + "val_auc": 0.8169267707082833, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1500, + 5653, + 3801, + 1927, + 24630, + 4674, + 3802, + 7258, + 1067, + 1149 + ] + }, + { + "test_f1": 0.7559585327588221, + "test_acc": 0.7559748427672957, + "test_auc": 0.8374790767757021, + "val_auc": 0.8515406162464986, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1500, + 5653, + 3801, + 1927, + 24630, + 4674, + 3802, + 7258, + 1067, + 1149, + 984, + 1180, + 16529, + 3245, + 27820, + 1551 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5fc1e774e4c3afc871d0bb63dbd78940e20b824e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7574184140992876, + "test_acc": 0.7574842767295598, + "test_auc": 0.7844324376049958, + "val_auc": 0.728891556622649, + "k": 1, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7032 + ] + }, + { + "test_f1": 0.7542125397673082, + "test_acc": 0.7544654088050314, + "test_auc": 0.8040259675820398, + "val_auc": 0.7406962785114046, + "k": 2, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7032, + 2472 + ] + }, + { + "test_f1": 0.7633424307784946, + "test_acc": 0.7635220125786164, + "test_auc": 0.8306173224551131, + "val_auc": 0.8237294917967186, + "k": 5, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7032, + 2472, + 1503, + 1134, + 7615 + ] + }, + { + "test_f1": 0.7635217132507586, + "test_acc": 0.7635220125786164, + "test_auc": 0.843021516112275, + "val_auc": 0.8415366146458583, + "k": 10, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7032, + 2472, + 1503, + 1134, + 7615, + 1928, + 5891, + 18220, + 1643, + 2471 + ] + }, + { + "test_f1": 0.7717081680003076, + "test_acc": 0.7718238993710692, + "test_auc": 0.849218739715596, + "val_auc": 0.851140456182473, + "k": 16, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7032, + 2472, + 1503, + 1134, + 7615, + 1928, + 5891, + 18220, + 1643, + 2471, + 349, + 32397, + 1048, + 24597, + 25162, + 17225 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5b8518d41147bcab7dbf6672ae42b10dbfb2aa58 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7056264149284225, + "test_acc": 0.7242767295597484, + "test_auc": 0.7260624643684959, + "val_auc": 0.7857142857142857, + "k": 1, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562 + ] + }, + { + "test_f1": 0.7874138807595709, + "test_acc": 0.7922012578616352, + "test_auc": 0.8016388466683342, + "val_auc": 0.8405362144857944, + "k": 2, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562, + 1438 + ] + }, + { + "test_f1": 0.8108374368360629, + "test_acc": 0.8138364779874214, + "test_auc": 0.888911982210323, + "val_auc": 0.9327731092436975, + "k": 5, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562, + 1438, + 7284, + 186, + 4675 + ] + }, + { + "test_f1": 0.820969678376029, + "test_acc": 0.8218867924528301, + "test_auc": 0.9119352501521459, + "val_auc": 0.9583833533413365, + "k": 10, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562, + 1438, + 7284, + 186, + 4675, + 100, + 10076, + 505, + 11157, + 1335 + ] + }, + { + "test_f1": 0.8438355715410555, + "test_acc": 0.8440251572327044, + "test_auc": 0.9273248955231135, + "val_auc": 0.9427771108443378, + "k": 16, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1562, + 1438, + 7284, + 186, + 4675, + 100, + 10076, + 505, + 11157, + 1335, + 794, + 152, + 253, + 1417, + 821, + 210 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e2c32df23f3a162e82dee5606b7838daab516a3c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7517372829953665, + "test_acc": 0.7529559748427673, + "test_auc": 0.8022247728950451, + "val_auc": 0.8471388555422169, + "k": 1, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1335 + ] + }, + { + "test_f1": 0.7897146252760728, + "test_acc": 0.790440251572327, + "test_auc": 0.8895238567793273, + "val_auc": 0.912765106042417, + "k": 2, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1335, + 444 + ] + }, + { + "test_f1": 0.8972585762361887, + "test_acc": 0.8976100628930818, + "test_auc": 0.9465132516285433, + "val_auc": 0.9581832733093237, + "k": 5, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1335, + 444, + 1116, + 18480, + 6553 + ] + }, + { + "test_f1": 0.9161848701275295, + "test_acc": 0.9162264150943397, + "test_auc": 0.9691467881268487, + "val_auc": 0.9891956782713085, + "k": 10, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1335, + 444, + 1116, + 18480, + 6553, + 1928, + 988, + 3481, + 7666, + 1981 + ] + }, + { + "test_f1": 0.9272864856701901, + "test_acc": 0.9272955974842767, + "test_auc": 0.9815906004724877, + "val_auc": 0.9959983993597439, + "k": 16, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1335, + 444, + 1116, + 18480, + 6553, + 1928, + 988, + 3481, + 7666, + 1981, + 32550, + 894, + 1960, + 993, + 821, + 6027 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d68761cb9c137ac3eec9051670a64426ffd1a80d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7688323917137476, + "test_acc": 0.7777777777777778, + "test_auc": 0.7751020408163266, + "val_auc": 0.7508012820512819, + "k": 1, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629 + ] + }, + { + "test_f1": 0.7365678057167419, + "test_acc": 0.7373737373737373, + "test_auc": 0.8077551020408162, + "val_auc": 0.8253205128205128, + "k": 2, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625 + ] + }, + { + "test_f1": 0.7572786690433749, + "test_acc": 0.7575757575757576, + "test_auc": 0.8359183673469387, + "val_auc": 0.8587740384615384, + "k": 5, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 530, + 567 + ] + }, + { + "test_f1": 0.7777777777777778, + "test_acc": 0.7777777777777778, + "test_auc": 0.8689795918367347, + "val_auc": 0.8697916666666666, + "k": 10, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 530, + 567, + 794, + 6400, + 2046, + 680, + 7669 + ] + }, + { + "test_f1": 0.8382848088730442, + "test_acc": 0.8383838383838383, + "test_auc": 0.8979591836734693, + "val_auc": 0.921474358974359, + "k": 16, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 530, + 567, + 794, + 6400, + 2046, + 680, + 7669, + 1398, + 1758, + 920, + 777, + 400, + 1182 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..67c22dd719876081a8f0b603d8a1c48b9ef3dca2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8464525565993019, + "test_acc": 0.8464818763326226, + "test_auc": 0.9066648481542099, + "val_auc": 0.9376, + "k": 1, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417 + ] + }, + { + "test_f1": 0.9039935554844599, + "test_acc": 0.9040511727078892, + "test_auc": 0.9563738861611202, + "val_auc": 0.9634, + "k": 2, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417, + 82 + ] + }, + { + "test_f1": 0.9253283209323577, + "test_acc": 0.9253731343283582, + "test_auc": 0.9720494635388252, + "val_auc": 0.9868000000000001, + "k": 5, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417, + 82, + 12485, + 2855, + 28188 + ] + }, + { + "test_f1": 0.95948606263797, + "test_acc": 0.9594882729211087, + "test_auc": 0.9903436988543372, + "val_auc": 0.9904, + "k": 10, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417, + 82, + 12485, + 2855, + 28188, + 32655, + 20154, + 1689, + 330, + 320 + ] + }, + { + "test_f1": 0.9658848614072495, + "test_acc": 0.9658848614072495, + "test_auc": 0.9931260229132568, + "val_auc": 0.9892, + "k": 16, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417, + 82, + 12485, + 2855, + 28188, + 32655, + 20154, + 1689, + 330, + 320, + 567, + 1301, + 1566, + 32722, + 9661, + 1927 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/saebench_sparse_probing_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/saebench_sparse_probing_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..c2ae6ae57054c95131e8ae43704f19773a281f27 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/saebench_sparse_probing_custom_sae_eval_results.json @@ -0,0 +1,782 @@ +{ + "eval_type_id": "sparse_probing", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "LabHC/bias_in_bios_class_set2", + "LabHC/bias_in_bios_class_set3", + "canrager/amazon_reviews_mcauley_1and5", + "canrager/amazon_reviews_mcauley_1and5_sentiment", + "codeparrot/github-code", + "fancyzhx/ag_news", + "Helsinki-NLP/europarl" + ], + "probe_train_set_size": 4000, + "probe_test_set_size": 1000, + "context_length": 128, + "sae_batch_size": 125, + "llm_batch_size": 128, + "llm_dtype": "float32", + "model_name": "gemma-2-2b", + "k_values": [ + 1, + 2, + 5, + 10 + ], + "lower_vram_usage": false + }, + "eval_id": "81b37442-d842-4673-82be-4ced479caa46", + "datetime_epoch_millis": 1776879657130, + "eval_result_metrics": { + "llm": { + "llm_test_accuracy": 0.9584312926977874, + "llm_top_1_test_accuracy": 0.66844375, + "llm_top_2_test_accuracy": 0.71781875, + "llm_top_5_test_accuracy": 0.786, + "llm_top_10_test_accuracy": 0.83568125, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null + }, + "sae": { + "sae_test_accuracy": 0.9568187899887561, + "sae_top_1_test_accuracy": 0.72328125, + "sae_top_2_test_accuracy": 0.7752875, + "sae_top_5_test_accuracy": 0.8784937500000001, + "sae_top_10_test_accuracy": 0.9060812500000001, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_results", + "llm_test_accuracy": 0.9674000382423401, + "llm_top_1_test_accuracy": 0.6487999999999999, + "llm_top_2_test_accuracy": 0.6956, + "llm_top_5_test_accuracy": 0.7882, + "llm_top_10_test_accuracy": 0.8400000000000001, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9666000485420227, + "sae_top_1_test_accuracy": 0.7916000000000001, + "sae_top_2_test_accuracy": 0.7999999999999999, + "sae_top_5_test_accuracy": 0.8832000000000001, + "sae_top_10_test_accuracy": 0.8984, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set2_results", + "llm_test_accuracy": 0.953600037097931, + "llm_top_1_test_accuracy": 0.6714, + "llm_top_2_test_accuracy": 0.7110000000000001, + "llm_top_5_test_accuracy": 0.7674, + "llm_top_10_test_accuracy": 0.807, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9478000402450562, + "sae_top_1_test_accuracy": 0.6968, + "sae_top_2_test_accuracy": 0.736, + "sae_top_5_test_accuracy": 0.8556000000000001, + "sae_top_10_test_accuracy": 0.9028, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set3_results", + "llm_test_accuracy": 0.9302000522613525, + "llm_top_1_test_accuracy": 0.6878, + "llm_top_2_test_accuracy": 0.7394000000000001, + "llm_top_5_test_accuracy": 0.7634, + "llm_top_10_test_accuracy": 0.8048, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9286000370979309, + "sae_top_1_test_accuracy": 0.6926, + "sae_top_2_test_accuracy": 0.734, + "sae_top_5_test_accuracy": 0.8418000000000001, + "sae_top_10_test_accuracy": 0.8777999999999999, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_results", + "llm_test_accuracy": 0.9136000394821167, + "llm_top_1_test_accuracy": 0.5955999999999999, + "llm_top_2_test_accuracy": 0.6424, + "llm_top_5_test_accuracy": 0.7004, + "llm_top_10_test_accuracy": 0.7674, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9192000508308411, + "sae_top_1_test_accuracy": 0.7085999999999999, + "sae_top_2_test_accuracy": 0.7382, + "sae_top_5_test_accuracy": 0.8218, + "sae_top_10_test_accuracy": 0.8426, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_sentiment_results", + "llm_test_accuracy": 0.9810000658035278, + "llm_top_1_test_accuracy": 0.674, + "llm_top_2_test_accuracy": 0.724, + "llm_top_5_test_accuracy": 0.764, + "llm_top_10_test_accuracy": 0.827, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9745000302791595, + "sae_top_1_test_accuracy": 0.544, + "sae_top_2_test_accuracy": 0.731, + "sae_top_5_test_accuracy": 0.924, + "sae_top_10_test_accuracy": 0.9424999999999999, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "codeparrot/github-code_results", + "llm_test_accuracy": 0.9700000405311584, + "llm_top_1_test_accuracy": 0.6452, + "llm_top_2_test_accuracy": 0.6862000000000001, + "llm_top_5_test_accuracy": 0.7644, + "llm_top_10_test_accuracy": 0.8066000000000001, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9690000414848328, + "sae_top_1_test_accuracy": 0.6437999999999999, + "sae_top_2_test_accuracy": 0.7256, + "sae_top_5_test_accuracy": 0.8414000000000001, + "sae_top_10_test_accuracy": 0.8876, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "fancyzhx/ag_news_results", + "llm_test_accuracy": 0.9522500485181808, + "llm_top_1_test_accuracy": 0.69575, + "llm_top_2_test_accuracy": 0.76175, + "llm_top_5_test_accuracy": 0.831, + "llm_top_10_test_accuracy": 0.87025, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9502500295639038, + "sae_top_1_test_accuracy": 0.81525, + "sae_top_2_test_accuracy": 0.8394999999999999, + "sae_top_5_test_accuracy": 0.87375, + "sae_top_10_test_accuracy": 0.9047499999999999, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "Helsinki-NLP/europarl_results", + "llm_test_accuracy": 0.9994000196456909, + "llm_top_1_test_accuracy": 0.729, + "llm_top_2_test_accuracy": 0.7822, + "llm_top_5_test_accuracy": 0.9092, + "llm_top_10_test_accuracy": 0.9624, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9986000418663025, + "sae_top_1_test_accuracy": 0.8936, + "sae_top_2_test_accuracy": 0.898, + "sae_top_5_test_accuracy": 0.9863999999999999, + "sae_top_10_test_accuracy": 0.9922000000000001, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_sparse_probing", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1_results": { + "sae_test_accuracy": { + "0": 0.9490000605583191, + "1": 0.9640000462532043, + "2": 0.9500000476837158, + "6": 0.9900000691413879, + "9": 0.9800000190734863 + }, + "llm_test_accuracy": { + "0": 0.9480000734329224, + "1": 0.9650000333786011, + "2": 0.9510000348091125, + "6": 0.9930000305175781, + "9": 0.9800000190734863 + }, + "llm_top_1_test_accuracy": { + "0": 0.586, + "1": 0.633, + "2": 0.689, + "6": 0.803, + "9": 0.533 + }, + "llm_top_2_test_accuracy": { + "0": 0.587, + "1": 0.645, + "2": 0.716, + "6": 0.82, + "9": 0.71 + }, + "llm_top_5_test_accuracy": { + "0": 0.72, + "1": 0.707, + "2": 0.768, + "6": 0.891, + "9": 0.855 + }, + "llm_top_10_test_accuracy": { + "0": 0.788, + "1": 0.796, + "2": 0.809, + "6": 0.902, + "9": 0.905 + }, + "sae_top_1_test_accuracy": { + "0": 0.562, + "1": 0.629, + "2": 0.86, + "6": 0.973, + "9": 0.934 + }, + "sae_top_2_test_accuracy": { + "0": 0.602, + "1": 0.617, + "2": 0.865, + "6": 0.976, + "9": 0.94 + }, + "sae_top_5_test_accuracy": { + "0": 0.845, + "1": 0.794, + "2": 0.86, + "6": 0.981, + "9": 0.936 + }, + "sae_top_10_test_accuracy": { + "0": 0.86, + "1": 0.811, + "2": 0.892, + "6": 0.978, + "9": 0.951 + } + }, + "LabHC/bias_in_bios_class_set2_results": { + "sae_test_accuracy": { + "11": 0.9600000381469727, + "13": 0.9490000605583191, + "14": 0.9340000152587891, + "18": 0.9300000667572021, + "19": 0.9660000205039978 + }, + "llm_test_accuracy": { + "11": 0.9630000591278076, + "13": 0.9520000219345093, + "14": 0.9650000333786011, + "18": 0.9330000281333923, + "19": 0.9550000429153442 + }, + "llm_top_1_test_accuracy": { + "11": 0.567, + "13": 0.682, + "14": 0.627, + "18": 0.693, + "19": 0.788 + }, + "llm_top_2_test_accuracy": { + "11": 0.682, + "13": 0.709, + "14": 0.676, + "18": 0.718, + "19": 0.77 + }, + "llm_top_5_test_accuracy": { + "11": 0.793, + "13": 0.75, + "14": 0.719, + "18": 0.731, + "19": 0.844 + }, + "llm_top_10_test_accuracy": { + "11": 0.838, + "13": 0.79, + "14": 0.789, + "18": 0.755, + "19": 0.863 + }, + "sae_top_1_test_accuracy": { + "11": 0.597, + "13": 0.677, + "14": 0.671, + "18": 0.698, + "19": 0.841 + }, + "sae_top_2_test_accuracy": { + "11": 0.734, + "13": 0.666, + "14": 0.732, + "18": 0.701, + "19": 0.847 + }, + "sae_top_5_test_accuracy": { + "11": 0.891, + "13": 0.745, + "14": 0.866, + "18": 0.887, + "19": 0.889 + }, + "sae_top_10_test_accuracy": { + "11": 0.929, + "13": 0.9, + "14": 0.911, + "18": 0.898, + "19": 0.876 + } + }, + "LabHC/bias_in_bios_class_set3_results": { + "sae_test_accuracy": { + "20": 0.9570000171661377, + "21": 0.9260000586509705, + "22": 0.9100000262260437, + "25": 0.9600000381469727, + "26": 0.89000004529953 + }, + "llm_test_accuracy": { + "20": 0.9600000381469727, + "21": 0.9290000200271606, + "22": 0.9110000729560852, + "25": 0.9630000591278076, + "26": 0.8880000710487366 + }, + "llm_top_1_test_accuracy": { + "20": 0.698, + "21": 0.77, + "22": 0.641, + "25": 0.699, + "26": 0.631 + }, + "llm_top_2_test_accuracy": { + "20": 0.809, + "21": 0.774, + "22": 0.673, + "25": 0.771, + "26": 0.67 + }, + "llm_top_5_test_accuracy": { + "20": 0.818, + "21": 0.797, + "22": 0.706, + "25": 0.811, + "26": 0.685 + }, + "llm_top_10_test_accuracy": { + "20": 0.88, + "21": 0.828, + "22": 0.75, + "25": 0.843, + "26": 0.723 + }, + "sae_top_1_test_accuracy": { + "20": 0.823, + "21": 0.667, + "22": 0.461, + "25": 0.865, + "26": 0.647 + }, + "sae_top_2_test_accuracy": { + "20": 0.851, + "21": 0.694, + "22": 0.617, + "25": 0.862, + "26": 0.646 + }, + "sae_top_5_test_accuracy": { + "20": 0.915, + "21": 0.793, + "22": 0.888, + "25": 0.873, + "26": 0.74 + }, + "sae_top_10_test_accuracy": { + "20": 0.934, + "21": 0.837, + "22": 0.893, + "25": 0.924, + "26": 0.801 + } + }, + "canrager/amazon_reviews_mcauley_1and5_results": { + "sae_test_accuracy": { + "1": 0.9500000476837158, + "2": 0.9300000667572021, + "3": 0.9260000586509705, + "5": 0.9150000214576721, + "6": 0.8750000596046448 + }, + "llm_test_accuracy": { + "1": 0.9530000686645508, + "2": 0.9290000200271606, + "3": 0.9200000166893005, + "5": 0.9110000729560852, + "6": 0.8550000190734863 + }, + "llm_top_1_test_accuracy": { + "1": 0.661, + "2": 0.596, + "3": 0.574, + "5": 0.567, + "6": 0.58 + }, + "llm_top_2_test_accuracy": { + "1": 0.739, + "2": 0.617, + "3": 0.607, + "5": 0.628, + "6": 0.621 + }, + "llm_top_5_test_accuracy": { + "1": 0.781, + "2": 0.739, + "3": 0.681, + "5": 0.636, + "6": 0.665 + }, + "llm_top_10_test_accuracy": { + "1": 0.823, + "2": 0.79, + "3": 0.751, + "5": 0.739, + "6": 0.734 + }, + "sae_top_1_test_accuracy": { + "1": 0.839, + "2": 0.611, + "3": 0.572, + "5": 0.74, + "6": 0.781 + }, + "sae_top_2_test_accuracy": { + "1": 0.837, + "2": 0.649, + "3": 0.572, + "5": 0.852, + "6": 0.781 + }, + "sae_top_5_test_accuracy": { + "1": 0.913, + "2": 0.876, + "3": 0.657, + "5": 0.879, + "6": 0.784 + }, + "sae_top_10_test_accuracy": { + "1": 0.925, + "2": 0.892, + "3": 0.74, + "5": 0.872, + "6": 0.784 + } + }, + "canrager/amazon_reviews_mcauley_1and5_sentiment_results": { + "sae_test_accuracy": { + "1.0": 0.9750000238418579, + "5.0": 0.9740000367164612 + }, + "llm_test_accuracy": { + "1.0": 0.9810000658035278, + "5.0": 0.9810000658035278 + }, + "llm_top_1_test_accuracy": { + "1.0": 0.674, + "5.0": 0.674 + }, + "llm_top_2_test_accuracy": { + "1.0": 0.724, + "5.0": 0.724 + }, + "llm_top_5_test_accuracy": { + "1.0": 0.764, + "5.0": 0.764 + }, + "llm_top_10_test_accuracy": { + "1.0": 0.827, + "5.0": 0.827 + }, + "sae_top_1_test_accuracy": { + "1.0": 0.544, + "5.0": 0.544 + }, + "sae_top_2_test_accuracy": { + "1.0": 0.731, + "5.0": 0.731 + }, + "sae_top_5_test_accuracy": { + "1.0": 0.924, + "5.0": 0.924 + }, + "sae_top_10_test_accuracy": { + "1.0": 0.942, + "5.0": 0.943 + } + }, + "codeparrot/github-code_results": { + "sae_test_accuracy": { + "C": 0.9530000686645508, + "Python": 0.9830000400543213, + "HTML": 0.9890000224113464, + "Java": 0.9650000333786011, + "PHP": 0.9550000429153442 + }, + "llm_test_accuracy": { + "C": 0.9530000686645508, + "Python": 0.9890000224113464, + "HTML": 0.9900000691413879, + "Java": 0.9610000252723694, + "PHP": 0.9570000171661377 + }, + "llm_top_1_test_accuracy": { + "C": 0.668, + "Python": 0.633, + "HTML": 0.714, + "Java": 0.634, + "PHP": 0.577 + }, + "llm_top_2_test_accuracy": { + "C": 0.668, + "Python": 0.674, + "HTML": 0.785, + "Java": 0.683, + "PHP": 0.621 + }, + "llm_top_5_test_accuracy": { + "C": 0.756, + "Python": 0.732, + "HTML": 0.896, + "Java": 0.734, + "PHP": 0.704 + }, + "llm_top_10_test_accuracy": { + "C": 0.805, + "Python": 0.743, + "HTML": 0.948, + "Java": 0.802, + "PHP": 0.735 + }, + "sae_top_1_test_accuracy": { + "C": 0.66, + "Python": 0.632, + "HTML": 0.709, + "Java": 0.616, + "PHP": 0.602 + }, + "sae_top_2_test_accuracy": { + "C": 0.645, + "Python": 0.64, + "HTML": 0.908, + "Java": 0.654, + "PHP": 0.781 + }, + "sae_top_5_test_accuracy": { + "C": 0.707, + "Python": 0.879, + "HTML": 0.917, + "Java": 0.781, + "PHP": 0.923 + }, + "sae_top_10_test_accuracy": { + "C": 0.879, + "Python": 0.894, + "HTML": 0.953, + "Java": 0.792, + "PHP": 0.92 + } + }, + "fancyzhx/ag_news_results": { + "sae_test_accuracy": { + "0": 0.937000036239624, + "1": 0.984000027179718, + "2": 0.9290000200271606, + "3": 0.9510000348091125 + }, + "llm_test_accuracy": { + "0": 0.9360000491142273, + "1": 0.9860000610351562, + "2": 0.9320000410079956, + "3": 0.9550000429153442 + }, + "llm_top_1_test_accuracy": { + "0": 0.809, + "1": 0.671, + "2": 0.67, + "3": 0.633 + }, + "llm_top_2_test_accuracy": { + "0": 0.804, + "1": 0.804, + "2": 0.702, + "3": 0.737 + }, + "llm_top_5_test_accuracy": { + "0": 0.826, + "1": 0.883, + "2": 0.764, + "3": 0.851 + }, + "llm_top_10_test_accuracy": { + "0": 0.863, + "1": 0.933, + "2": 0.801, + "3": 0.884 + }, + "sae_top_1_test_accuracy": { + "0": 0.848, + "1": 0.941, + "2": 0.795, + "3": 0.677 + }, + "sae_top_2_test_accuracy": { + "0": 0.856, + "1": 0.963, + "2": 0.815, + "3": 0.724 + }, + "sae_top_5_test_accuracy": { + "0": 0.873, + "1": 0.964, + "2": 0.841, + "3": 0.817 + }, + "sae_top_10_test_accuracy": { + "0": 0.907, + "1": 0.968, + "2": 0.871, + "3": 0.873 + } + }, + "Helsinki-NLP/europarl_results": { + "sae_test_accuracy": { + "en": 0.9980000257492065, + "fr": 1.0, + "de": 0.999000072479248, + "es": 0.999000072479248, + "nl": 0.9970000386238098 + }, + "llm_test_accuracy": { + "en": 1.0, + "fr": 1.0, + "de": 1.0, + "es": 0.999000072479248, + "nl": 0.9980000257492065 + }, + "llm_top_1_test_accuracy": { + "en": 0.755, + "fr": 0.59, + "de": 0.751, + "es": 0.907, + "nl": 0.642 + }, + "llm_top_2_test_accuracy": { + "en": 0.825, + "fr": 0.593, + "de": 0.827, + "es": 0.916, + "nl": 0.75 + }, + "llm_top_5_test_accuracy": { + "en": 0.891, + "fr": 0.899, + "de": 0.919, + "es": 0.978, + "nl": 0.859 + }, + "llm_top_10_test_accuracy": { + "en": 0.978, + "fr": 0.966, + "de": 0.963, + "es": 0.99, + "nl": 0.915 + }, + "sae_top_1_test_accuracy": { + "en": 1.0, + "fr": 0.995, + "de": 0.903, + "es": 0.889, + "nl": 0.681 + }, + "sae_top_2_test_accuracy": { + "en": 0.999, + "fr": 0.995, + "de": 0.915, + "es": 0.891, + "nl": 0.69 + }, + "sae_top_5_test_accuracy": { + "en": 0.999, + "fr": 0.997, + "de": 0.945, + "es": 0.993, + "nl": 0.998 + }, + "sae_top_10_test_accuracy": { + "en": 0.998, + "fr": 0.994, + "de": 0.975, + "es": 0.995, + "nl": 0.999 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/scr/saebench_scr_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/scr/saebench_scr_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..2c867b4426128bcbf7bb09bb7f9cd324c6894cbb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/scr/saebench_scr_custom_sae_eval_results.json @@ -0,0 +1,326 @@ +{ + "eval_type_id": "scr", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": true, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "a2c1f281-07fc-4e96-b2fe-ee93ca2cd5ef", + "datetime_epoch_millis": 1776879209120, + "eval_result_metrics": { + "scr_metrics": { + "scr_dir1_threshold_2": 0.1692460812131599, + "scr_metric_threshold_2": 0.19027546592797184, + "scr_dir2_threshold_2": 0.191625049754217, + "scr_dir1_threshold_5": 0.22611947159445542, + "scr_metric_threshold_5": 0.2388952041052821, + "scr_dir2_threshold_5": 0.24414094159052138, + "scr_dir1_threshold_10": 0.3033770862103188, + "scr_metric_threshold_10": 0.3157847187706869, + "scr_dir2_threshold_10": 0.3202438683368537, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results", + "scr_dir1_threshold_2": 0.4487182328294704, + "scr_metric_threshold_2": 0.02849734852006005, + "scr_dir2_threshold_2": 0.02849734852006005, + "scr_dir1_threshold_5": 0.5384614208981502, + "scr_metric_threshold_5": 0.06994820573050735, + "scr_dir2_threshold_5": 0.06994820573050735, + "scr_dir1_threshold_10": 0.5641028776049328, + "scr_metric_threshold_10": 0.17098439995268688, + "scr_dir2_threshold_10": 0.17098439995268688, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results", + "scr_dir1_threshold_2": 0.16666674192503994, + "scr_metric_threshold_2": 0.339449522889458, + "scr_dir2_threshold_2": 0.339449522889458, + "scr_dir1_threshold_5": 0.23484857036936357, + "scr_metric_threshold_5": 0.3700305314291443, + "scr_dir2_threshold_5": 0.3700305314291443, + "scr_dir1_threshold_10": 0.3030303988136872, + "scr_metric_threshold_10": 0.3883792823747072, + "scr_dir2_threshold_10": 0.3883792823747072, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results", + "scr_dir1_threshold_2": 0.3437501746229541, + "scr_metric_threshold_2": 0.047738670514110945, + "scr_dir2_threshold_2": 0.047738670514110945, + "scr_dir1_threshold_5": 0.3906245634426147, + "scr_metric_threshold_5": 0.05778894058451181, + "scr_dir2_threshold_5": 0.05778894058451181, + "scr_dir1_threshold_10": 0.5156247962732202, + "scr_metric_threshold_10": 0.14321616130272372, + "scr_dir2_threshold_10": 0.14321616130272372, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results", + "scr_dir1_threshold_2": 0.14093939327267527, + "scr_metric_threshold_2": 0.04777065348328394, + "scr_dir2_threshold_2": 0.04777065348328394, + "scr_dir1_threshold_5": 0.20134210468416536, + "scr_metric_threshold_5": 0.11783439248632226, + "scr_dir2_threshold_5": 0.11783439248632226, + "scr_dir1_threshold_10": 0.32885885099037726, + "scr_metric_threshold_10": 0.19745233811549862, + "scr_dir2_threshold_10": 0.19745233811549862, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results", + "scr_dir1_threshold_2": 0.03157916861558172, + "scr_metric_threshold_2": 0.5119047252994956, + "scr_dir2_threshold_2": 0.5119047252994956, + "scr_dir1_threshold_5": 0.03684225881529179, + "scr_metric_threshold_5": 0.5555554241520353, + "scr_dir2_threshold_5": 0.5555554241520353, + "scr_dir1_threshold_10": 0.09473687842942387, + "scr_metric_threshold_10": 0.603174561876354, + "scr_dir2_threshold_10": 0.603174561876354, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results", + "scr_dir1_threshold_2": 0.08743156050198193, + "scr_metric_threshold_2": 0.15267182345378855, + "scr_dir2_threshold_2": 0.15267182345378855, + "scr_dir1_threshold_5": 0.20218582972611174, + "scr_metric_threshold_5": 0.19847321124088896, + "scr_dir2_threshold_5": 0.19847321124088896, + "scr_dir1_threshold_10": 0.2950818016891243, + "scr_metric_threshold_10": 0.24045808212745792, + "scr_dir2_threshold_10": 0.24045808212745792, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results", + "scr_dir1_threshold_2": 0.032557992538054095, + "scr_metric_threshold_2": 0.29184559786405584, + "scr_dir2_threshold_2": 0.29184559786405584, + "scr_dir1_threshold_5": 0.08372094828480038, + "scr_metric_threshold_5": 0.4206008506837011, + "scr_dir2_threshold_5": 0.4206008506837011, + "scr_dir1_threshold_10": 0.13953472898742536, + "scr_metric_threshold_10": 0.5965665675217078, + "scr_dir2_threshold_10": 0.5965665675217078, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results", + "scr_dir1_threshold_2": 0.10232538539952167, + "scr_metric_threshold_2": 0.10232538539952167, + "scr_dir2_threshold_2": 0.1131220560094828, + "scr_dir1_threshold_5": 0.12093007653514554, + "scr_metric_threshold_5": 0.12093007653514554, + "scr_dir2_threshold_5": 0.16289597641705988, + "scr_dir1_threshold_10": 0.18604635689435922, + "scr_metric_threshold_10": 0.18604635689435922, + "scr_dir2_threshold_10": 0.22171955342369326, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_scr", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": null +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/tpp/saebench_tpp_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/tpp/saebench_tpp_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..d7cfff06de0d3bc2fd75a8e870174a1fea15315e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-12207-tokens-50000000/tpp/saebench_tpp_custom_sae_eval_results.json @@ -0,0 +1,297 @@ +{ + "eval_type_id": "tpp", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": false, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "7cb425b4-5b15-421d-ace0-7ec84d22f3c7", + "datetime_epoch_millis": 1776879398918, + "eval_result_metrics": { + "tpp_metrics": { + "tpp_threshold_2_total_metric": 0.013050007820129394, + "tpp_threshold_2_intended_diff_only": 0.01310001015663147, + "tpp_threshold_2_unintended_diff_only": 5.0002336502075174e-05, + "tpp_threshold_5_total_metric": 0.019749999046325684, + "tpp_threshold_5_intended_diff_only": 0.02290000319480896, + "tpp_threshold_5_unintended_diff_only": 0.0031500041484832765, + "tpp_threshold_10_total_metric": 0.036524996161460876, + "tpp_threshold_10_intended_diff_only": 0.041000002622604365, + "tpp_threshold_10_unintended_diff_only": 0.004475006461143493, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results", + "tpp_threshold_2_total_metric": 0.010800004005432129, + "tpp_threshold_2_intended_diff_only": 0.010000014305114746, + "tpp_threshold_2_unintended_diff_only": -0.0007999897003173828, + "tpp_threshold_5_total_metric": 0.015700006484985353, + "tpp_threshold_5_intended_diff_only": 0.017800021171569824, + "tpp_threshold_5_unintended_diff_only": 0.0021000146865844727, + "tpp_threshold_10_total_metric": 0.03234998881816864, + "tpp_threshold_10_intended_diff_only": 0.03500000238418579, + "tpp_threshold_10_unintended_diff_only": 0.002650013566017151, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results", + "tpp_threshold_2_total_metric": 0.01530001163482666, + "tpp_threshold_2_intended_diff_only": 0.016200006008148193, + "tpp_threshold_2_unintended_diff_only": 0.0008999943733215332, + "tpp_threshold_5_total_metric": 0.023799991607666014, + "tpp_threshold_5_intended_diff_only": 0.027999985218048095, + "tpp_threshold_5_unintended_diff_only": 0.00419999361038208, + "tpp_threshold_10_total_metric": 0.04070000350475311, + "tpp_threshold_10_intended_diff_only": 0.047000002861022946, + "tpp_threshold_10_unintended_diff_only": 0.006299999356269836, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_tpp", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1": { + "0": { + "tpp_threshold_2_total_metric": 0.013999968767166138, + "tpp_threshold_2_intended_diff_only": 0.014999985694885254, + "tpp_threshold_2_unintended_diff_only": 0.0010000169277191162, + "tpp_threshold_5_total_metric": 0.021999970078468323, + "tpp_threshold_5_intended_diff_only": 0.02399998903274536, + "tpp_threshold_5_unintended_diff_only": 0.0020000189542770386, + "tpp_threshold_10_total_metric": 0.020249947905540466, + "tpp_threshold_10_intended_diff_only": 0.02499997615814209, + "tpp_threshold_10_unintended_diff_only": 0.0047500282526016235 + }, + "1": { + "tpp_threshold_2_total_metric": -0.0027499794960021973, + "tpp_threshold_2_intended_diff_only": -0.001999974250793457, + "tpp_threshold_2_unintended_diff_only": 0.0007500052452087402, + "tpp_threshold_5_total_metric": 0.0012499988079071045, + "tpp_threshold_5_intended_diff_only": 0.0040000081062316895, + "tpp_threshold_5_unintended_diff_only": 0.002750009298324585, + "tpp_threshold_10_total_metric": 0.022749990224838257, + "tpp_threshold_10_intended_diff_only": 0.02799999713897705, + "tpp_threshold_10_unintended_diff_only": 0.005250006914138794 + }, + "2": { + "tpp_threshold_2_total_metric": 0.014500021934509277, + "tpp_threshold_2_intended_diff_only": 0.01100003719329834, + "tpp_threshold_2_unintended_diff_only": -0.0034999847412109375, + "tpp_threshold_5_total_metric": 0.02825002372264862, + "tpp_threshold_5_intended_diff_only": 0.029000043869018555, + "tpp_threshold_5_unintended_diff_only": 0.0007500201463699341, + "tpp_threshold_10_total_metric": 0.0532500296831131, + "tpp_threshold_10_intended_diff_only": 0.053000032901763916, + "tpp_threshold_10_unintended_diff_only": -0.00024999678134918213 + }, + "6": { + "tpp_threshold_2_total_metric": 0.015000015497207642, + "tpp_threshold_2_intended_diff_only": 0.012000024318695068, + "tpp_threshold_2_unintended_diff_only": -0.0029999911785125732, + "tpp_threshold_5_total_metric": 0.009499996900558472, + "tpp_threshold_5_intended_diff_only": 0.013000011444091797, + "tpp_threshold_5_unintended_diff_only": 0.003500014543533325, + "tpp_threshold_10_total_metric": 0.01749996840953827, + "tpp_threshold_10_intended_diff_only": 0.018999993801116943, + "tpp_threshold_10_unintended_diff_only": 0.0015000253915786743 + }, + "9": { + "tpp_threshold_2_total_metric": 0.013249993324279785, + "tpp_threshold_2_intended_diff_only": 0.013999998569488525, + "tpp_threshold_2_unintended_diff_only": 0.0007500052452087402, + "tpp_threshold_5_total_metric": 0.01750004291534424, + "tpp_threshold_5_intended_diff_only": 0.01900005340576172, + "tpp_threshold_5_unintended_diff_only": 0.0015000104904174805, + "tpp_threshold_10_total_metric": 0.04800000786781311, + "tpp_threshold_10_intended_diff_only": 0.050000011920928955, + "tpp_threshold_10_unintended_diff_only": 0.0020000040531158447 + } + }, + "canrager/amazon_reviews_mcauley_1and5": { + "1": { + "tpp_threshold_2_total_metric": 0.0104999840259552, + "tpp_threshold_2_intended_diff_only": 0.006999969482421875, + "tpp_threshold_2_unintended_diff_only": -0.003500014543533325, + "tpp_threshold_5_total_metric": 0.01774999499320984, + "tpp_threshold_5_intended_diff_only": 0.018999993801116943, + "tpp_threshold_5_unintended_diff_only": 0.0012499988079071045, + "tpp_threshold_10_total_metric": 0.025999978184700012, + "tpp_threshold_10_intended_diff_only": 0.0339999794960022, + "tpp_threshold_10_unintended_diff_only": 0.008000001311302185 + }, + "2": { + "tpp_threshold_2_total_metric": 0.004250049591064453, + "tpp_threshold_2_intended_diff_only": 0.01100003719329834, + "tpp_threshold_2_unintended_diff_only": 0.006749987602233887, + "tpp_threshold_5_total_metric": 0.005750060081481934, + "tpp_threshold_5_intended_diff_only": 0.012000024318695068, + "tpp_threshold_5_unintended_diff_only": 0.006249964237213135, + "tpp_threshold_10_total_metric": 0.0155000239610672, + "tpp_threshold_10_intended_diff_only": 0.017000019550323486, + "tpp_threshold_10_unintended_diff_only": 0.0014999955892562866 + }, + "3": { + "tpp_threshold_2_total_metric": 0.01475001871585846, + "tpp_threshold_2_intended_diff_only": 0.008000016212463379, + "tpp_threshold_2_unintended_diff_only": -0.006750002503395081, + "tpp_threshold_5_total_metric": 0.015249982476234436, + "tpp_threshold_5_intended_diff_only": 0.015999972820281982, + "tpp_threshold_5_unintended_diff_only": 0.0007499903440475464, + "tpp_threshold_10_total_metric": 0.011500030755996704, + "tpp_threshold_10_intended_diff_only": 0.021000027656555176, + "tpp_threshold_10_unintended_diff_only": 0.009499996900558472 + }, + "5": { + "tpp_threshold_2_total_metric": -0.006999999284744263, + "tpp_threshold_2_intended_diff_only": 0.0, + "tpp_threshold_2_unintended_diff_only": 0.006999999284744263, + "tpp_threshold_5_total_metric": 0.0009999722242355347, + "tpp_threshold_5_intended_diff_only": 0.001999974250793457, + "tpp_threshold_5_unintended_diff_only": 0.0010000020265579224, + "tpp_threshold_10_total_metric": 0.007749989628791809, + "tpp_threshold_10_intended_diff_only": 0.009999990463256836, + "tpp_threshold_10_unintended_diff_only": 0.002250000834465027 + }, + "6": { + "tpp_threshold_2_total_metric": 0.05400000512599945, + "tpp_threshold_2_intended_diff_only": 0.05500000715255737, + "tpp_threshold_2_unintended_diff_only": 0.0010000020265579224, + "tpp_threshold_5_total_metric": 0.07924994826316833, + "tpp_threshold_5_intended_diff_only": 0.09099996089935303, + "tpp_threshold_5_unintended_diff_only": 0.011750012636184692, + "tpp_threshold_10_total_metric": 0.14274999499320984, + "tpp_threshold_10_intended_diff_only": 0.15299999713897705, + "tpp_threshold_10_unintended_diff_only": 0.010250002145767212 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..31ba4fed269a04c1e36f502ac3c0263c3f724995 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.979470590342848, + "test_acc": 0.9794749403341289, + "test_auc": 0.9821841940258244, + "val_auc": 0.9862719862719863, + "k": 1, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431 + ] + }, + { + "test_f1": 0.979470590342848, + "test_acc": 0.9794749403341289, + "test_auc": 0.9941563317949502, + "val_auc": 0.9931359931359931, + "k": 2, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928 + ] + }, + { + "test_f1": 0.9809052680340768, + "test_acc": 0.9809069212410502, + "test_auc": 0.9979949984324533, + "val_auc": 0.9978549978549979, + "k": 5, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472 + ] + }, + { + "test_f1": 0.9923626710511473, + "test_acc": 0.99236276849642, + "test_auc": 0.9993843733823283, + "val_auc": 0.9982839982839983, + "k": 10, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472, + 461, + 1411, + 5951, + 542, + 786 + ] + }, + { + "test_f1": 0.9937947211270238, + "test_acc": 0.9937947494033413, + "test_auc": 0.9996081133299795, + "val_auc": 0.9982839982839983, + "k": 16, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472, + 461, + 1411, + 5951, + 542, + 786, + 126, + 1319, + 1634, + 1848, + 4674, + 1872 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..263f1eb3242d565f3f9d2c721c5254f3c264ebad --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.927371238389374, + "test_acc": 0.9274141283214518, + "test_auc": 0.9619211306223664, + "val_auc": 0.9607275733774286, + "k": 1, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441 + ] + }, + { + "test_f1": 0.946208232511368, + "test_acc": 0.9462086843810759, + "test_auc": 0.9643841185997595, + "val_auc": 0.9594873914840842, + "k": 2, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441, + 6838 + ] + }, + { + "test_f1": 0.9455445397139227, + "test_acc": 0.9455605962410888, + "test_auc": 0.97025261587468, + "val_auc": 0.9681686647374949, + "k": 5, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441, + 6838, + 1500, + 4110, + 943 + ] + }, + { + "test_f1": 0.9533115329848558, + "test_acc": 0.9533376539209333, + "test_auc": 0.9707734387075528, + "val_auc": 0.972302604381976, + "k": 10, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441, + 6838, + 1500, + 4110, + 943, + 1697, + 1180, + 908, + 30737, + 1417 + ] + }, + { + "test_f1": 0.9565576216532059, + "test_acc": 0.9565780946208684, + "test_auc": 0.9864720469345376, + "val_auc": 0.9909053327821414, + "k": 16, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441, + 6838, + 1500, + 4110, + 943, + 1697, + 1180, + 908, + 30737, + 1417, + 1411, + 654, + 698, + 601, + 1619, + 1703 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..52ff2c50bc49c306551eed6f782f8bb971accfae --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5622891387052189, + "test_acc": 0.5821383647798742, + "test_auc": 0.5891159741539321, + "val_auc": 0.6200480192076832, + "k": 1, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921 + ] + }, + { + "test_f1": 0.5666067502623415, + "test_acc": 0.5844025157232704, + "test_auc": 0.6068666655190328, + "val_auc": 0.6282513005202082, + "k": 2, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777 + ] + }, + { + "test_f1": 0.5774382716716295, + "test_acc": 0.5901886792452831, + "test_auc": 0.6214450771058156, + "val_auc": 0.6106442577030813, + "k": 5, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 2046 + ] + }, + { + "test_f1": 0.5931428275591205, + "test_acc": 0.599748427672956, + "test_auc": 0.6276741981835654, + "val_auc": 0.6680672268907563, + "k": 10, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 2046, + 1380, + 505, + 969, + 559, + 1335 + ] + }, + { + "test_f1": 0.5924677208295186, + "test_acc": 0.5967295597484277, + "test_auc": 0.6258398402493471, + "val_auc": 0.6340536214485795, + "k": 16, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 2046, + 1380, + 505, + 969, + 559, + 1335, + 1178, + 152, + 1533, + 521, + 26301, + 988 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5cd57a9341e6a66a2a458f21b42d85ee24d1e595 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6792809560411771, + "test_acc": 0.6847798742138365, + "test_auc": 0.695234947683079, + "val_auc": 0.7482993197278913, + "k": 1, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567 + ] + }, + { + "test_f1": 0.6934597353819044, + "test_acc": 0.6945911949685535, + "test_auc": 0.7351145119331995, + "val_auc": 0.7462985194077632, + "k": 2, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470 + ] + }, + { + "test_f1": 0.7171601443752754, + "test_acc": 0.7182389937106918, + "test_auc": 0.7624141932622407, + "val_auc": 0.8103241296518607, + "k": 5, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 4505, + 1164, + 433 + ] + }, + { + "test_f1": 0.7206808548396748, + "test_acc": 0.7215094339622642, + "test_auc": 0.7781461035969213, + "val_auc": 0.7981192476990796, + "k": 10, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 4505, + 1164, + 433, + 1345, + 1629, + 559, + 1731, + 6411 + ] + }, + { + "test_f1": 0.7263478071970517, + "test_acc": 0.7270440251572327, + "test_auc": 0.7813509137360652, + "val_auc": 0.7975190076030413, + "k": 16, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 4505, + 1164, + 433, + 1345, + 1629, + 559, + 1731, + 6411, + 628, + 794, + 993, + 1335, + 1928, + 1551 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a8125ec1ccc3fb2df71278c3f2688f8d4d5ed10e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7536095027057746, + "test_acc": 0.7547169811320755, + "test_auc": 0.815060215343394, + "val_auc": 0.8557422969187676, + "k": 1, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411 + ] + }, + { + "test_f1": 0.7315418828461587, + "test_acc": 0.7338364779874214, + "test_auc": 0.821108330911488, + "val_auc": 0.8667466986794718, + "k": 2, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786 + ] + }, + { + "test_f1": 0.7525654097165525, + "test_acc": 0.7532075471698113, + "test_auc": 0.8300338011967121, + "val_auc": 0.8813525410164066, + "k": 5, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417 + ] + }, + { + "test_f1": 0.7825567094810887, + "test_acc": 0.7826415094339623, + "test_auc": 0.8848891031139023, + "val_auc": 0.9265706282513005, + "k": 10, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417, + 1758, + 1335, + 181, + 5054, + 1183 + ] + }, + { + "test_f1": 0.7972025169156662, + "test_acc": 0.7972327044025157, + "test_auc": 0.8955879210846356, + "val_auc": 0.9161664665866347, + "k": 16, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417, + 1758, + 1335, + 181, + 5054, + 1183, + 1360, + 1108, + 12220, + 126, + 191, + 19471 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3394fa65bedb3cb51bb1607d705cc56daa869dec --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8461051660207854, + "test_acc": 0.8462893081761006, + "test_auc": 0.8935836964413557, + "val_auc": 0.9339735894357744, + "k": 1, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8759200924194657, + "test_acc": 0.8759748427672956, + "test_auc": 0.9368473042583635, + "val_auc": 0.9627851140456183, + "k": 2, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8754140593689842, + "test_acc": 0.8754716981132076, + "test_auc": 0.936106067709731, + "val_auc": 0.9615846338535414, + "k": 5, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 4675, + 22141, + 100 + ] + }, + { + "test_f1": 0.8493435156972853, + "test_acc": 0.849811320754717, + "test_auc": 0.9293344364121315, + "val_auc": 0.9587835134053622, + "k": 10, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 4675, + 22141, + 100, + 821, + 6319, + 505, + 6745, + 7284 + ] + }, + { + "test_f1": 0.8920736138002565, + "test_acc": 0.8920754716981132, + "test_auc": 0.9476413083432654, + "val_auc": 0.9595838335334134, + "k": 16, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 4675, + 22141, + 100, + 821, + 6319, + 505, + 6745, + 7284, + 1697, + 3654, + 7143, + 1438, + 3473, + 1202 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..757b09669d7728cabbf7237ae02786ab207f52b3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5106615262968283, + "test_acc": 0.5261538461538462, + "test_auc": 0.53538728245868, + "val_auc": 0.5287637698898409, + "k": 1, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264 + ] + }, + { + "test_f1": 0.5580661976547998, + "test_acc": 0.56, + "test_auc": 0.5755545830948935, + "val_auc": 0.554671562627499, + "k": 2, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945 + ] + }, + { + "test_f1": 0.5554860324172624, + "test_acc": 0.5620512820512821, + "test_auc": 0.5831012051031743, + "val_auc": 0.5675234598123214, + "k": 5, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 913 + ] + }, + { + "test_f1": 0.5533260509531696, + "test_acc": 0.556923076923077, + "test_auc": 0.6111480459151042, + "val_auc": 0.6185230518155855, + "k": 10, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 913, + 1851, + 1523, + 4531, + 470, + 26171 + ] + }, + { + "test_f1": 0.5924191080116761, + "test_acc": 0.5969230769230769, + "test_auc": 0.6440548523916922, + "val_auc": 0.6454508363933088, + "k": 16, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 913, + 1851, + 1523, + 4531, + 470, + 26171, + 7073, + 1895, + 1132, + 1447, + 551, + 386 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e8a8d55e285d5dfc0f8237151b8636c36b28297a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.4906281230709832, + "test_acc": 0.5169230769230769, + "test_auc": 0.5189265156360454, + "val_auc": 0.5185638514891882, + "k": 1, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341 + ] + }, + { + "test_f1": 0.49995277596897036, + "test_acc": 0.5353846153846153, + "test_auc": 0.5623758710068335, + "val_auc": 0.5487556099551204, + "k": 2, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 1409 + ] + }, + { + "test_f1": 0.5148869751374495, + "test_acc": 0.5158974358974359, + "test_auc": 0.5362582892920861, + "val_auc": 0.5257037943696451, + "k": 5, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 1409, + 1264, + 239, + 456 + ] + }, + { + "test_f1": 0.5102647227543808, + "test_acc": 0.5117948717948718, + "test_auc": 0.5185162587942236, + "val_auc": 0.5267237862097103, + "k": 10, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 1409, + 1264, + 239, + 456, + 238, + 400, + 538, + 1158, + 326 + ] + }, + { + "test_f1": 0.5368892793545521, + "test_acc": 0.5374358974358975, + "test_auc": 0.5427298279866698, + "val_auc": 0.5960832313341493, + "k": 16, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 1409, + 1264, + 239, + 456, + 238, + 400, + 538, + 1158, + 326, + 8141, + 1478, + 827, + 347, + 4201, + 1482 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7f49cd47aa0ea61cf7b257674752bfb99410cb87 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5394522980612879, + "test_acc": 0.5394871794871795, + "test_auc": 0.5455721715420607, + "val_auc": 0.6101591187270502, + "k": 1, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815 + ] + }, + { + "test_f1": 0.5398322464902607, + "test_acc": 0.5405128205128205, + "test_auc": 0.5579282660652372, + "val_auc": 0.6258669930640556, + "k": 2, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156 + ] + }, + { + "test_f1": 0.5354029257396614, + "test_acc": 0.5374358974358975, + "test_auc": 0.5507414077490154, + "val_auc": 0.627906976744186, + "k": 5, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 1770, + 698, + 1409 + ] + }, + { + "test_f1": 0.5304429949476936, + "test_acc": 0.5333333333333333, + "test_auc": 0.5494327936176658, + "val_auc": 0.6466748266013872, + "k": 10, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 1770, + 698, + 1409, + 16809, + 191, + 1611, + 538, + 1684 + ] + }, + { + "test_f1": 0.5326668254140436, + "test_acc": 0.5374358974358975, + "test_auc": 0.5503290470259534, + "val_auc": 0.6221950224398205, + "k": 16, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 1770, + 698, + 1409, + 16809, + 191, + 1611, + 538, + 1684, + 1265, + 12594, + 360, + 24861, + 400, + 848 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9478eccc3aa48a05b313067c09e9a3becc75c889 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.848175159876744, + "test_acc": 0.8484848484848485, + "test_auc": 0.887142857142857, + "val_auc": 0.8762019230769229, + "k": 1, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997 + ] + }, + { + "test_f1": 0.848175159876744, + "test_acc": 0.8484848484848485, + "test_auc": 0.8918367346938775, + "val_auc": 0.8954326923076924, + "k": 2, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335 + ] + }, + { + "test_f1": 0.848175159876744, + "test_acc": 0.8484848484848485, + "test_auc": 0.8879591836734693, + "val_auc": 0.8986378205128205, + "k": 5, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 1156, + 314 + ] + }, + { + "test_f1": 0.7675344854073037, + "test_acc": 0.7676767676767676, + "test_auc": 0.889795918367347, + "val_auc": 0.9475160256410255, + "k": 10, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 1156, + 314, + 1499, + 2780, + 851, + 2412, + 1279 + ] + }, + { + "test_f1": 0.7877488779805815, + "test_acc": 0.7878787878787878, + "test_auc": 0.9053061224489796, + "val_auc": 0.9499198717948718, + "k": 16, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 1156, + 314, + 1499, + 2780, + 851, + 2412, + 1279, + 1473, + 7481, + 3526, + 1158, + 5888, + 1684 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..14639e6ed328ceb509d368e5778f9ee33751d84e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8880672805906451, + "test_acc": 0.8888888888888888, + "test_auc": 0.9124489795918367, + "val_auc": 0.9459134615384615, + "k": 1, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8684184718931782, + "test_acc": 0.8686868686868687, + "test_auc": 0.9220408163265306, + "val_auc": 0.9627403846153846, + "k": 2, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335 + ] + }, + { + "test_f1": 0.89867992527567, + "test_acc": 0.898989898989899, + "test_auc": 0.956938775510204, + "val_auc": 0.9815705128205129, + "k": 5, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4889, + 4201 + ] + }, + { + "test_f1": 0.8782658413503368, + "test_acc": 0.8787878787878788, + "test_auc": 0.9661224489795919, + "val_auc": 0.9887820512820512, + "k": 10, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4889, + 4201, + 2832, + 1116, + 1924, + 581, + 2900 + ] + }, + { + "test_f1": 0.8886617839096123, + "test_acc": 0.8888888888888888, + "test_auc": 0.9622448979591837, + "val_auc": 0.9923878205128205, + "k": 16, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4889, + 4201, + 2832, + 1116, + 1924, + 581, + 2900, + 7997, + 3526, + 461, + 4531, + 1851, + 472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1a023d8669e83774283782af69456b67b6e9c191 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8681220810253067, + "test_acc": 0.8686868686868687, + "test_auc": 0.8914285714285715, + "val_auc": 0.8325320512820513, + "k": 1, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388 + ] + }, + { + "test_f1": 0.8780918288711562, + "test_acc": 0.8787878787878788, + "test_auc": 0.9175510204081633, + "val_auc": 0.8818108974358975, + "k": 2, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335 + ] + }, + { + "test_f1": 0.8681220810253067, + "test_acc": 0.8686868686868687, + "test_auc": 0.9381632653061224, + "val_auc": 0.9451121794871795, + "k": 5, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 5888, + 337, + 4201 + ] + }, + { + "test_f1": 0.8382848088730442, + "test_acc": 0.8383838383838383, + "test_auc": 0.9495918367346938, + "val_auc": 0.9537259615384616, + "k": 10, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 5888, + 337, + 4201, + 2900, + 1596, + 7384, + 2832, + 2000 + ] + }, + { + "test_f1": 0.8988661121014063, + "test_acc": 0.898989898989899, + "test_auc": 0.9526530612244899, + "val_auc": 0.9471153846153846, + "k": 16, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 5888, + 337, + 4201, + 2900, + 1596, + 7384, + 2832, + 2000, + 3825, + 1499, + 7481, + 1847, + 5463, + 7368 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..912b996402e153e0b9eeaa7b730f7fefb769a5e5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6358012901405214, + "test_acc": 0.657997399219766, + "test_auc": 0.6712268951014684, + "val_auc": 0.7406, + "k": 1, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.6480792473532627, + "test_acc": 0.6740355439965323, + "test_auc": 0.7666617564608039, + "val_auc": 0.8462000000000001, + "k": 2, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388 + ] + }, + { + "test_f1": 0.7361438115342405, + "test_acc": 0.7368877329865626, + "test_auc": 0.8333279471381267, + "val_auc": 0.9304000000000001, + "k": 5, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 789, + 5352, + 2335 + ] + }, + { + "test_f1": 0.784068702900731, + "test_acc": 0.7841352405721717, + "test_auc": 0.8685551669144317, + "val_auc": 0.9359999999999999, + "k": 10, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 789, + 5352, + 2335, + 337, + 7384, + 2000, + 7481, + 1602 + ] + }, + { + "test_f1": 0.7896884665416892, + "test_acc": 0.7897702644126572, + "test_auc": 0.884245905113779, + "val_auc": 0.9456, + "k": 16, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 789, + 5352, + 2335, + 337, + 7384, + 2000, + 7481, + 1602, + 2659, + 4201, + 4889, + 2330, + 1596, + 239 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..736ae362b897b83c2238f9749c2cf0ec6e6c6da1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6319106716867092, + "test_acc": 0.6527958387516255, + "test_auc": 0.6685806448703631, + "val_auc": 0.7098, + "k": 1, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.7008508062610861, + "test_acc": 0.7043779800606849, + "test_auc": 0.7455383514635169, + "val_auc": 0.7556, + "k": 2, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384 + ] + }, + { + "test_f1": 0.7622065360200792, + "test_acc": 0.7633289986996099, + "test_auc": 0.8591771747577341, + "val_auc": 0.8916, + "k": 5, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 4201, + 3526, + 5388 + ] + }, + { + "test_f1": 0.795491933508102, + "test_acc": 0.7962722149978326, + "test_auc": 0.9000354737321522, + "val_auc": 0.9316, + "k": 10, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 4201, + 3526, + 5388, + 7481, + 3842, + 2900, + 1924, + 1091 + ] + }, + { + "test_f1": 0.8049374508925177, + "test_acc": 0.8049414824447334, + "test_auc": 0.9071212014171455, + "val_auc": 0.942, + "k": 16, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 4201, + 3526, + 5388, + 7481, + 3842, + 2900, + 1924, + 1091, + 6244, + 2832, + 789, + 1614, + 119, + 5709 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9ece054be123af482d1df3931c05c393aa3fdd26 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8106260347436919, + "test_acc": 0.811876896402254, + "test_auc": 0.8420336669768114, + "val_auc": 0.8728, + "k": 1, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8103845942579552, + "test_acc": 0.8123103597745991, + "test_auc": 0.8923548846276987, + "val_auc": 0.9286, + "k": 2, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789 + ] + }, + { + "test_f1": 0.8180954921465378, + "test_acc": 0.8209796272214998, + "test_auc": 0.8880345297701272, + "val_auc": 0.9232, + "k": 5, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789, + 4889, + 5352, + 12946 + ] + }, + { + "test_f1": 0.8354736738719162, + "test_acc": 0.8370177719982661, + "test_auc": 0.9279297018853687, + "val_auc": 0.9312, + "k": 10, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789, + 4889, + 5352, + 12946, + 4531, + 4201, + 1924, + 2335, + 14154 + ] + }, + { + "test_f1": 0.8170081419691883, + "test_acc": 0.8201127004768097, + "test_auc": 0.9147660913208103, + "val_auc": 0.9548, + "k": 16, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789, + 4889, + 5352, + 12946, + 4531, + 4201, + 1924, + 2335, + 14154, + 1851, + 5388, + 1091, + 1158, + 559, + 737 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..75f4f3ee8d7c4702e9ed55cfb55af03455959cd3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9569205391904891, + "test_acc": 0.9569230769230769, + "test_auc": 0.9808715117649041, + "val_auc": 0.9946960424316605, + "k": 1, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958 + ] + }, + { + "test_f1": 0.9548691362947981, + "test_acc": 0.9548717948717949, + "test_auc": 0.9800594136062207, + "val_auc": 0.9951040391676866, + "k": 2, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818 + ] + }, + { + "test_f1": 0.9733333333333334, + "test_acc": 0.9733333333333334, + "test_auc": 0.9922366782239875, + "val_auc": 1.0, + "k": 5, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 2780 + ] + }, + { + "test_f1": 0.9794865321227618, + "test_acc": 0.9794871794871794, + "test_auc": 0.9935579156427777, + "val_auc": 1.0, + "k": 10, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 2780, + 386, + 730, + 1499, + 1279, + 4586 + ] + }, + { + "test_f1": 0.9784610853107364, + "test_acc": 0.9784615384615385, + "test_auc": 0.9946729727000371, + "val_auc": 1.0, + "k": 16, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 2780, + 386, + 730, + 1499, + 1279, + 4586, + 3475, + 5463, + 1158, + 1596, + 779, + 12605 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..88c8008bc46d135f694b7207f0f345f54cfc4391 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.926839190384715, + "test_acc": 0.9271794871794872, + "test_auc": 0.9286721143164911, + "val_auc": 0.9534883720930232, + "k": 1, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065 + ] + }, + { + "test_f1": 0.9278594089471172, + "test_acc": 0.9282051282051282, + "test_auc": 0.9740949102905039, + "val_auc": 0.9869441044471644, + "k": 2, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508 + ] + }, + { + "test_f1": 0.925776755025442, + "test_acc": 0.9261538461538461, + "test_auc": 0.9808399535463022, + "val_auc": 0.9885760913912689, + "k": 5, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 1279 + ] + }, + { + "test_f1": 0.9425641025641026, + "test_acc": 0.9425641025641026, + "test_auc": 0.9921209647557814, + "val_auc": 0.9991840065279478, + "k": 10, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 1279, + 992, + 3475, + 1499, + 2708, + 581 + ] + }, + { + "test_f1": 0.9743574637800257, + "test_acc": 0.9743589743589743, + "test_auc": 0.9967011142155048, + "val_auc": 1.0, + "k": 16, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 1279, + 992, + 3475, + 1499, + 2708, + 581, + 5406, + 778, + 3958, + 1482, + 6828, + 305 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d11c272a0727b6d3f85cac658638ed009159c8db --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.956909206525174, + "test_acc": 0.9569230769230769, + "test_auc": 0.9762577002053388, + "val_auc": 0.9979600163198694, + "k": 1, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037 + ] + }, + { + "test_f1": 0.9599856045610812, + "test_acc": 0.96, + "test_auc": 0.9803728919109974, + "val_auc": 0.9983680130558955, + "k": 2, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508 + ] + }, + { + "test_f1": 0.9599856045610812, + "test_acc": 0.96, + "test_auc": 0.9800425825562997, + "val_auc": 0.9979600163198694, + "k": 5, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 5472, + 1028, + 26034 + ] + }, + { + "test_f1": 0.9753799018576346, + "test_acc": 0.9753846153846154, + "test_auc": 0.9945341165381897, + "val_auc": 1.0, + "k": 10, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 5472, + 1028, + 26034, + 1253, + 1279, + 493, + 778, + 1079 + ] + }, + { + "test_f1": 0.9774327638756616, + "test_acc": 0.9774358974358974, + "test_auc": 0.9942690275019355, + "val_auc": 1.0, + "k": 16, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 5472, + 1028, + 26034, + 1253, + 1279, + 493, + 778, + 1079, + 581, + 582, + 305, + 3475, + 12813, + 238 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6402d3d468e515984816cd1a33e81c2e0c1533ae --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6022641618977247, + "test_acc": 0.6293888166449935, + "test_auc": 0.6337378491193947, + "val_auc": 0.6542, + "k": 1, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351 + ] + }, + { + "test_f1": 0.6355333648959068, + "test_acc": 0.6423927178153446, + "test_auc": 0.667131633099397, + "val_auc": 0.7052000000000002, + "k": 2, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1441 + ] + }, + { + "test_f1": 0.6661630129272125, + "test_acc": 0.6697009102730819, + "test_auc": 0.7331781608072379, + "val_auc": 0.7667999999999999, + "k": 5, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1441, + 1894, + 559, + 3078 + ] + }, + { + "test_f1": 0.6864628871378918, + "test_acc": 0.6883398352839185, + "test_auc": 0.7530058727064204, + "val_auc": 0.7874, + "k": 10, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1441, + 1894, + 559, + 3078, + 654, + 619, + 343, + 1960, + 1602 + ] + }, + { + "test_f1": 0.7087066065272408, + "test_acc": 0.7100130039011704, + "test_auc": 0.7826290695209994, + "val_auc": 0.8194, + "k": 16, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1441, + 1894, + 559, + 3078, + 654, + 619, + 343, + 1960, + 1602, + 7595, + 2046, + 19018, + 1305, + 4510, + 3301 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2a922a51b61ff88aa19fab67b16b0a1de140bb41 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5966157688298253, + "test_acc": 0.6272214997832684, + "test_auc": 0.6329077487557888, + "val_auc": 0.634, + "k": 1, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274 + ] + }, + { + "test_f1": 0.5991899287350257, + "test_acc": 0.6280884265279584, + "test_auc": 0.6461807116090794, + "val_auc": 0.6632, + "k": 2, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121 + ] + }, + { + "test_f1": 0.6681508621492285, + "test_acc": 0.671434763762462, + "test_auc": 0.7188075414749557, + "val_auc": 0.7196, + "k": 5, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 698, + 3078 + ] + }, + { + "test_f1": 0.7422318773610836, + "test_acc": 0.7425227568270482, + "test_auc": 0.8125408661903767, + "val_auc": 0.8174, + "k": 10, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 698, + 3078, + 52, + 1524, + 1602, + 4510, + 649 + ] + }, + { + "test_f1": 0.7581101677489881, + "test_acc": 0.7581274382314694, + "test_auc": 0.8404061592019012, + "val_auc": 0.8400000000000001, + "k": 16, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 698, + 3078, + 52, + 1524, + 1602, + 4510, + 649, + 2046, + 1895, + 19018, + 1158, + 831, + 1351 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b425de5a58f4789c5d15b8d12527b6631dfaf1e2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.566390460639583, + "test_acc": 0.5864759427828349, + "test_auc": 0.5950579529552175, + "val_auc": 0.584, + "k": 1, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.5698911088901407, + "test_acc": 0.5960121369744257, + "test_auc": 0.6613017657200492, + "val_auc": 0.7036, + "k": 2, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508 + ] + }, + { + "test_f1": 0.6414652989147998, + "test_acc": 0.6454269614217598, + "test_auc": 0.7111468687667317, + "val_auc": 0.7535999999999999, + "k": 5, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121 + ] + }, + { + "test_f1": 0.6538495988633334, + "test_acc": 0.6558300823580407, + "test_auc": 0.7295428548237513, + "val_auc": 0.8144, + "k": 10, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 567, + 52, + 7893, + 925, + 1441 + ] + }, + { + "test_f1": 0.6675656110642338, + "test_acc": 0.6688339835283919, + "test_auc": 0.745393299973996, + "val_auc": 0.8200000000000001, + "k": 16, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 567, + 52, + 7893, + 925, + 1441, + 1524, + 3274, + 4654, + 51, + 3301, + 4510 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..aa7daea686856bfc8ee097ce482a02388ce9bb99 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6177425937522635, + "test_acc": 0.6183368869936035, + "test_auc": 0.645426441171122, + "val_auc": 0.6749999999999999, + "k": 1, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048 + ] + }, + { + "test_f1": 0.6585225783248485, + "test_acc": 0.6588486140724946, + "test_auc": 0.7011183851609384, + "val_auc": 0.7248, + "k": 2, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 696 + ] + }, + { + "test_f1": 0.7228069376795992, + "test_acc": 0.7228144989339019, + "test_auc": 0.7603746135661027, + "val_auc": 0.7236, + "k": 5, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 696, + 1504, + 1499, + 1902 + ] + }, + { + "test_f1": 0.7343320522803057, + "test_acc": 0.7356076759061834, + "test_auc": 0.8019821785779233, + "val_auc": 0.7816, + "k": 10, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 696, + 1504, + 1499, + 1902, + 927, + 1818, + 1927, + 7129, + 4672 + ] + }, + { + "test_f1": 0.7177801922499488, + "test_acc": 0.7185501066098081, + "test_auc": 0.7889434442625931, + "val_auc": 0.7852, + "k": 16, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 696, + 1504, + 1499, + 1902, + 927, + 1818, + 1927, + 7129, + 4672, + 437, + 741, + 246, + 29085, + 6905, + 499 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..916cd83bdb44693252e1a38c2665e739734d8cac --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9494949494949495, + "test_acc": 0.9494949494949495, + "test_auc": 0.9722448979591837, + "val_auc": 0.9520933977455718, + "k": 1, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299 + ] + }, + { + "test_f1": 0.9494949494949495, + "test_acc": 0.9494949494949495, + "test_auc": 0.976326530612245, + "val_auc": 0.9561191626409019, + "k": 2, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 23633 + ] + }, + { + "test_f1": 0.979793856936714, + "test_acc": 0.9797979797979798, + "test_auc": 0.9995918367346939, + "val_auc": 0.9913446054750403, + "k": 5, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 23633, + 114, + 432, + 1915 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 1.0, + "val_auc": 1.0, + "k": 10, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 23633, + 114, + 432, + 1915, + 6055, + 6877, + 1610, + 2535, + 1232 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 1.0, + "val_auc": 1.0, + "k": 16, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 23633, + 114, + 432, + 1915, + 6055, + 6877, + 1610, + 2535, + 1232, + 1500, + 27804, + 19940, + 373, + 1420, + 3959 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..30b86e47458b508f10d30a997bcfa508ca67ab67 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7187700897378317, + "test_acc": 0.7373737373737373, + "test_auc": 0.74, + "val_auc": 0.75, + "k": 1, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 278 + ] + }, + { + "test_f1": 0.8270130700037241, + "test_acc": 0.8282828282828283, + "test_auc": 0.8451020408163264, + "val_auc": 0.8502415458937198, + "k": 2, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 278, + 23791 + ] + }, + { + "test_f1": 0.8787631416202846, + "test_acc": 0.8787878787878788, + "test_auc": 0.9740816326530611, + "val_auc": 0.9537037037037037, + "k": 5, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 278, + 23791, + 429, + 1657, + 1396 + ] + }, + { + "test_f1": 0.9595132999388319, + "test_acc": 0.9595959595959596, + "test_auc": 0.9914285714285714, + "val_auc": 0.980475040257649, + "k": 10, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 278, + 23791, + 429, + 1657, + 1396, + 19000, + 253, + 521, + 114, + 4580 + ] + }, + { + "test_f1": 0.9696969696969697, + "test_acc": 0.9696969696969697, + "test_auc": 0.9926530612244898, + "val_auc": 0.9939613526570049, + "k": 16, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 278, + 23791, + 429, + 1657, + 1396, + 19000, + 253, + 521, + 114, + 4580, + 1299, + 4314, + 1411, + 1610, + 1420, + 5387 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..04a484cfc0c4d3c26472f4eb7587c74c0a8f1b3b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9494330117733285, + "test_acc": 0.9494949494949495, + "test_auc": 0.986326530612245, + "val_auc": 0.8796296296296297, + "k": 1, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1610 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 0.9991836734693877, + "val_auc": 0.9716183574879228, + "k": 2, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1610, + 1396 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 0.9991836734693877, + "val_auc": 0.9842995169082126, + "k": 5, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1610, + 1396, + 1244, + 419, + 1550 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 0.9995918367346939, + "val_auc": 0.9867149758454107, + "k": 10, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1610, + 1396, + 1244, + 419, + 1550, + 1501, + 177, + 5948, + 7577, + 907 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 1.0, + "val_auc": 0.998389694041868, + "k": 16, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1610, + 1396, + 1244, + 419, + 1550, + 1501, + 177, + 5948, + 7577, + 907, + 1360, + 6662, + 7891, + 1960, + 15752, + 4580 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fcd087ceb02abf4eb3c459af7e650e279ffb3d37 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9333306993803626, + "test_acc": 0.9333333333333333, + "test_auc": 0.9483249051833123, + "val_auc": 0.9342532467532467, + "k": 1, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752 + ] + }, + { + "test_f1": 0.9422153739444982, + "test_acc": 0.9422222222222222, + "test_auc": 0.9612041719342603, + "val_auc": 0.9815340909090909, + "k": 2, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1369 + ] + }, + { + "test_f1": 0.9333333333333333, + "test_acc": 0.9333333333333333, + "test_auc": 0.9598214285714285, + "val_auc": 0.9823457792207793, + "k": 5, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1369, + 959, + 1309, + 22580 + ] + }, + { + "test_f1": 0.9422153739444982, + "test_acc": 0.9422222222222222, + "test_auc": 0.9638906447534765, + "val_auc": 0.9866071428571429, + "k": 10, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1369, + 959, + 1309, + 22580, + 5563, + 3301, + 943, + 1915, + 969 + ] + }, + { + "test_f1": 0.9599905155209548, + "test_acc": 0.96, + "test_auc": 0.9795353982300885, + "val_auc": 0.9918831168831169, + "k": 16, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1369, + 959, + 1309, + 22580, + 5563, + 3301, + 943, + 1915, + 969, + 1073, + 1541, + 1182, + 1295, + 4850, + 10260 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d500f06c9584f68aaf2a9140e047e8cbfee5ba7d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7955555555555556, + "test_acc": 0.7955555555555556, + "test_auc": 0.8346634007585335, + "val_auc": 0.8305600649350648, + "k": 1, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406 + ] + }, + { + "test_f1": 0.8080930505154662, + "test_acc": 0.8088888888888889, + "test_auc": 0.8693505056890013, + "val_auc": 0.8392857142857144, + "k": 2, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854 + ] + }, + { + "test_f1": 0.8075895111429195, + "test_acc": 0.8088888888888889, + "test_auc": 0.882822376738306, + "val_auc": 0.8662743506493508, + "k": 5, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1804, + 1182 + ] + }, + { + "test_f1": 0.8528800935005248, + "test_acc": 0.8533333333333334, + "test_auc": 0.9171934260429835, + "val_auc": 0.8695211038961039, + "k": 10, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1804, + 1182, + 1232, + 1570, + 311, + 3481, + 1420 + ] + }, + { + "test_f1": 0.8350473775611946, + "test_acc": 0.8355555555555556, + "test_auc": 0.9223293299620733, + "val_auc": 0.8772321428571429, + "k": 16, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1804, + 1182, + 1232, + 1570, + 311, + 3481, + 1420, + 800, + 322, + 1858, + 920, + 6662, + 748 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..99e1ee0af1d98539e0e922b2ec28338e19105e68 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9155455465342666, + "test_acc": 0.9155555555555556, + "test_auc": 0.955515170670038, + "val_auc": 0.9375, + "k": 1, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9154854480881879, + "test_acc": 0.9155555555555556, + "test_auc": 0.9647993046776233, + "val_auc": 0.9561688311688312, + "k": 2, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16 + ] + }, + { + "test_f1": 0.9510917908760189, + "test_acc": 0.9511111111111111, + "test_auc": 0.9842762326169405, + "val_auc": 0.9650974025974026, + "k": 5, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 21467 + ] + }, + { + "test_f1": 0.9599952588846526, + "test_acc": 0.96, + "test_auc": 0.99423198482933, + "val_auc": 0.9910714285714286, + "k": 10, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 21467, + 1835, + 79, + 1960, + 4580, + 1915 + ] + }, + { + "test_f1": 0.9599841925349245, + "test_acc": 0.96, + "test_auc": 0.9950221238938053, + "val_auc": 0.9943181818181819, + "k": 16, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 21467, + 1835, + 79, + 1960, + 4580, + 1915, + 969, + 15752, + 6718, + 1417, + 7095, + 1420 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e53e5919b7dd756f416744ff9839143faa38a718 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.65631184797218, + "test_acc": 0.6614651061985263, + "test_auc": 0.6807709824870993, + "val_auc": 0.7288000000000001, + "k": 1, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649 + ] + }, + { + "test_f1": 0.7002916269860426, + "test_acc": 0.706111833550065, + "test_auc": 0.754299311118159, + "val_auc": 0.812, + "k": 2, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394 + ] + }, + { + "test_f1": 0.7227037700417082, + "test_acc": 0.7238838318162115, + "test_auc": 0.7730154626390954, + "val_auc": 0.8231999999999999, + "k": 5, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 943, + 7614, + 28053 + ] + }, + { + "test_f1": 0.7376485615061654, + "test_acc": 0.7386215864759428, + "test_auc": 0.8080810965591982, + "val_auc": 0.866, + "k": 10, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 943, + 7614, + 28053, + 1928, + 1560, + 28397, + 811, + 253 + ] + }, + { + "test_f1": 0.7422530585829447, + "test_acc": 0.7429562201993931, + "test_auc": 0.8153216460412968, + "val_auc": 0.8684, + "k": 16, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 943, + 7614, + 28053, + 1928, + 1560, + 28397, + 811, + 253, + 1848, + 7246, + 1737, + 17225, + 2041, + 1551 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9e8e774159dca31c59dac644cc0c26cd9b29ef49 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5990258883734136, + "test_acc": 0.5990463805808409, + "test_auc": 0.6326537207585967, + "val_auc": 0.6412, + "k": 1, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943 + ] + }, + { + "test_f1": 0.6265226396150242, + "test_acc": 0.6267880364109233, + "test_auc": 0.6651888450143624, + "val_auc": 0.6946, + "k": 2, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253 + ] + }, + { + "test_f1": 0.6416791787040564, + "test_acc": 0.6419592544429996, + "test_auc": 0.6974049311493942, + "val_auc": 0.742, + "k": 5, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 7649, + 1011, + 7614 + ] + }, + { + "test_f1": 0.6410120945336101, + "test_acc": 0.6419592544429996, + "test_auc": 0.7057499011695809, + "val_auc": 0.7328, + "k": 10, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 7649, + 1011, + 7614, + 32305, + 28053, + 5394, + 1551, + 1928 + ] + }, + { + "test_f1": 0.6848394411060057, + "test_acc": 0.6848721283051582, + "test_auc": 0.7541805643029036, + "val_auc": 0.7772, + "k": 16, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 7649, + 1011, + 7614, + 32305, + 28053, + 5394, + 1551, + 1928, + 1411, + 844, + 1749, + 635, + 1417, + 56 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..406cd0c0b015122bd42a67abf6f3344f0def2dd3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.610375909132735, + "test_acc": 0.623753792804508, + "test_auc": 0.6438091573335177, + "val_auc": 0.7044, + "k": 1, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253 + ] + }, + { + "test_f1": 0.6207677846566636, + "test_acc": 0.6341569137407889, + "test_auc": 0.7036507881631973, + "val_auc": 0.7456, + "k": 2, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394 + ] + }, + { + "test_f1": 0.7067892645171567, + "test_acc": 0.7108799306458604, + "test_auc": 0.7956209481407104, + "val_auc": 0.7728, + "k": 5, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 844, + 7649, + 1541 + ] + }, + { + "test_f1": 0.7165024827544869, + "test_acc": 0.7191157347204161, + "test_auc": 0.8067929190823125, + "val_auc": 0.7934, + "k": 10, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 844, + 7649, + 1541, + 794, + 1099, + 28053, + 4675, + 28397 + ] + }, + { + "test_f1": 0.7352308961439565, + "test_acc": 0.7368877329865626, + "test_auc": 0.8189592067111491, + "val_auc": 0.8082, + "k": 16, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 844, + 7649, + 1541, + 794, + 1099, + 28053, + 4675, + 28397, + 100, + 1960, + 1737, + 505, + 7614, + 1857 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..15b3b188133ed1be78a3a4765d94152da639c647 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7925857986109961, + "test_acc": 0.7945762711864407, + "test_auc": 0.8211437270410697, + "val_auc": 0.8099838969404187, + "k": 1, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7925857986109961, + "test_acc": 0.7945762711864407, + "test_auc": 0.8211437270410697, + "val_auc": 0.8099838969404187, + "k": 2, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 32278 + ] + }, + { + "test_f1": 0.7925857986109961, + "test_acc": 0.7945762711864407, + "test_auc": 0.871644181163657, + "val_auc": 0.8454106280193237, + "k": 5, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 32278, + 1847, + 508, + 5020 + ] + }, + { + "test_f1": 0.8377907462233519, + "test_acc": 0.8379661016949153, + "test_auc": 0.920171867933062, + "val_auc": 0.931561996779388, + "k": 10, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 32278, + 1847, + 508, + 5020, + 266, + 4413, + 1433, + 776, + 6838 + ] + }, + { + "test_f1": 0.8718552183967047, + "test_acc": 0.871864406779661, + "test_auc": 0.9456634050736707, + "val_auc": 0.9428341384863124, + "k": 16, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 32278, + 1847, + 508, + 5020, + 266, + 4413, + 1433, + 776, + 6838, + 1697, + 1528, + 1628, + 27673, + 1180, + 1585 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..305349b5a1208eeeb10f5f291794c2c636aca101 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7913670634072635, + "test_acc": 0.7972881355932203, + "test_auc": 0.8034568473228829, + "val_auc": 0.7157809983896941, + "k": 1, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528 + ] + }, + { + "test_f1": 0.8020291309890475, + "test_acc": 0.8067796610169492, + "test_auc": 0.8127700374697098, + "val_auc": 0.7616747181964574, + "k": 2, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791 + ] + }, + { + "test_f1": 0.8124120785328041, + "test_acc": 0.8142372881355933, + "test_auc": 0.8805124414880513, + "val_auc": 0.8345410628019324, + "k": 5, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1815, + 1628, + 4185 + ] + }, + { + "test_f1": 0.8128827683615819, + "test_acc": 0.8142372881355933, + "test_auc": 0.8880918761697794, + "val_auc": 0.8317230273752013, + "k": 10, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1815, + 1628, + 4185, + 416, + 266, + 1103, + 31149, + 16223 + ] + }, + { + "test_f1": 0.8240429481094771, + "test_acc": 0.8250847457627118, + "test_auc": 0.9106867730821133, + "val_auc": 0.8454106280193237, + "k": 16, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1815, + 1628, + 4185, + 416, + 266, + 1103, + 31149, + 16223, + 1866, + 4558, + 20997, + 2023, + 1789, + 203 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..45650b69e16d637b6c90bcad50683443627fccf6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.645479066643475, + "test_acc": 0.6508474576271186, + "test_auc": 0.6836640890153813, + "val_auc": 0.7043075684380032, + "k": 1, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020 + ] + }, + { + "test_f1": 0.6949141023854907, + "test_acc": 0.6983050847457627, + "test_auc": 0.7612620563112008, + "val_auc": 0.8200483091787439, + "k": 2, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266 + ] + }, + { + "test_f1": 0.7009534285314507, + "test_acc": 0.704406779661017, + "test_auc": 0.7954830430258171, + "val_auc": 0.8655394524959743, + "k": 5, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 32278, + 431 + ] + }, + { + "test_f1": 0.7828596986817326, + "test_acc": 0.7844067796610169, + "test_auc": 0.858560118844065, + "val_auc": 0.927133655394525, + "k": 10, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 32278, + 431, + 373, + 1628, + 27673, + 18827, + 895 + ] + }, + { + "test_f1": 0.8119733402526075, + "test_acc": 0.8128813559322033, + "test_auc": 0.8811449037149802, + "val_auc": 0.9162640901771337, + "k": 16, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 32278, + 431, + 373, + 1628, + 27673, + 18827, + 895, + 5804, + 16223, + 416, + 776, + 4413, + 508 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9c7cba11dcdf261a20327450bbff2c65ae10f5d7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5793154031479381, + "test_acc": 0.5830903790087464, + "test_auc": 0.5932952536379709, + "val_auc": 0.5372148859543817, + "k": 1, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381 + ] + }, + { + "test_f1": 0.5686259938766673, + "test_acc": 0.5743440233236151, + "test_auc": 0.60874813001496, + "val_auc": 0.5904361744697879, + "k": 2, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 993 + ] + }, + { + "test_f1": 0.5928028354198822, + "test_acc": 0.597667638483965, + "test_auc": 0.6262749898000817, + "val_auc": 0.5916366546618648, + "k": 5, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 993, + 1122, + 219, + 799 + ] + }, + { + "test_f1": 0.6471703426868548, + "test_acc": 0.6472303206997084, + "test_auc": 0.7185332517339862, + "val_auc": 0.7090836334533813, + "k": 10, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 993, + 1122, + 219, + 799, + 1799, + 5897, + 275, + 1411, + 1508 + ] + }, + { + "test_f1": 0.6879936088222601, + "test_acc": 0.6880466472303207, + "test_auc": 0.7612369101047192, + "val_auc": 0.7527010804321729, + "k": 16, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 993, + 1122, + 219, + 799, + 1799, + 5897, + 275, + 1411, + 1508, + 1697, + 801, + 8006, + 1039, + 304, + 506 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d72060dba361ec5e055c8bbc646a797037325499 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7230964460628838, + "test_acc": 0.7376093294460642, + "test_auc": 0.7440500475996192, + "val_auc": 0.6996798719487796, + "k": 1, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7790790177841551, + "test_acc": 0.7842565597667639, + "test_auc": 0.7999286005711954, + "val_auc": 0.7979191676670668, + "k": 2, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275 + ] + }, + { + "test_f1": 0.8204445472880801, + "test_acc": 0.8221574344023324, + "test_auc": 0.8785869713042296, + "val_auc": 0.838735494197679, + "k": 5, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 191, + 433, + 8006 + ] + }, + { + "test_f1": 0.766083755259536, + "test_acc": 0.7667638483965015, + "test_auc": 0.8629130966952264, + "val_auc": 0.8727490996398559, + "k": 10, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 191, + 433, + 8006, + 534, + 1928, + 506, + 799, + 1122 + ] + }, + { + "test_f1": 0.8918469353587224, + "test_acc": 0.892128279883382, + "test_auc": 0.9492724058207535, + "val_auc": 0.9211684673869547, + "k": 16, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 191, + 433, + 8006, + 534, + 1928, + 506, + 799, + 1122, + 1799, + 801, + 1411, + 1960, + 635, + 1239 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..15185a2f959161abf405f50a40de45d3a650a40f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6789778340798748, + "test_acc": 0.685131195335277, + "test_auc": 0.6880524955800353, + "val_auc": 0.6948779511804721, + "k": 1, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.6846434006259734, + "test_acc": 0.685131195335277, + "test_auc": 0.733000135998912, + "val_auc": 0.7583033213285315, + "k": 2, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 433 + ] + }, + { + "test_f1": 0.7462348867161539, + "test_acc": 0.7463556851311953, + "test_auc": 0.8063715490276078, + "val_auc": 0.7827130852340938, + "k": 5, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 433, + 191, + 534, + 1239 + ] + }, + { + "test_f1": 0.8221544110258081, + "test_acc": 0.8221574344023324, + "test_auc": 0.871515027879777, + "val_auc": 0.8479391756702682, + "k": 10, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 433, + 191, + 534, + 1239, + 5897, + 275, + 8006, + 635, + 1464 + ] + }, + { + "test_f1": 0.7957273577186644, + "test_acc": 0.7959183673469388, + "test_auc": 0.8712090303277574, + "val_auc": 0.8527410964385754, + "k": 16, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 433, + 191, + 534, + 1239, + 5897, + 275, + 8006, + 635, + 1464, + 4003, + 3233, + 1195, + 1411, + 1596, + 142 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c0b93654b89cbce50b322c0a6e66bae827286a46 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320 + ] + }, + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 2, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 5, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 1073, + 859, + 987 + ] + }, + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 10, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 1073, + 859, + 987, + 5883, + 1703, + 1030, + 1848, + 1011 + ] + }, + { + "test_f1": 0.46624047786838485, + "test_acc": 0.47474747474747475, + "test_auc": 0.44816326530612244, + "val_auc": 0.5680808080808081, + "k": 16, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 1073, + 859, + 987, + 5883, + 1703, + 1030, + 1848, + 1011, + 13612, + 4558, + 4712, + 4264, + 4393, + 113 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..87330c9a92d99045d77b9d7a4c5c4ca2a35ff10a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5251556083443396, + "test_acc": 0.5252525252525253, + "test_auc": 0.5571428571428572, + "val_auc": 0.5925252525252526, + "k": 1, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520 + ] + }, + { + "test_f1": 0.5252525252525253, + "test_acc": 0.5252525252525253, + "test_auc": 0.5344897959183674, + "val_auc": 0.5915151515151514, + "k": 2, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1788 + ] + }, + { + "test_f1": 0.5145573380867499, + "test_acc": 0.5151515151515151, + "test_auc": 0.5708163265306122, + "val_auc": 0.637979797979798, + "k": 5, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1788, + 6534, + 4712, + 5653 + ] + }, + { + "test_f1": 0.5047473819546904, + "test_acc": 0.5050505050505051, + "test_auc": 0.5224489795918367, + "val_auc": 0.6117171717171718, + "k": 10, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1788, + 6534, + 4712, + 5653, + 5129, + 31245, + 1030, + 1818, + 1034 + ] + }, + { + "test_f1": 0.48453298938141237, + "test_acc": 0.48484848484848486, + "test_auc": 0.5108163265306123, + "val_auc": 0.6258585858585859, + "k": 16, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1788, + 6534, + 4712, + 5653, + 5129, + 31245, + 1030, + 1818, + 1034, + 2046, + 1366, + 5551, + 7666, + 347, + 1194 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e9cf627f835be8fa5338e7cacf4144cdd4612efa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.60581934482108, + "test_acc": 0.6060606060606061, + "test_auc": 0.5889795918367347, + "val_auc": 0.5692929292929292, + "k": 1, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666 + ] + }, + { + "test_f1": 0.5655678969958852, + "test_acc": 0.5656565656565656, + "test_auc": 0.603265306122449, + "val_auc": 0.5806060606060607, + "k": 2, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 1030 + ] + }, + { + "test_f1": 0.5352587095444239, + "test_acc": 0.5353535353535354, + "test_auc": 0.5861224489795919, + "val_auc": 0.6157575757575757, + "k": 5, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 1030, + 31245, + 2779, + 27174 + ] + }, + { + "test_f1": 0.5856049522478021, + "test_acc": 0.5858585858585859, + "test_auc": 0.6185714285714287, + "val_auc": 0.6844444444444444, + "k": 10, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 1030, + 31245, + 2779, + 27174, + 5883, + 7026, + 2595, + 111, + 4985 + ] + }, + { + "test_f1": 0.6352477309924118, + "test_acc": 0.6363636363636364, + "test_auc": 0.6477551020408163, + "val_auc": 0.6864646464646463, + "k": 16, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 1030, + 31245, + 2779, + 27174, + 5883, + 7026, + 2595, + 111, + 4985, + 1180, + 1116, + 1717, + 373, + 4520, + 10308 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7a6c5b47798f599a525dd3ff7e41dbd9efca8aa0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5020682833445714, + "test_acc": 0.551864406779661, + "test_auc": 0.553082701790383, + "val_auc": 0.6332528180354268, + "k": 1, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217 + ] + }, + { + "test_f1": 0.5863740889049524, + "test_acc": 0.5877966101694915, + "test_auc": 0.6018613510422757, + "val_auc": 0.6404991948470209, + "k": 2, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1319 + ] + }, + { + "test_f1": 0.5781363900265303, + "test_acc": 0.5783050847457627, + "test_auc": 0.613387239706861, + "val_auc": 0.606682769726248, + "k": 5, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1319, + 1960, + 1541, + 186 + ] + }, + { + "test_f1": 0.5863170855250341, + "test_acc": 0.5864406779661017, + "test_auc": 0.6298551587958213, + "val_auc": 0.5865539452495974, + "k": 10, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1319, + 1960, + 1541, + 186, + 539, + 311, + 20, + 1812, + 100 + ] + }, + { + "test_f1": 0.575114762246228, + "test_acc": 0.5755932203389831, + "test_auc": 0.6140739392468552, + "val_auc": 0.6175523349436393, + "k": 16, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1319, + 1960, + 1541, + 186, + 539, + 311, + 20, + 1812, + 100, + 253, + 6088, + 1616, + 836, + 373, + 3780 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ac61b53b764041b14fd829b81c649cc862c44ce2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.619613460785858, + "test_acc": 0.6196610169491525, + "test_auc": 0.6451500810801867, + "val_auc": 0.6207729468599034, + "k": 1, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319 + ] + }, + { + "test_f1": 0.655550156350676, + "test_acc": 0.655593220338983, + "test_auc": 0.7065614278937904, + "val_auc": 0.6674718196457328, + "k": 2, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541 + ] + }, + { + "test_f1": 0.671155190447643, + "test_acc": 0.671864406779661, + "test_auc": 0.7241784425985373, + "val_auc": 0.6904186795491144, + "k": 5, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1874, + 1960 + ] + }, + { + "test_f1": 0.6644026165666426, + "test_acc": 0.6650847457627118, + "test_auc": 0.727348107945123, + "val_auc": 0.7355072463768116, + "k": 10, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1874, + 1960, + 186, + 1335, + 908, + 1255, + 100 + ] + }, + { + "test_f1": 0.6774053028835297, + "test_acc": 0.6779661016949152, + "test_auc": 0.7323912587836869, + "val_auc": 0.7413446054750402, + "k": 16, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1874, + 1960, + 186, + 1335, + 908, + 1255, + 100, + 253, + 311, + 765, + 1500, + 3780, + 6478 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2ace61764b384f5cbe5cdbe36d5a17e486b547e9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6005326288434044, + "test_acc": 0.6081355932203389, + "test_auc": 0.6186032145260394, + "val_auc": 0.6497584541062802, + "k": 1, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.5955433552564344, + "test_acc": 0.6047457627118644, + "test_auc": 0.6186032145260394, + "val_auc": 0.6497584541062802, + "k": 2, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255 + ] + }, + { + "test_f1": 0.6127813760605284, + "test_acc": 0.6149152542372881, + "test_auc": 0.6440873974547072, + "val_auc": 0.6851851851851852, + "k": 5, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 186, + 12485 + ] + }, + { + "test_f1": 0.609489708964015, + "test_acc": 0.6115254237288136, + "test_auc": 0.6351924781120266, + "val_auc": 0.6835748792270531, + "k": 10, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 186, + 12485, + 4511, + 521, + 1942, + 567, + 1500 + ] + }, + { + "test_f1": 0.6156542747235041, + "test_acc": 0.6183050847457627, + "test_auc": 0.6504947546083331, + "val_auc": 0.6884057971014492, + "k": 16, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 186, + 12485, + 4511, + 521, + 1942, + 567, + 1500, + 311, + 908, + 1596, + 6212, + 583, + 598 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5543b28dd9bdb1a6aefa57d9abe910c601bd136f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5508684863523573, + "test_acc": 0.5508684863523573, + "test_auc": 0.5750701935865228, + "val_auc": 0.5592236894757904, + "k": 1, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067 + ] + }, + { + "test_f1": 0.5562473678976209, + "test_acc": 0.5583126550868487, + "test_auc": 0.5853159942859958, + "val_auc": 0.5370148059223689, + "k": 2, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 5731 + ] + }, + { + "test_f1": 0.5978762386315494, + "test_acc": 0.598014888337469, + "test_auc": 0.6377518348849811, + "val_auc": 0.611844737895158, + "k": 5, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 5731, + 1268, + 24630, + 1441 + ] + }, + { + "test_f1": 0.575682382133995, + "test_acc": 0.575682382133995, + "test_auc": 0.6241687601595981, + "val_auc": 0.5894357743097239, + "k": 10, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 5731, + 1268, + 24630, + 1441, + 8013, + 1619, + 1101, + 124, + 1551 + ] + }, + { + "test_f1": 0.6015052511887001, + "test_acc": 0.6029776674937966, + "test_auc": 0.6510024136742033, + "val_auc": 0.6674669867947178, + "k": 16, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 5731, + 1268, + 24630, + 1441, + 8013, + 1619, + 1101, + 124, + 1551, + 10, + 1674, + 1955, + 2480, + 1120, + 61 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..809494add206fb4fcdf2b593314a8997ae9a3e00 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5616392620052463, + "test_acc": 0.5905707196029777, + "test_auc": 0.6037879907393724, + "val_auc": 0.6330532212885155, + "k": 1, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575 + ] + }, + { + "test_f1": 0.6008073619954063, + "test_acc": 0.6054590570719603, + "test_auc": 0.6560637407024285, + "val_auc": 0.6606642657062826, + "k": 2, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 1625 + ] + }, + { + "test_f1": 0.6015052511887001, + "test_acc": 0.6029776674937966, + "test_auc": 0.6460765479533028, + "val_auc": 0.6670668267306923, + "k": 5, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 1625, + 28711, + 583, + 434 + ] + }, + { + "test_f1": 0.6268263877875335, + "test_acc": 0.6277915632754343, + "test_auc": 0.6602630412294961, + "val_auc": 0.6712685074029612, + "k": 10, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 1625, + 28711, + 583, + 434, + 1381, + 1619, + 10, + 4079, + 1551 + ] + }, + { + "test_f1": 0.6166715411532361, + "test_acc": 0.6178660049627791, + "test_auc": 0.6780454164819468, + "val_auc": 0.6834733893557423, + "k": 16, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 1625, + 28711, + 583, + 434, + 1381, + 1619, + 10, + 4079, + 1551, + 1799, + 124, + 680, + 1653, + 433, + 765 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8c9758f513018f339b5e360ff74b34bca53aa5e5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.547433974162972, + "test_acc": 0.5483870967741935, + "test_auc": 0.5719668981823555, + "val_auc": 0.5468187274909964, + "k": 1, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619 + ] + }, + { + "test_f1": 0.56254271106352, + "test_acc": 0.5632754342431762, + "test_auc": 0.605967686320871, + "val_auc": 0.6346538615446178, + "k": 2, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042 + ] + }, + { + "test_f1": 0.6374715635885743, + "test_acc": 0.6377171215880894, + "test_auc": 0.7130436924289444, + "val_auc": 0.6774709883953581, + "k": 5, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 125, + 32279 + ] + }, + { + "test_f1": 0.6476296617731545, + "test_acc": 0.6476426799007444, + "test_auc": 0.7373035811043791, + "val_auc": 0.7022809123649459, + "k": 10, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 125, + 32279, + 1662, + 347, + 1067, + 635, + 1625 + ] + }, + { + "test_f1": 0.6946838010862356, + "test_acc": 0.6947890818858561, + "test_auc": 0.7471060538889709, + "val_auc": 0.6906762705082033, + "k": 16, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 125, + 32279, + 1662, + 347, + 1067, + 635, + 1625, + 7319, + 772, + 10, + 1872, + 2536, + 2480 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7bd41ef8d119cd0931cde63e313e78c4226658c0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8321968160523074, + "test_acc": 0.8340026773761714, + "test_auc": 0.8541848862381902, + "val_auc": 0.8402020202020202, + "k": 1, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731 + ] + }, + { + "test_f1": 0.8508814373961711, + "test_acc": 0.8527443105756358, + "test_auc": 0.9180621066364639, + "val_auc": 0.896969696969697, + "k": 2, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648 + ] + }, + { + "test_f1": 0.8664429651567891, + "test_acc": 0.8674698795180723, + "test_auc": 0.9546637324196069, + "val_auc": 0.9632323232323231, + "k": 5, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648, + 23226, + 1274, + 737 + ] + }, + { + "test_f1": 0.8637448432407647, + "test_acc": 0.8647925033467202, + "test_auc": 0.9578966609797709, + "val_auc": 0.9636363636363636, + "k": 10, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648, + 23226, + 1274, + 737, + 23697, + 6567, + 1158, + 13701, + 292 + ] + }, + { + "test_f1": 0.8745957720282534, + "test_acc": 0.8755020080321285, + "test_auc": 0.9617568206907428, + "val_auc": 0.9660606060606061, + "k": 16, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648, + 23226, + 1274, + 737, + 23697, + 6567, + 1158, + 13701, + 292, + 5872, + 508, + 10069, + 1101, + 21549, + 2046 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b8451bc8124dcf7f7a0b4f17cd59caf70cbb91c3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8723309734723812, + "test_acc": 0.8741633199464525, + "test_auc": 0.8746684635345728, + "val_auc": 0.8777777777777778, + "k": 1, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226 + ] + }, + { + "test_f1": 0.8723309734723812, + "test_acc": 0.8741633199464525, + "test_auc": 0.9222520107238606, + "val_auc": 0.9220202020202021, + "k": 2, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648 + ] + }, + { + "test_f1": 0.8747338912574095, + "test_acc": 0.8755020080321285, + "test_auc": 0.9662657166205504, + "val_auc": 0.9632323232323232, + "k": 5, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648, + 4731, + 4505, + 1065 + ] + }, + { + "test_f1": 0.8747338912574095, + "test_acc": 0.8755020080321285, + "test_auc": 0.9662657166205504, + "val_auc": 0.9632323232323232, + "k": 10, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648, + 4731, + 4505, + 1065, + 32, + 1380, + 6567, + 1682, + 1447 + ] + }, + { + "test_f1": 0.8900320924855419, + "test_acc": 0.8902275769745649, + "test_auc": 0.9699072414732406, + "val_auc": 0.9660606060606061, + "k": 16, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648, + 4731, + 4505, + 1065, + 32, + 1380, + 6567, + 1682, + 1447, + 805, + 7481, + 10069, + 3255, + 1744, + 1274 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9f893f5de98a20aab00049d138afa10bf8e3f44e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8702978752875538, + "test_acc": 0.8714859437751004, + "test_auc": 0.8846826568794713, + "val_auc": 0.8715151515151515, + "k": 1, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648 + ] + }, + { + "test_f1": 0.8772644864456856, + "test_acc": 0.8781793842034806, + "test_auc": 0.9169832690570744, + "val_auc": 0.8933333333333334, + "k": 2, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505 + ] + }, + { + "test_f1": 0.902143185030189, + "test_acc": 0.9022757697456493, + "test_auc": 0.9780074837636736, + "val_auc": 0.9696969696969697, + "k": 5, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 4731, + 23226, + 32 + ] + }, + { + "test_f1": 0.8941981344820025, + "test_acc": 0.8942436412315931, + "test_auc": 0.9756634313486545, + "val_auc": 0.972929292929293, + "k": 10, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 4731, + 23226, + 32, + 2832, + 1682, + 1305, + 1158, + 1065 + ] + }, + { + "test_f1": 0.8955430171268094, + "test_acc": 0.8955823293172691, + "test_auc": 0.97589281874095, + "val_auc": 0.972929292929293, + "k": 16, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 4731, + 23226, + 32, + 2832, + 1682, + 1305, + 1158, + 1065, + 1101, + 1380, + 22848, + 6828, + 1643, + 23 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..af1b338d39e1df32f335dcded99a60544e117886 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7172943190903961, + "test_acc": 0.7364102564102564, + "test_auc": 0.7357146463796412, + "val_auc": 0.7127702978376174, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518 + ] + }, + { + "test_f1": 0.7568705418570647, + "test_acc": 0.7661538461538462, + "test_auc": 0.7811837277409365, + "val_auc": 0.7637698898408813, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541 + ] + }, + { + "test_f1": 0.7565016011778936, + "test_acc": 0.7651282051282051, + "test_auc": 0.8566920254485475, + "val_auc": 0.8514891880864953, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 821 + ] + }, + { + "test_f1": 0.7725073091850416, + "test_acc": 0.7774358974358975, + "test_auc": 0.876657858417208, + "val_auc": 0.8763769889840881, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 821, + 1335, + 374, + 253, + 505, + 2632 + ] + }, + { + "test_f1": 0.7896065196819055, + "test_acc": 0.7928205128205128, + "test_auc": 0.8911241458242165, + "val_auc": 0.890656874745002, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 821, + 1335, + 374, + 253, + 505, + 2632, + 2581, + 7813, + 1030, + 549, + 4895, + 1408 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4383cfdaf12be05871b2354d0e0d2e1c360d72bf --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8771398255480949, + "test_acc": 0.8783193277310924, + "test_auc": 0.8841790138186867, + "val_auc": 0.8924126856684063, + "k": 1, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852 + ] + }, + { + "test_f1": 0.9010985310834012, + "test_acc": 0.9015126050420168, + "test_auc": 0.9132517662031512, + "val_auc": 0.9498193496587716, + "k": 2, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782 + ] + }, + { + "test_f1": 0.911979809585346, + "test_acc": 0.9122689075630253, + "test_auc": 0.9639975667252387, + "val_auc": 0.9877559213167403, + "k": 5, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 60, + 1318, + 173 + ] + }, + { + "test_f1": 0.9197332663446008, + "test_acc": 0.92, + "test_auc": 0.976585605715484, + "val_auc": 0.9811320754716982, + "k": 10, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 60, + 1318, + 173, + 1275, + 11077, + 8815, + 1688, + 1283 + ] + }, + { + "test_f1": 0.9204139296216219, + "test_acc": 0.920672268907563, + "test_auc": 0.9783332790998691, + "val_auc": 0.9779205138498595, + "k": 16, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 60, + 1318, + 173, + 1275, + 11077, + 8815, + 1688, + 1283, + 259, + 42, + 6919, + 23720, + 98, + 1152 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..69e2a355bc0b49936d548901df1a1c5df8930e9d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7489895150382037, + "test_acc": 0.7552941176470588, + "test_auc": 0.7604297730871857, + "val_auc": 0.7671617824167001, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021 + ] + }, + { + "test_f1": 0.7764676984720892, + "test_acc": 0.7771428571428571, + "test_auc": 0.8811256697832831, + "val_auc": 0.8879967884383781, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914 + ] + }, + { + "test_f1": 0.780980512879357, + "test_acc": 0.7825210084033614, + "test_auc": 0.894484501883709, + "val_auc": 0.9305499799277398, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 4790, + 1180 + ] + }, + { + "test_f1": 0.8045683752253956, + "test_acc": 0.8053781512605042, + "test_auc": 0.9051406996839997, + "val_auc": 0.9217181854676837, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 4790, + 1180, + 6919, + 1275, + 1067, + 8815, + 1794 + ] + }, + { + "test_f1": 0.8331159805466657, + "test_acc": 0.8332773109243697, + "test_auc": 0.9294969484637468, + "val_auc": 0.9554395824969892, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 4790, + 1180, + 6919, + 1275, + 1067, + 8815, + 1794, + 1500, + 1091, + 2962, + 1152, + 98, + 347 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9b47a0afba4d9dea9366e4bf86ca24296d7b7c5f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7656268434631189, + "test_acc": 0.7727731092436975, + "test_auc": 0.7840834725325582, + "val_auc": 0.7615415495784825, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235 + ] + }, + { + "test_f1": 0.8049208624641269, + "test_acc": 0.8084033613445378, + "test_auc": 0.8330047689292869, + "val_auc": 0.7960658370132476, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 764 + ] + }, + { + "test_f1": 0.8151914849229767, + "test_acc": 0.8181512605042017, + "test_auc": 0.8589844060712556, + "val_auc": 0.8362103572862305, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 764, + 5189, + 8189, + 3494 + ] + }, + { + "test_f1": 0.8316029383512124, + "test_acc": 0.8326050420168067, + "test_auc": 0.9226156257457101, + "val_auc": 0.9191087916499399, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 764, + 5189, + 8189, + 3494, + 42, + 692, + 576, + 1021, + 30916 + ] + }, + { + "test_f1": 0.8425351327392072, + "test_acc": 0.8433613445378151, + "test_auc": 0.9357017087156793, + "val_auc": 0.9205138498594941, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 764, + 5189, + 8189, + 3494, + 42, + 692, + 576, + 1021, + 30916, + 17338, + 1521, + 8568, + 2504, + 2046, + 964 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..31e44ff1adfb31bf0705e7c68efb7c8361247f07 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.945210009747314, + "test_acc": 0.9452100840336134, + "test_auc": 0.9864714623511291, + "val_auc": 0.9851465274989964, + "k": 1, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9872264491265804, + "test_acc": 0.9872268907563025, + "test_auc": 0.9994924651640382, + "val_auc": 0.9995985547972703, + "k": 2, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847 + ] + }, + { + "test_f1": 0.9842000204583102, + "test_acc": 0.9842016806722689, + "test_auc": 0.9995087352033032, + "val_auc": 0.9995985547972703, + "k": 5, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1411, + 786 + ] + }, + { + "test_f1": 0.9996638655462184, + "test_acc": 0.9996638655462184, + "test_auc": 0.9999954805446486, + "val_auc": 1.0, + "k": 10, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1411, + 786, + 1758, + 461, + 279, + 1697, + 431 + ] + }, + { + "test_f1": 1.0, + "test_acc": 1.0, + "test_auc": 0.9999999999999999, + "val_auc": 1.0, + "k": 16, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1411, + 786, + 1758, + 461, + 279, + 1697, + 431, + 27673, + 1634, + 4148, + 5020, + 1360, + 4674 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..385e9442b12be7e2661e3dce6814a6ec91fe9d5e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9381474028619566, + "test_acc": 0.9381512605042017, + "test_auc": 0.9867372063257912, + "val_auc": 0.9871537535126456, + "k": 1, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9623463044128202, + "test_acc": 0.9623529411764706, + "test_auc": 0.9956965746144001, + "val_auc": 0.9967884383781613, + "k": 2, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.96300616706425, + "test_acc": 0.9630252100840336, + "test_auc": 0.9966623822229935, + "val_auc": 0.9971898835808912, + "k": 5, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758 + ] + }, + { + "test_f1": 0.9915960689392034, + "test_acc": 0.9915966386554622, + "test_auc": 0.9991874019278189, + "val_auc": 0.9991971095945403, + "k": 10, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758, + 431, + 1847, + 279, + 1628, + 126 + ] + }, + { + "test_f1": 0.998655462184874, + "test_acc": 0.998655462184874, + "test_auc": 0.9999864416339458, + "val_auc": 1.0, + "k": 16, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758, + 431, + 1847, + 279, + 1628, + 126, + 461, + 1634, + 1472, + 1914, + 1551, + 1360 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bfe45d61ef636880e15e18323907ddd1ff4bcc19 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9287382397456757, + "test_acc": 0.9287394957983194, + "test_auc": 0.979435348287307, + "val_auc": 0.9775190686471297, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9660481185143814, + "test_acc": 0.9660504201680672, + "test_auc": 0.9960260429095169, + "val_auc": 0.997591328783621, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.975453332112914, + "test_acc": 0.9754621848739495, + "test_auc": 0.999100628385072, + "val_auc": 1.0, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 1411 + ] + }, + { + "test_f1": 0.9979831864411416, + "test_acc": 0.9979831932773109, + "test_auc": 0.9997695077770787, + "val_auc": 1.0, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 1411, + 461, + 1758, + 279, + 431, + 4148 + ] + }, + { + "test_f1": 0.9973109237620901, + "test_acc": 0.9973109243697479, + "test_auc": 0.9999823741241296, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 1411, + 461, + 1758, + 279, + 431, + 4148, + 27673, + 1360, + 1634, + 1697, + 1914, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..586facff463176a60cd1fc81ca7667e3749b59cb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7292671647643266, + "test_acc": 0.730062893081761, + "test_auc": 0.7959966644355312, + "val_auc": 0.7921168467386954, + "k": 1, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.753476220414611, + "test_acc": 0.7539622641509434, + "test_auc": 0.82081112745927, + "val_auc": 0.8497398959583834, + "k": 2, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 20310 + ] + }, + { + "test_f1": 0.8491380128864392, + "test_acc": 0.8493081761006289, + "test_auc": 0.9199858182816071, + "val_auc": 0.9631852741096438, + "k": 5, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 20310, + 1585, + 776, + 508 + ] + }, + { + "test_f1": 0.8595101109765322, + "test_acc": 0.859622641509434, + "test_auc": 0.9282762503556822, + "val_auc": 0.9631852741096439, + "k": 10, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 20310, + 1585, + 776, + 508, + 32700, + 10363, + 1866, + 1499, + 1327 + ] + }, + { + "test_f1": 0.8585329789810889, + "test_acc": 0.8586163522012579, + "test_auc": 0.930739950523473, + "val_auc": 0.9627851140456183, + "k": 16, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 20310, + 1585, + 776, + 508, + 32700, + 10363, + 1866, + 1499, + 1327, + 1608, + 4605, + 320, + 19688, + 28521, + 2837 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..80c67cf141290cb8131f1c59a17854f4f50b7aa3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7193998399203513, + "test_acc": 0.7194928684627575, + "test_auc": 0.717279485633916, + "val_auc": 0.7398959583833533, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.7606973058637084, + "test_acc": 0.7606973058637084, + "test_auc": 0.8203887884267631, + "val_auc": 0.8105242096838736, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 913 + ] + }, + { + "test_f1": 0.7939674636579838, + "test_acc": 0.7939778129952456, + "test_auc": 0.8523508137432187, + "val_auc": 0.8595438175270108, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 913, + 508, + 28521, + 1327 + ] + }, + { + "test_f1": 0.8096248457730458, + "test_acc": 0.8098256735340729, + "test_auc": 0.8617540687160941, + "val_auc": 0.86234493797519, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 913, + 508, + 28521, + 1327, + 1866, + 5186, + 1608, + 1681, + 11098 + ] + }, + { + "test_f1": 0.8095836703522572, + "test_acc": 0.8098256735340729, + "test_auc": 0.8802692384970866, + "val_auc": 0.8875550220088035, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 913, + 508, + 28521, + 1327, + 1866, + 5186, + 1608, + 1681, + 11098, + 1585, + 431, + 21219, + 435, + 20310, + 19666 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ac2dd9e5326974be734ecf894b48ce897b4590e1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7027436572606287, + "test_acc": 0.7034813925570228, + "test_auc": 0.7123570374469654, + "val_auc": 0.7351046698872785, + "k": 1, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.717850558496101, + "test_acc": 0.7178871548619448, + "test_auc": 0.7858732244973252, + "val_auc": 0.7844202898550724, + "k": 2, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501 + ] + }, + { + "test_f1": 0.7514898559152999, + "test_acc": 0.751500600240096, + "test_auc": 0.8273928933776055, + "val_auc": 0.8361513687600645, + "k": 5, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501, + 1447, + 1528, + 1417 + ] + }, + { + "test_f1": 0.7778759523596435, + "test_acc": 0.7779111644657863, + "test_auc": 0.8547748339789708, + "val_auc": 0.8079710144927535, + "k": 10, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501, + 1447, + 1528, + 1417, + 2837, + 5186, + 32700, + 347, + 1866 + ] + }, + { + "test_f1": 0.7790536937833429, + "test_acc": 0.7791116446578632, + "test_auc": 0.8602656336469287, + "val_auc": 0.8140096618357487, + "k": 16, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501, + 1447, + 1528, + 1417, + 2837, + 5186, + 32700, + 347, + 1866, + 776, + 913, + 1377, + 1804, + 559, + 3507 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9b545bf1bdb0ce8055755909b19f3bf2a9e9c5e6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7266378263981824, + "test_acc": 0.7272727272727273, + "test_auc": 0.7463116993503734, + "val_auc": 0.7102841136454582, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7017869040824486, + "test_acc": 0.7024793388429752, + "test_auc": 0.7959140307206605, + "val_auc": 0.7591036414565826, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 6501 + ] + }, + { + "test_f1": 0.7437977641763308, + "test_acc": 0.743801652892562, + "test_auc": 0.797872017485277, + "val_auc": 0.787515006002401, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 6501, + 5186, + 3507, + 1377 + ] + }, + { + "test_f1": 0.7437977641763308, + "test_acc": 0.743801652892562, + "test_auc": 0.7984639669722543, + "val_auc": 0.8083233293317328, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 6501, + 5186, + 3507, + 1377, + 508, + 1681, + 1528, + 12612, + 913 + ] + }, + { + "test_f1": 0.7603269406810835, + "test_acc": 0.7603305785123967, + "test_auc": 0.834041648958776, + "val_auc": 0.8083233293317327, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 6501, + 5186, + 3507, + 1377, + 508, + 1681, + 1528, + 12612, + 913, + 1447, + 601, + 3301, + 1585, + 19688, + 1417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6d30b1c5a3388008814d76f9b746e5a9e0e0aa90 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6277772470975926, + "test_acc": 0.6324528301886793, + "test_auc": 0.6421917260989185, + "val_auc": 0.6422569027611045, + "k": 1, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6212506993290443, + "test_acc": 0.6254088050314466, + "test_auc": 0.6878887821139216, + "val_auc": 0.750500200080032, + "k": 2, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 4605 + ] + }, + { + "test_f1": 0.6638949040220143, + "test_acc": 0.6638993710691824, + "test_auc": 0.7171258552826775, + "val_auc": 0.7733093237294918, + "k": 5, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 4605, + 2837, + 1585, + 776 + ] + }, + { + "test_f1": 0.6804590407877987, + "test_acc": 0.6810062893081761, + "test_auc": 0.7463548275055467, + "val_auc": 0.8075230092036815, + "k": 10, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 4605, + 2837, + 1585, + 776, + 27673, + 12749, + 1804, + 32700, + 126 + ] + }, + { + "test_f1": 0.6726938069345733, + "test_acc": 0.6729559748427673, + "test_auc": 0.7427104144747702, + "val_auc": 0.8163265306122449, + "k": 16, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 4605, + 2837, + 1585, + 776, + 27673, + 12749, + 1804, + 32700, + 126, + 1411, + 1866, + 1499, + 1447, + 20310, + 25969 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..392215a3c74048ef757212e60cb558414c9dfc2f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7995115567489557, + "test_acc": 0.800251572327044, + "test_auc": 0.8401121373434367, + "val_auc": 0.8279311724689875, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293 + ] + }, + { + "test_f1": 0.8072552692199677, + "test_acc": 0.8072955974842767, + "test_auc": 0.8874122946030486, + "val_auc": 0.8823529411764706, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062 + ] + }, + { + "test_f1": 0.8208670119866824, + "test_acc": 0.8208805031446541, + "test_auc": 0.9063247122392127, + "val_auc": 0.9179671868747499, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1166 + ] + }, + { + "test_f1": 0.8371884995527431, + "test_acc": 0.8372327044025157, + "test_auc": 0.9172744063778747, + "val_auc": 0.9327731092436975, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1166, + 1313, + 1441, + 1116, + 1305, + 409 + ] + }, + { + "test_f1": 0.8422501137218609, + "test_acc": 0.8422641509433962, + "test_auc": 0.9207611041184196, + "val_auc": 0.9379751900760304, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1166, + 1313, + 1441, + 1116, + 1305, + 409, + 1928, + 14, + 4110, + 765, + 524, + 1737 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9f9a2e3ac883c5d87b0bd57d4af0d22dd5b0ad26 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6612819716267991, + "test_acc": 0.6666666666666666, + "test_auc": 0.676530612244898, + "val_auc": 0.6057808109193095, + "k": 1, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377 + ] + }, + { + "test_f1": 0.6500763448437867, + "test_acc": 0.6565656565656566, + "test_auc": 0.7, + "val_auc": 0.6573665194700923, + "k": 2, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027 + ] + }, + { + "test_f1": 0.6356196994494867, + "test_acc": 0.6363636363636364, + "test_auc": 0.7279591836734693, + "val_auc": 0.6918908069048575, + "k": 5, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 681 + ] + }, + { + "test_f1": 0.6963497495412388, + "test_acc": 0.696969696969697, + "test_auc": 0.730204081632653, + "val_auc": 0.7129666800481734, + "k": 10, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 681, + 1339, + 474, + 1551, + 524, + 505 + ] + }, + { + "test_f1": 0.6864846729946372, + "test_acc": 0.6868686868686869, + "test_auc": 0.7085714285714286, + "val_auc": 0.7442794058611, + "k": 16, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 681, + 1339, + 474, + 1551, + 524, + 505, + 56, + 1790, + 495, + 3418, + 1221, + 100 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..73ec703de684c26da99b614b402aa1eb866c0815 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6691599188332341, + "test_acc": 0.6713928273561302, + "test_auc": 0.6969616026711185, + "val_auc": 0.8055555555555555, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6853491196980299, + "test_acc": 0.6889074228523769, + "test_auc": 0.7164788536449638, + "val_auc": 0.8623188405797102, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114 + ] + }, + { + "test_f1": 0.733574226372386, + "test_acc": 0.7364470391993327, + "test_auc": 0.7880272676683361, + "val_auc": 0.856682769726248, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 238 + ] + }, + { + "test_f1": 0.769298605324354, + "test_acc": 0.7706422018348624, + "test_auc": 0.8491624930439622, + "val_auc": 0.8856682769726248, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 238, + 6086, + 30704, + 3276, + 801, + 4481 + ] + }, + { + "test_f1": 0.7758926184571007, + "test_acc": 0.7773144286905754, + "test_auc": 0.8715498052309405, + "val_auc": 0.9086151368760065, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 238, + 6086, + 30704, + 3276, + 801, + 4481, + 363, + 756, + 1670, + 4534, + 6719, + 1335 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ce786a65ef4bae56983062e3b64b37ec1b8ecc24 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.559525184520645, + "test_acc": 0.5861635220125786, + "test_auc": 0.6362333538219755, + "val_auc": 0.5496198479391757, + "k": 2, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727 + ] + }, + { + "test_f1": 0.6245373754086625, + "test_acc": 0.6254088050314466, + "test_auc": 0.6826896203592971, + "val_auc": 0.586234493797519, + "k": 5, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 486, + 6400 + ] + }, + { + "test_f1": 0.6371294135799064, + "test_acc": 0.6377358490566037, + "test_auc": 0.6945632020608806, + "val_auc": 0.5978391356542617, + "k": 10, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 486, + 6400, + 1928, + 588, + 293, + 1438, + 7143 + ] + }, + { + "test_f1": 0.627419685881474, + "test_acc": 0.6274213836477988, + "test_auc": 0.6898481984002658, + "val_auc": 0.5906362545018007, + "k": 16, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 486, + 6400, + 1928, + 588, + 293, + 1438, + 7143, + 765, + 1848, + 140, + 2034, + 1608, + 238 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ffd624ac66c53153006b97237a7118e1c0624fa4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5058526940303268, + "test_acc": 0.5463794135248354, + "test_auc": 0.5492622410681031, + "val_auc": 0.5616218386190285, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498 + ] + }, + { + "test_f1": 0.5540470697042121, + "test_acc": 0.5541591861160982, + "test_auc": 0.5839569664498754, + "val_auc": 0.5234845443596949, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727 + ] + }, + { + "test_f1": 0.5570270496306249, + "test_acc": 0.5577498503889886, + "test_auc": 0.5869230438644243, + "val_auc": 0.560417503010839, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 462, + 293, + 5392 + ] + }, + { + "test_f1": 0.5552193286773577, + "test_acc": 0.5565529622980251, + "test_auc": 0.5771423659857319, + "val_auc": 0.5796868727418707, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 462, + 293, + 5392, + 860, + 150, + 2797, + 1336, + 672 + ] + }, + { + "test_f1": 0.5467957452936197, + "test_acc": 0.5469778575703171, + "test_auc": 0.5770578460304272, + "val_auc": 0.5692492974708953, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 462, + 293, + 5392, + 860, + 150, + 2797, + 1336, + 672, + 6142, + 6417, + 5037, + 4218, + 4913, + 100 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b435a5f8dfc0b0ac7770fd21a14b8fe2ac00cfd6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5697039951616865, + "test_acc": 0.600503144654088, + "test_auc": 0.6092508751553103, + "val_auc": 0.6178471388555422, + "k": 1, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659 + ] + }, + { + "test_f1": 0.5936241570004224, + "test_acc": 0.610062893081761, + "test_auc": 0.6158109958188992, + "val_auc": 0.6798719487795117, + "k": 2, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794 + ] + }, + { + "test_f1": 0.6039108196302213, + "test_acc": 0.6203773584905661, + "test_auc": 0.6233204460785853, + "val_auc": 0.7034813925570228, + "k": 5, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 6400 + ] + }, + { + "test_f1": 0.5930007374287459, + "test_acc": 0.6075471698113207, + "test_auc": 0.6825991176044692, + "val_auc": 0.7835134053621449, + "k": 10, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 6400, + 777, + 15353, + 7143, + 1377, + 9390 + ] + }, + { + "test_f1": 0.6230474548226522, + "test_acc": 0.6296855345911949, + "test_auc": 0.7095826088893704, + "val_auc": 0.8263305322128852, + "k": 16, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 6400, + 777, + 15353, + 7143, + 1377, + 9390, + 3053, + 121, + 1596, + 4252, + 598, + 448 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6d781354536f4c654fb67c4543c7b6e83ef7bc23 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.994968466843654, + "test_acc": 0.9949685534591195, + "test_auc": 0.9994625528713297, + "val_auc": 1.0, + "k": 1, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9982389874693722, + "test_acc": 0.9982389937106918, + "test_auc": 0.9999827854899908, + "val_auc": 1.0, + "k": 2, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625 + ] + }, + { + "test_f1": 0.9989937094180884, + "test_acc": 0.9989937106918239, + "test_auc": 0.9999870891174931, + "val_auc": 1.0, + "k": 5, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1525 + ] + }, + { + "test_f1": 0.9987421359765241, + "test_acc": 0.9987421383647799, + "test_auc": 0.9999903801267596, + "val_auc": 1.0, + "k": 10, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1525, + 1551, + 347, + 6718, + 1255, + 1232 + ] + }, + { + "test_f1": 0.9992452824456872, + "test_acc": 0.999245283018868, + "test_auc": 0.9999906332813185, + "val_auc": 1.0, + "k": 16, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1525, + 1551, + 347, + 6718, + 1255, + 1232, + 1042, + 15467, + 1758, + 374, + 1472, + 583 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b7f00fe98ee5c2fee1fce22a1921dbd4e7d483dc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6454654818916407, + "test_acc": 0.6455345911949686, + "test_auc": 0.6905336143686477, + "val_auc": 0.7639055622248899, + "k": 1, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6478866733191273, + "test_acc": 0.6483018867924528, + "test_auc": 0.6983833043555748, + "val_auc": 0.7511004401760704, + "k": 2, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928 + ] + }, + { + "test_f1": 0.6879465109312953, + "test_acc": 0.6880503144654088, + "test_auc": 0.760061627945833, + "val_auc": 0.8427370948379351, + "k": 5, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 1616 + ] + }, + { + "test_f1": 0.7016287619580798, + "test_acc": 0.7016352201257862, + "test_auc": 0.7728618819104865, + "val_auc": 0.8419367747098839, + "k": 10, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 1616, + 6400, + 56, + 1255, + 1135, + 152 + ] + }, + { + "test_f1": 0.7109420790532165, + "test_acc": 0.7109433962264151, + "test_auc": 0.7964176604670803, + "val_auc": 0.8627450980392157, + "k": 16, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 1616, + 6400, + 56, + 1255, + 1135, + 152, + 1691, + 5827, + 406, + 1327, + 19000, + 121 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d0b3c29f64e13abab17bec6b5aa1be5a60165564 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6660453644628225, + "test_acc": 0.6676729559748428, + "test_auc": 0.7159418514104253, + "val_auc": 0.716486594637855, + "k": 1, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6659475681980369, + "test_acc": 0.6684276729559748, + "test_auc": 0.7159418514104253, + "val_auc": 0.716486594637855, + "k": 2, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143 + ] + }, + { + "test_f1": 0.6901387998707214, + "test_acc": 0.6913207547169812, + "test_auc": 0.7620473723063088, + "val_auc": 0.7438975590236094, + "k": 5, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 1011, + 1928 + ] + }, + { + "test_f1": 0.7092232131356375, + "test_acc": 0.709685534591195, + "test_auc": 0.7804943652858266, + "val_auc": 0.7603041216486595, + "k": 10, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 1011, + 1928, + 19940, + 6400, + 13614, + 30704, + 598 + ] + }, + { + "test_f1": 0.7123069856938826, + "test_acc": 0.7124528301886792, + "test_auc": 0.7847960941289407, + "val_auc": 0.767106842737095, + "k": 16, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 1011, + 1928, + 19940, + 6400, + 13614, + 30704, + 598, + 7068, + 253, + 1019, + 1960, + 1135, + 9390 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0f91ad4d56ce38d55f17f0a24554d5453b2c8bec --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7809133805266738, + "test_acc": 0.7816031537450723, + "test_auc": 0.8030043326354481, + "val_auc": 0.7152, + "k": 1, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062 + ] + }, + { + "test_f1": 0.7825113212873673, + "test_acc": 0.7865965834428383, + "test_auc": 0.8836566371211982, + "val_auc": 0.8062, + "k": 2, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448 + ] + }, + { + "test_f1": 0.8419214024448101, + "test_acc": 0.8420499342969776, + "test_auc": 0.911647611580144, + "val_auc": 0.8384, + "k": 5, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 140 + ] + }, + { + "test_f1": 0.8431009659135266, + "test_acc": 0.8431011826544021, + "test_auc": 0.9193594374480937, + "val_auc": 0.8648, + "k": 10, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 140, + 30127, + 1030, + 628, + 817, + 1662 + ] + }, + { + "test_f1": 0.8507224239269006, + "test_acc": 0.8507227332457293, + "test_auc": 0.9255649804144543, + "val_auc": 0.8884, + "k": 16, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 140, + 30127, + 1030, + 628, + 817, + 1662, + 1313, + 2046, + 1942, + 24630, + 2831, + 296 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..71f4a5f493bdb401001cfdb9efa5604e21314b86 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5149712021145977, + "test_acc": 0.5197484276729559, + "test_auc": 0.5343626935240025, + "val_auc": 0.6042416966786714, + "k": 1, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011 + ] + }, + { + "test_f1": 0.5238815835630615, + "test_acc": 0.5260377358490566, + "test_auc": 0.5381450757894118, + "val_auc": 0.6498599439775911, + "k": 2, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411 + ] + }, + { + "test_f1": 0.5238815835630615, + "test_acc": 0.5260377358490566, + "test_auc": 0.5381450757894118, + "val_auc": 0.6498599439775911, + "k": 5, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928 + ] + }, + { + "test_f1": 0.5238815835630615, + "test_acc": 0.5260377358490566, + "test_auc": 0.5381450757894118, + "val_auc": 0.6498599439775911, + "k": 10, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928, + 1703, + 786, + 4880, + 1202, + 6718 + ] + }, + { + "test_f1": 0.5238815835630615, + "test_acc": 0.5260377358490566, + "test_auc": 0.5381450757894118, + "val_auc": 0.6498599439775911, + "k": 16, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928, + 1703, + 786, + 4880, + 1202, + 6718, + 18461, + 1596, + 1417, + 1643, + 13614, + 448 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6556efc9335fbaea866b3555819d184e8c42418a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9932073485568238, + "test_acc": 0.9932075471698113, + "test_auc": 0.9968912620159811, + "val_auc": 1.0, + "k": 1, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.9942138035199232, + "test_acc": 0.9942138364779874, + "test_auc": 0.9986960008668012, + "val_auc": 1.0, + "k": 2, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988 + ] + }, + { + "test_f1": 0.9927043184801099, + "test_acc": 0.9927044025157232, + "test_auc": 0.9994030615499743, + "val_auc": 1.0, + "k": 5, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 1744, + 152, + 487 + ] + }, + { + "test_f1": 0.9904399454202186, + "test_acc": 0.9904402515723271, + "test_auc": 0.9988658675758628, + "val_auc": 1.0, + "k": 10, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 1744, + 152, + 487, + 292, + 6935, + 6475, + 821, + 559 + ] + }, + { + "test_f1": 0.992704291699109, + "test_acc": 0.9927044025157232, + "test_auc": 0.9988119456548046, + "val_auc": 1.0, + "k": 16, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 1744, + 152, + 487, + 292, + 6935, + 6475, + 821, + 559, + 347, + 3417, + 771, + 1960, + 334, + 1132 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fa7f58c3d372a4e43ede69832a20077dda892ae9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.861585354755887, + "test_acc": 0.8618867924528302, + "test_auc": 0.8942769602010655, + "val_auc": 0.9251700680272109, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487 + ] + }, + { + "test_f1": 0.8960406240088914, + "test_acc": 0.8961006289308177, + "test_auc": 0.9402104625741363, + "val_auc": 0.9703881552621048, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744 + ] + }, + { + "test_f1": 0.9073497872999161, + "test_acc": 0.9074213836477988, + "test_auc": 0.9642452095562809, + "val_auc": 0.9725890356142457, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 771, + 1524 + ] + }, + { + "test_f1": 0.9310139254615947, + "test_acc": 0.9310691823899371, + "test_auc": 0.9758755097267044, + "val_auc": 0.9719887955182074, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 771, + 1524, + 7368, + 6475, + 559, + 152, + 1028 + ] + }, + { + "test_f1": 0.9413423289104685, + "test_acc": 0.9413836477987422, + "test_auc": 0.9845487114939259, + "val_auc": 0.9807923169267707, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 771, + 1524, + 7368, + 6475, + 559, + 152, + 1028, + 878, + 3417, + 24947, + 1288, + 6797, + 6935 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a109fe22f113770bdeb9ca8164353307729ddfbd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8629332075500492, + "test_acc": 0.8643037974683544, + "test_auc": 0.8700546374118602, + "val_auc": 0.8291316526610645, + "k": 1, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104 + ] + }, + { + "test_f1": 0.8662615937094683, + "test_acc": 0.8673417721518988, + "test_auc": 0.8984613743852266, + "val_auc": 0.8739495798319328, + "k": 2, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292 + ] + }, + { + "test_f1": 0.8656113139156743, + "test_acc": 0.8668354430379747, + "test_auc": 0.9367721677352137, + "val_auc": 0.9019607843137254, + "k": 5, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 337, + 1708, + 4555 + ] + }, + { + "test_f1": 0.864180536804401, + "test_acc": 0.8643037974683544, + "test_auc": 0.9409766232274632, + "val_auc": 0.9199679871948779, + "k": 10, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 337, + 1708, + 4555, + 1815, + 968, + 918, + 1851, + 1745 + ] + }, + { + "test_f1": 0.8621448203464447, + "test_acc": 0.8622784810126582, + "test_auc": 0.9427096792718292, + "val_auc": 0.9315726290516206, + "k": 16, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 337, + 1708, + 4555, + 1815, + 968, + 918, + 1851, + 1745, + 1438, + 3385, + 1380, + 1627, + 677, + 6747 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6e5061eb4ccd0ed2b5a3b0829632344e62b01417 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.746306171763726, + "test_acc": 0.7497435897435898, + "test_auc": 0.7778490759753593, + "val_auc": 0.7384740922072623, + "k": 1, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1804 + ] + }, + { + "test_f1": 0.7783747205269684, + "test_acc": 0.7815384615384615, + "test_auc": 0.8518068132090081, + "val_auc": 0.8223174214606284, + "k": 2, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1804, + 468 + ] + }, + { + "test_f1": 0.8577799776386027, + "test_acc": 0.8584615384615385, + "test_auc": 0.8928493284411081, + "val_auc": 0.8849449204406364, + "k": 5, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1804, + 468, + 1351, + 1474, + 26988 + ] + }, + { + "test_f1": 0.8629131725046423, + "test_acc": 0.8635897435897436, + "test_auc": 0.9088219948160368, + "val_auc": 0.8808649530803753, + "k": 10, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1804, + 468, + 1351, + 1474, + 26988, + 1132, + 343, + 3417, + 805, + 1305 + ] + }, + { + "test_f1": 0.86205042874373, + "test_acc": 0.8625641025641025, + "test_auc": 0.9222237183155484, + "val_auc": 0.8977968176254589, + "k": 16, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1804, + 468, + 1351, + 1474, + 26988, + 1132, + 343, + 3417, + 805, + 1305, + 930, + 7595, + 1473, + 21810, + 1551, + 6475 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8fe1182629aad91af959168748aa392ed9a148e3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8995124547916965, + "test_acc": 0.899874213836478, + "test_auc": 0.911359703262352, + "val_auc": 0.969187675070028, + "k": 1, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.8914228249105614, + "test_acc": 0.8915723270440251, + "test_auc": 0.9500542510219849, + "val_auc": 0.9823929571828731, + "k": 2, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279 + ] + }, + { + "test_f1": 0.9178934174585688, + "test_acc": 0.9179874213836478, + "test_auc": 0.9751536395018323, + "val_auc": 0.9875950380152061, + "k": 5, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279, + 1447, + 238, + 1556 + ] + }, + { + "test_f1": 0.9307840721430317, + "test_acc": 0.9308176100628931, + "test_auc": 0.9772965928434219, + "val_auc": 0.9843937575030012, + "k": 10, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279, + 1447, + 238, + 1556, + 27068, + 8129, + 2860, + 11381, + 6828 + ] + }, + { + "test_f1": 0.9383527629974818, + "test_acc": 0.9383647798742139, + "test_auc": 0.9830475049592978, + "val_auc": 0.992797118847539, + "k": 16, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279, + 1447, + 238, + 1556, + 27068, + 8129, + 2860, + 11381, + 6828, + 1478, + 1012, + 88, + 1101, + 32735, + 7740 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ab3a9d7eb7f9e909f87009595aa91770b1da9a9b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8461201369714404, + "test_acc": 0.8482051282051282, + "test_auc": 0.8520719022452621, + "val_auc": 0.890452876376989, + "k": 1, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628 + ] + }, + { + "test_f1": 0.8903582928496795, + "test_acc": 0.8912820512820513, + "test_auc": 0.9134105598007204, + "val_auc": 0.9051407588739291, + "k": 2, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416 + ] + }, + { + "test_f1": 0.8893028015988149, + "test_acc": 0.8902564102564102, + "test_auc": 0.9125669034234355, + "val_auc": 0.9145246838025296, + "k": 5, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 238 + ] + }, + { + "test_f1": 0.9009757920351851, + "test_acc": 0.9015384615384615, + "test_auc": 0.9148390951627562, + "val_auc": 0.9277845777233783, + "k": 10, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 238, + 737, + 4211, + 805, + 930, + 334 + ] + }, + { + "test_f1": 0.9011190938662798, + "test_acc": 0.9015384615384615, + "test_auc": 0.9302815834651764, + "val_auc": 0.9163606691146471, + "k": 16, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 238, + 737, + 4211, + 805, + 930, + 334, + 1132, + 1274, + 1524, + 2062, + 878, + 1288 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9a101c33e3a105782926549da3bc8146b0b16e71 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8666018371178635, + "test_acc": 0.8666666666666667, + "test_auc": 0.91641700609284, + "val_auc": 0.9394124847001224, + "k": 1, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524 + ] + }, + { + "test_f1": 0.8696160386508646, + "test_acc": 0.8697435897435898, + "test_auc": 0.9229895310869491, + "val_auc": 0.9398204814361485, + "k": 2, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947 + ] + }, + { + "test_f1": 0.8686134644051721, + "test_acc": 0.8687179487179487, + "test_auc": 0.9265703369576194, + "val_auc": 0.9396164830681355, + "k": 5, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 888 + ] + }, + { + "test_f1": 0.8716693293807722, + "test_acc": 0.8717948717948718, + "test_auc": 0.9360062443195206, + "val_auc": 0.949000407996736, + "k": 10, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 888, + 969, + 1474, + 3203, + 6631, + 862 + ] + }, + { + "test_f1": 0.8696473319444868, + "test_acc": 0.8697435897435898, + "test_auc": 0.937914464604302, + "val_auc": 0.9551203590371278, + "k": 16, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 888, + 969, + 1474, + 3203, + 6631, + 862, + 1214, + 1260, + 776, + 1682, + 805, + 3417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..16450aa3ae8b622a67a87bed992052b34c8b22be --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8044354871357218, + "test_acc": 0.8071794871794872, + "test_auc": 0.8314033729424042, + "val_auc": 0.8119135046919624, + "k": 1, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.8064197341120418, + "test_acc": 0.8092307692307692, + "test_auc": 0.8424340222843101, + "val_auc": 0.8164014687882497, + "k": 2, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 10639 + ] + }, + { + "test_f1": 0.8144418875636049, + "test_acc": 0.8153846153846154, + "test_auc": 0.8980816810852661, + "val_auc": 0.8745410036719706, + "k": 5, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 10639, + 1815, + 1524, + 334 + ] + }, + { + "test_f1": 0.8756145606230952, + "test_acc": 0.8758974358974358, + "test_auc": 0.9448846231527922, + "val_auc": 0.8884128926968585, + "k": 10, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 10639, + 1815, + 1524, + 334, + 1866, + 487, + 930, + 6935, + 913 + ] + }, + { + "test_f1": 0.8830089914800513, + "test_acc": 0.8830769230769231, + "test_auc": 0.9514045511158986, + "val_auc": 0.9220726234190126, + "k": 16, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 10639, + 1815, + 1524, + 334, + 1866, + 487, + 930, + 6935, + 913, + 1744, + 2046, + 1221, + 1158, + 1148, + 821 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..34cacf9ad912adee484dc90d5e2ec4129e057871 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8191546384810471, + "test_acc": 0.8205128205128205, + "test_auc": 0.8614720436260814, + "val_auc": 0.8616891064871481, + "k": 1, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52 + ] + }, + { + "test_f1": 0.8221575943609841, + "test_acc": 0.8235897435897436, + "test_auc": 0.8614720436260814, + "val_auc": 0.8616891064871481, + "k": 2, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230 + ] + }, + { + "test_f1": 0.8221575943609841, + "test_acc": 0.8235897435897436, + "test_auc": 0.8614720436260814, + "val_auc": 0.8616891064871481, + "k": 5, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 2131, + 18545, + 2454 + ] + }, + { + "test_f1": 0.8191739465772615, + "test_acc": 0.8215384615384616, + "test_auc": 0.8614720436260814, + "val_auc": 0.8616891064871481, + "k": 10, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 2131, + 18545, + 2454, + 1274, + 1193, + 2046, + 888, + 2884 + ] + }, + { + "test_f1": 0.8320392680663173, + "test_acc": 0.8328205128205128, + "test_auc": 0.8926347325546167, + "val_auc": 0.9281925744594044, + "k": 16, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 2131, + 18545, + 2454, + 1274, + 1193, + 2046, + 888, + 2884, + 17546, + 238, + 821, + 2484, + 1682, + 1851 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d5fc02f2d21ea8d94891fcdd6223b62e2908b592 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8735935028014741, + "test_acc": 0.8737967914438503, + "test_auc": 0.9198443419535496, + "val_auc": 0.9191919191919192, + "k": 1, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839 + ] + }, + { + "test_f1": 0.8746957519446293, + "test_acc": 0.8748663101604278, + "test_auc": 0.9355313969874997, + "val_auc": 0.9321212121212121, + "k": 2, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682 + ] + }, + { + "test_f1": 0.881157001692009, + "test_acc": 0.881283422459893, + "test_auc": 0.941415472464723, + "val_auc": 0.9365656565656566, + "k": 5, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 3417 + ] + }, + { + "test_f1": 0.9058681400678664, + "test_acc": 0.9058823529411765, + "test_auc": 0.9529868775050788, + "val_auc": 0.9503030303030303, + "k": 10, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 3417, + 930, + 340, + 343, + 1895, + 874 + ] + }, + { + "test_f1": 0.9101468609738725, + "test_acc": 0.9101604278074866, + "test_auc": 0.9588938304141731, + "val_auc": 0.9474747474747476, + "k": 16, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 3417, + 930, + 340, + 343, + 1895, + 874, + 238, + 1441, + 334, + 888, + 456, + 1744 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8be3d10ad1d8f5abec9fbb9dd56513eca0009fad --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9245601931534215, + "test_acc": 0.9245977011494253, + "test_auc": 0.9592958561069322, + "val_auc": 0.9874188311688312, + "k": 1, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554 + ] + }, + { + "test_f1": 0.9544531037439785, + "test_acc": 0.9544827586206897, + "test_auc": 0.9841078048054549, + "val_auc": 0.9971590909090909, + "k": 2, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 14949 + ] + }, + { + "test_f1": 0.9724128833712203, + "test_acc": 0.9724137931034482, + "test_auc": 0.9960385775745442, + "val_auc": 1.0, + "k": 5, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 14949, + 1072, + 760, + 1065 + ] + }, + { + "test_f1": 0.97747083575023, + "test_acc": 0.9774712643678161, + "test_auc": 0.9960478786730883, + "val_auc": 1.0, + "k": 10, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 14949, + 1072, + 760, + 1065, + 1101, + 1682, + 192, + 1622, + 1876 + ] + }, + { + "test_f1": 0.9802298599823935, + "test_acc": 0.9802298850574712, + "test_auc": 0.9952767330483252, + "val_auc": 1.0, + "k": 16, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 14949, + 1072, + 760, + 1065, + 1101, + 1682, + 192, + 1622, + 1876, + 763, + 1684, + 49, + 432, + 1872, + 7847 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f02f10d0562601271992e1a7e5d5585db7260f42 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9756330292151327, + "test_acc": 0.9756330625895844, + "test_auc": 0.9887637628040418, + "val_auc": 1.0, + "k": 1, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286 + ] + }, + { + "test_f1": 0.9827992905873214, + "test_acc": 0.9827998088867654, + "test_auc": 0.992454084418561, + "val_auc": 1.0, + "k": 2, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905 + ] + }, + { + "test_f1": 0.9847101035571257, + "test_acc": 0.9847109412326803, + "test_auc": 0.9939342307348137, + "val_auc": 1.0, + "k": 5, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 304, + 1874, + 2874 + ] + }, + { + "test_f1": 0.9880549974637796, + "test_acc": 0.9880554228380315, + "test_auc": 0.9961165562720401, + "val_auc": 1.0, + "k": 10, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 304, + 1874, + 2874, + 1233, + 581, + 1079, + 1684, + 3877 + ] + }, + { + "test_f1": 0.9885327295026292, + "test_acc": 0.9885332059245103, + "test_auc": 0.9975200016070682, + "val_auc": 1.0, + "k": 16, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 304, + 1874, + 2874, + 1233, + 581, + 1079, + 1684, + 3877, + 1682, + 6080, + 760, + 1156, + 239, + 1596 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0304268f1d11c0e567c377c7728498fae6a5503e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9908041894831376, + "test_acc": 0.9908045977011494, + "test_auc": 0.9978971061745765, + "val_auc": 1.0, + "k": 1, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559 + ] + }, + { + "test_f1": 0.9921836899446597, + "test_acc": 0.992183908045977, + "test_auc": 0.998857655987878, + "val_auc": 1.0, + "k": 2, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300 + ] + }, + { + "test_f1": 0.9931034045399655, + "test_acc": 0.993103448275862, + "test_auc": 0.9997598625466746, + "val_auc": 1.0, + "k": 5, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 334, + 1934 + ] + }, + { + "test_f1": 0.9926436781609196, + "test_acc": 0.9926436781609196, + "test_auc": 0.9994351696520374, + "val_auc": 1.0, + "k": 10, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 334, + 1934, + 3825, + 2236, + 1592, + 982, + 1872 + ] + }, + { + "test_f1": 0.9931034453601555, + "test_acc": 0.993103448275862, + "test_auc": 0.9995797594566804, + "val_auc": 1.0, + "k": 16, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 334, + 1934, + 3825, + 2236, + 1592, + 982, + 1872, + 1158, + 119, + 1410, + 913, + 1769, + 1682 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0f340d660833950eac4fb8ec82279b084ff59272 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9393103191693833, + "test_acc": 0.9393103448275862, + "test_auc": 0.9783030737594025, + "val_auc": 0.952922077922078, + "k": 1, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095 + ] + }, + { + "test_f1": 0.9595394601259536, + "test_acc": 0.9595402298850575, + "test_auc": 0.9872490394501867, + "val_auc": 0.9715909090909092, + "k": 2, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023 + ] + }, + { + "test_f1": 0.9668944103001095, + "test_acc": 0.9668965517241379, + "test_auc": 0.9885029966448402, + "val_auc": 0.971185064935065, + "k": 5, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 1301 + ] + }, + { + "test_f1": 0.981147751531588, + "test_acc": 0.9811494252873563, + "test_auc": 0.9956310203474215, + "val_auc": 0.9841720779220778, + "k": 10, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 1301, + 1682, + 494, + 211, + 1079, + 1701 + ] + }, + { + "test_f1": 0.9788485090351811, + "test_acc": 0.9788505747126437, + "test_auc": 0.9960537975539802, + "val_auc": 0.987012987012987, + "k": 16, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 1301, + 1682, + 494, + 211, + 1079, + 1701, + 292, + 2011, + 1193, + 235, + 586, + 888 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3b33c809e1d0eed4a948edb82d1660bc7322c753 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9682711537000532, + "test_acc": 0.9682758620689655, + "test_auc": 0.9933623978570268, + "val_auc": 0.9878246753246753, + "k": 1, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32 + ] + }, + { + "test_f1": 0.9825286248348366, + "test_acc": 0.9825287356321839, + "test_auc": 0.9973415769251583, + "val_auc": 0.9943181818181819, + "k": 2, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683 + ] + }, + { + "test_f1": 0.9889654192727017, + "test_acc": 0.9889655172413793, + "test_auc": 0.9990123924454786, + "val_auc": 0.9967532467532467, + "k": 5, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 5094, + 816 + ] + }, + { + "test_f1": 0.9889654192727017, + "test_acc": 0.9889655172413793, + "test_auc": 0.9990123924454787, + "val_auc": 0.9967532467532467, + "k": 10, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 5094, + 816, + 1047, + 1558, + 1458, + 232, + 1627 + ] + }, + { + "test_f1": 0.993103419118721, + "test_acc": 0.993103448275862, + "test_auc": 0.9994867484712378, + "val_auc": 0.9967532467532468, + "k": 16, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 5094, + 816, + 1047, + 1558, + 1458, + 232, + 1627, + 495, + 868, + 1023, + 87, + 7481, + 1682 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..77137e1b36b79941e3bf7e3d18900f1320d9ca29 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8214457098955654, + "test_acc": 0.8216352201257862, + "test_auc": 0.870667892609811, + "val_auc": 0.9191676670668267, + "k": 1, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.8932651063285781, + "test_acc": 0.8933333333333333, + "test_auc": 0.9504991701593558, + "val_auc": 0.990796318527411, + "k": 2, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028 + ] + }, + { + "test_f1": 0.9149628701778163, + "test_acc": 0.9149685534591195, + "test_auc": 0.9655443987528595, + "val_auc": 0.9955982392957183, + "k": 5, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 778, + 1839, + 1305 + ] + }, + { + "test_f1": 0.9219989733601995, + "test_acc": 0.9220125786163522, + "test_auc": 0.9680031624067506, + "val_auc": 0.995998399359744, + "k": 10, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 778, + 1839, + 1305, + 3417, + 1478, + 1602, + 5539, + 4663 + ] + }, + { + "test_f1": 0.9222511014116447, + "test_acc": 0.9222641509433962, + "test_auc": 0.9704006626573737, + "val_auc": 0.9951980792316927, + "k": 16, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 778, + 1839, + 1305, + 3417, + 1478, + 1602, + 5539, + 4663, + 1895, + 654, + 30649, + 343, + 8038, + 1065 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6a3382448ee9897993917d10a05e729ec1845cda --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9990804593813543, + "test_acc": 0.999080459770115, + "test_auc": 0.9999890077926296, + "val_auc": 1.0, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999813978029115, + "val_auc": 1.0, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379 + ] + }, + { + "test_f1": 0.999080459770115, + "test_acc": 0.999080459770115, + "test_auc": 0.9990758090264624, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379, + 682, + 4787, + 547 + ] + }, + { + "test_f1": 0.999080459770115, + "test_acc": 0.999080459770115, + "test_auc": 0.9990783456897018, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379, + 682, + 4787, + 547, + 1840, + 6777, + 7651, + 24163, + 1162 + ] + }, + { + "test_f1": 0.9977011465095805, + "test_acc": 0.9977011494252873, + "test_auc": 0.9994875940256507, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379, + 682, + 4787, + 547, + 1840, + 6777, + 7651, + 24163, + 1162, + 84, + 990, + 1931, + 3910, + 613, + 1589 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d850fc2745bc312eb156a9e225a4735f8ebe2902 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.98252669674377, + "test_acc": 0.9825287356321839, + "test_auc": 0.9936963918502084, + "val_auc": 1.0, + "k": 1, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211 + ] + }, + { + "test_f1": 0.9852864045778265, + "test_acc": 0.9852873563218391, + "test_auc": 0.9946324205855295, + "val_auc": 1.0, + "k": 2, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586 + ] + }, + { + "test_f1": 0.9940228975345435, + "test_acc": 0.9940229885057471, + "test_auc": 0.9970363317820228, + "val_auc": 1.0, + "k": 5, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 21724, + 10, + 559 + ] + }, + { + "test_f1": 0.9935631421929442, + "test_acc": 0.9935632183908046, + "test_auc": 0.9972688592456301, + "val_auc": 1.0, + "k": 10, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 21724, + 10, + 559, + 1934, + 3223, + 878, + 620, + 495 + ] + }, + { + "test_f1": 0.9921836502898747, + "test_acc": 0.992183908045977, + "test_auc": 0.9967902754478055, + "val_auc": 1.0, + "k": 16, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 21724, + 10, + 559, + 1934, + 3223, + 878, + 620, + 495, + 1682, + 1874, + 1603, + 760, + 1872, + 1023 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f0e36d2ff8e19a7cd46778e5abb4240e807b6bd7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8941626676303573, + "test_acc": 0.8942528735632184, + "test_auc": 0.947365928892256, + "val_auc": 0.9577922077922078, + "k": 1, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979 + ] + }, + { + "test_f1": 0.9204461149695229, + "test_acc": 0.9204597701149425, + "test_auc": 0.9639523242599708, + "val_auc": 0.9898538961038961, + "k": 2, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 495 + ] + }, + { + "test_f1": 0.9374639672934927, + "test_acc": 0.9374712643678161, + "test_auc": 0.9808608758590832, + "val_auc": 0.9959415584415584, + "k": 5, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 495, + 1682, + 1027, + 1708 + ] + }, + { + "test_f1": 0.950793613615002, + "test_acc": 0.9508045977011494, + "test_auc": 0.9908874600898318, + "val_auc": 1.0, + "k": 10, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 495, + 1682, + 1027, + 1708, + 1631, + 17284, + 117, + 1707, + 425 + ] + }, + { + "test_f1": 0.959070163846223, + "test_acc": 0.9590804597701149, + "test_auc": 0.9933514056496564, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 495, + 1682, + 1027, + 1708, + 1631, + 17284, + 117, + 1707, + 425, + 713, + 1840, + 1841, + 2024, + 1079, + 87 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c4f583fcc289ecdb9143a14f947cf280cd788c01 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9627508095765662, + "test_acc": 0.9627586206896551, + "test_auc": 0.9765827933059149, + "val_auc": 0.9886363636363636, + "k": 1, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734 + ] + }, + { + "test_f1": 0.9650506004693761, + "test_acc": 0.9650574712643678, + "test_auc": 0.98399661440013, + "val_auc": 0.9983766233766234, + "k": 2, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734, + 1961 + ] + }, + { + "test_f1": 0.9655125026827402, + "test_acc": 0.9655172413793104, + "test_auc": 0.9863442962281509, + "val_auc": 0.9979707792207793, + "k": 5, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734, + 1961, + 7182, + 248, + 2689 + ] + }, + { + "test_f1": 0.9701134262804696, + "test_acc": 0.9701149425287356, + "test_auc": 0.987442248633584, + "val_auc": 0.9987824675324676, + "k": 10, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734, + 1961, + 7182, + 248, + 2689, + 732, + 1264, + 1536, + 1441, + 859 + ] + }, + { + "test_f1": 0.9714906369225534, + "test_acc": 0.9714942528735632, + "test_auc": 0.9905357094539748, + "val_auc": 1.0, + "k": 16, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8734, + 1961, + 7182, + 248, + 2689, + 732, + 1264, + 1536, + 1441, + 859, + 653, + 1682, + 1482, + 1960, + 1366, + 4861 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8a0e18d10e8bb2650e574b4352a13625cfa1d1da --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9558515980105186, + "test_acc": 0.9558620689655173, + "test_auc": 0.9807881581795551, + "val_auc": 0.9817370129870129, + "k": 1, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786 + ] + }, + { + "test_f1": 0.9581556123122761, + "test_acc": 0.9581609195402299, + "test_auc": 0.9882552492017966, + "val_auc": 0.984577922077922, + "k": 2, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558 + ] + }, + { + "test_f1": 0.9609153596541594, + "test_acc": 0.960919540229885, + "test_auc": 0.9883787001461118, + "val_auc": 0.9862012987012987, + "k": 5, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 271 + ] + }, + { + "test_f1": 0.9682687788732253, + "test_acc": 0.9682758620689655, + "test_auc": 0.9932744601980626, + "val_auc": 0.9898538961038961, + "k": 10, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 271, + 653, + 1627, + 1910, + 292, + 760 + ] + }, + { + "test_f1": 0.9668844974181595, + "test_acc": 0.9668965517241379, + "test_auc": 0.9919985185886683, + "val_auc": 0.9890422077922078, + "k": 16, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 271, + 653, + 1627, + 1910, + 292, + 760, + 276, + 327, + 1051, + 460, + 1649, + 67 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..798be0ae3cc32bc4cee30ef155dcd9b6fed235de --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9282738606783723, + "test_acc": 0.9282758620689655, + "test_auc": 0.9652532097245522, + "val_auc": 0.9598214285714286, + "k": 1, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856 + ] + }, + { + "test_f1": 0.971494072098944, + "test_acc": 0.9714942528735632, + "test_auc": 0.987877286379133, + "val_auc": 0.991883116883117, + "k": 2, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856, + 11831 + ] + }, + { + "test_f1": 0.9737927045745373, + "test_acc": 0.9737931034482759, + "test_auc": 0.990961868878186, + "val_auc": 0.994724025974026, + "k": 5, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856, + 11831, + 1168, + 1167, + 7739 + ] + }, + { + "test_f1": 0.9774673967254718, + "test_acc": 0.9774712643678161, + "test_auc": 0.9949926267655176, + "val_auc": 0.9967532467532467, + "k": 10, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856, + 11831, + 1168, + 1167, + 7739, + 1037, + 822, + 517, + 18934, + 1252 + ] + }, + { + "test_f1": 0.9742508051991596, + "test_acc": 0.9742528735632184, + "test_auc": 0.9946062083987229, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1856, + 11831, + 1168, + 1167, + 7739, + 1037, + 822, + 517, + 18934, + 1252, + 1536, + 586, + 1449, + 581, + 1101, + 779 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f8379b595c4ff73d7a9e0aa6fb5038e229d2de2c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9581587967926576, + "test_acc": 0.9581609195402299, + "test_auc": 0.9741260349586016, + "val_auc": 0.9859983766233766, + "k": 1, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840 + ] + }, + { + "test_f1": 0.9586206896551724, + "test_acc": 0.9586206896551724, + "test_auc": 0.9939551315006223, + "val_auc": 0.9987824675324676, + "k": 2, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495 + ] + }, + { + "test_f1": 0.9751714585742641, + "test_acc": 0.9751724137931035, + "test_auc": 0.9967556077168678, + "val_auc": 0.9991883116883117, + "k": 5, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495, + 1765, + 789, + 1763 + ] + }, + { + "test_f1": 0.9802290073941264, + "test_acc": 0.9802298850574712, + "test_auc": 0.9967750554683695, + "val_auc": 0.9991883116883117, + "k": 10, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495, + 1765, + 789, + 1763, + 1682, + 901, + 1797, + 434, + 1310 + ] + }, + { + "test_f1": 0.9843677169574855, + "test_acc": 0.984367816091954, + "test_auc": 0.9972189715352562, + "val_auc": 1.0, + "k": 16, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495, + 1765, + 789, + 1763, + 1682, + 901, + 1797, + 434, + 1310, + 2120, + 27505, + 1815, + 10075, + 677, + 1684 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cdecf73a4a662fd192aa13821aa5c33017bdd061 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9567713563852994, + "test_acc": 0.9567816091954023, + "test_auc": 0.9681796735483522, + "val_auc": 0.971185064935065, + "k": 1, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879 + ] + }, + { + "test_f1": 0.9645965336659029, + "test_acc": 0.9645977011494253, + "test_auc": 0.9870600580388549, + "val_auc": 0.9833603896103895, + "k": 2, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293 + ] + }, + { + "test_f1": 0.9875854038565993, + "test_acc": 0.9875862068965517, + "test_auc": 0.9958618567022024, + "val_auc": 1.0, + "k": 5, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 277, + 4455, + 31202 + ] + }, + { + "test_f1": 0.9894252605317521, + "test_acc": 0.9894252873563218, + "test_auc": 0.9966397667622708, + "val_auc": 1.0, + "k": 10, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 277, + 4455, + 31202, + 1283, + 990, + 7367, + 1051, + 1501 + ] + }, + { + "test_f1": 0.9852872941198196, + "test_acc": 0.9852873563218391, + "test_auc": 0.9941648289950754, + "val_auc": 1.0, + "k": 16, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 277, + 4455, + 31202, + 1283, + 990, + 7367, + 1051, + 1501, + 91, + 30210, + 30, + 1867, + 1484, + 1404 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6879f74862b02a51516746260697f41eda52aa1e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9705745882396393, + "test_acc": 0.9705747126436781, + "test_auc": 0.9799032854862275, + "val_auc": 1.0, + "k": 1, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439 + ] + }, + { + "test_f1": 0.9737930369699938, + "test_acc": 0.9737931034482759, + "test_auc": 0.984942367011202, + "val_auc": 1.0, + "k": 2, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 822 + ] + }, + { + "test_f1": 0.9733332995110996, + "test_acc": 0.9733333333333334, + "test_auc": 0.9858331585854212, + "val_auc": 1.0, + "k": 5, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 822, + 1795, + 26232, + 1121 + ] + }, + { + "test_f1": 0.9751722563442415, + "test_acc": 0.9751724137931035, + "test_auc": 0.9812963363818389, + "val_auc": 1.0, + "k": 10, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 822, + 1795, + 26232, + 1121, + 5535, + 493, + 5797, + 1790, + 802 + ] + }, + { + "test_f1": 0.9797699866508635, + "test_acc": 0.9797701149425287, + "test_auc": 0.9936887818604903, + "val_auc": 1.0, + "k": 16, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 822, + 1795, + 26232, + 1121, + 5535, + 493, + 5797, + 1790, + 802, + 1441, + 1631, + 345, + 2020, + 5397, + 495 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2dc97eab6f826acaad707341935d59814e6ae0c1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9544692262053773, + "test_acc": 0.9544827586206897, + "test_auc": 0.9830703095405596, + "val_auc": 0.9839691558441559, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272 + ] + }, + { + "test_f1": 0.9553844548243395, + "test_acc": 0.9554022988505747, + "test_auc": 0.9870110158828941, + "val_auc": 0.9906655844155844, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866 + ] + }, + { + "test_f1": 0.9586102778948411, + "test_acc": 0.9586206896551724, + "test_auc": 0.9928660574165269, + "val_auc": 0.9951298701298702, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 1072, + 5703 + ] + }, + { + "test_f1": 0.978390201611706, + "test_acc": 0.9783908045977011, + "test_auc": 0.9958483278315926, + "val_auc": 0.9967532467532467, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 1072, + 5703, + 165, + 918, + 1682, + 1787, + 1726 + ] + }, + { + "test_f1": 0.9839076717870553, + "test_acc": 0.9839080459770115, + "test_auc": 0.9975123789166082, + "val_auc": 0.9991883116883117, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 1072, + 5703, + 165, + 918, + 1682, + 1787, + 1726, + 1079, + 508, + 1181, + 1748, + 495, + 410 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ed6700b34c18f7082da5242b88d3532333227337 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5639757822652414, + "test_acc": 0.5645283018867925, + "test_auc": 0.5721374041936571, + "val_auc": 0.5854341736694677, + "k": 1, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556 + ] + }, + { + "test_f1": 0.5793531582263137, + "test_acc": 0.5811320754716981, + "test_auc": 0.6222532730352928, + "val_auc": 0.6306522609043617, + "k": 2, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556, + 776 + ] + }, + { + "test_f1": 0.6062179813780186, + "test_acc": 0.6062893081761006, + "test_auc": 0.6582143844445637, + "val_auc": 0.6790716286514606, + "k": 5, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556, + 776, + 1279, + 1132, + 6475 + ] + }, + { + "test_f1": 0.6146385814117148, + "test_acc": 0.6148427672955975, + "test_auc": 0.6597798922371674, + "val_auc": 0.6730692276910765, + "k": 10, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556, + 776, + 1279, + 1132, + 6475, + 1815, + 17275, + 238, + 1895, + 581 + ] + }, + { + "test_f1": 0.621743011894438, + "test_acc": 0.6218867924528302, + "test_auc": 0.6741340088847123, + "val_auc": 0.6850740296118447, + "k": 16, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556, + 776, + 1279, + 1132, + 6475, + 1815, + 17275, + 238, + 1895, + 581, + 7893, + 925, + 6758, + 7740, + 2126, + 1158 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fc6c908a74b8dd8cc493612c3f1aa4d85fb0ca95 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8928735632183908, + "test_acc": 0.8928735632183908, + "test_auc": 0.9462667081552032, + "val_auc": 0.9821428571428571, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217 + ] + }, + { + "test_f1": 0.9678060062059142, + "test_acc": 0.967816091954023, + "test_auc": 0.9798259172574273, + "val_auc": 0.9991883116883117, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 10963 + ] + }, + { + "test_f1": 0.970571564886127, + "test_acc": 0.9705747126436781, + "test_auc": 0.992974288381406, + "val_auc": 0.9987824675324676, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 10963, + 3772, + 1942, + 352 + ] + }, + { + "test_f1": 0.9788505657711487, + "test_acc": 0.9788505747126437, + "test_auc": 0.994132697927377, + "val_auc": 0.9983766233766234, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 10963, + 3772, + 1942, + 352, + 1682, + 1261, + 3243, + 111, + 1047 + ] + }, + { + "test_f1": 0.9774712357938903, + "test_acc": 0.9774712643678161, + "test_auc": 0.9943516965203745, + "val_auc": 0.9991883116883117, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 10963, + 3772, + 1942, + 352, + 1682, + 1261, + 3243, + 111, + 1047, + 1898, + 760, + 581, + 1966, + 1726, + 285 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6b5a35764c307b923b7f04072658190fa26a98de --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9862068790574652, + "test_acc": 0.9862068965517241, + "test_auc": 0.9986145590941068, + "val_auc": 1.0, + "k": 1, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833 + ] + }, + { + "test_f1": 0.9871264204536344, + "test_acc": 0.9871264367816092, + "test_auc": 0.9986822034471562, + "val_auc": 1.0, + "k": 2, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675 + ] + }, + { + "test_f1": 0.9894252873563218, + "test_acc": 0.9894252873563218, + "test_auc": 0.9990191568807836, + "val_auc": 1.0, + "k": 5, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682 + ] + }, + { + "test_f1": 0.9889655172413793, + "test_acc": 0.9889655172413793, + "test_auc": 0.9994258685534931, + "val_auc": 1.0, + "k": 10, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682, + 1588, + 913, + 829, + 1281, + 1278 + ] + }, + { + "test_f1": 0.9940229809249097, + "test_acc": 0.9940229885057471, + "test_auc": 0.9993912008225553, + "val_auc": 1.0, + "k": 16, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682, + 1588, + 913, + 829, + 1281, + 1278, + 491, + 361, + 1181, + 859, + 1079, + 1515 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..64eaeaa695611a1fdaa9794408a1c12c4e4295c5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8574032617158527, + "test_acc": 0.8574712643678161, + "test_auc": 0.8979923156014936, + "val_auc": 0.9348620129870129, + "k": 1, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020 + ] + }, + { + "test_f1": 0.910800072312905, + "test_acc": 0.9108045977011494, + "test_auc": 0.9582727352670599, + "val_auc": 0.9797077922077921, + "k": 2, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760 + ] + }, + { + "test_f1": 0.9429458117463396, + "test_acc": 0.9429885057471264, + "test_auc": 0.9829916729801396, + "val_auc": 0.9841720779220778, + "k": 5, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 913 + ] + }, + { + "test_f1": 0.958147808281168, + "test_acc": 0.9581609195402299, + "test_auc": 0.990523026137778, + "val_auc": 0.9931006493506493, + "k": 10, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 913, + 49, + 170, + 18571, + 1960, + 1120 + ] + }, + { + "test_f1": 0.9609072934494426, + "test_acc": 0.960919540229885, + "test_auc": 0.9931433992640295, + "val_auc": 0.9951298701298701, + "k": 16, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 913, + 49, + 170, + 18571, + 1960, + 1120, + 2817, + 364, + 901, + 2023, + 3842, + 2030 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b23ae40ba89e74ebbaa786cd9635595c422ac186 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.941124782869319, + "test_acc": 0.9411494252873563, + "test_auc": 0.9539891566102061, + "val_auc": 0.9435876623376624, + "k": 1, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123 + ] + }, + { + "test_f1": 0.9420560308081254, + "test_acc": 0.9420689655172414, + "test_auc": 0.9662573055901293, + "val_auc": 0.9606331168831169, + "k": 2, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 866 + ] + }, + { + "test_f1": 0.9655128672903374, + "test_acc": 0.9655172413793104, + "test_auc": 0.9838329996211916, + "val_auc": 0.9788961038961039, + "k": 5, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 866, + 1763, + 1491, + 1682 + ] + }, + { + "test_f1": 0.9659698755106825, + "test_acc": 0.9659770114942529, + "test_auc": 0.991119141999026, + "val_auc": 0.9821428571428571, + "k": 10, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 866, + 1763, + 1491, + 1682, + 1111, + 204, + 4322, + 679, + 167 + ] + }, + { + "test_f1": 0.9696503222670553, + "test_acc": 0.9696551724137931, + "test_auc": 0.9938054683695005, + "val_auc": 0.9906655844155844, + "k": 16, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 866, + 1763, + 1491, + 1682, + 1111, + 204, + 4322, + 679, + 167, + 7316, + 1486, + 523, + 687, + 144, + 1121 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b72249178b4786a3ce33f1f4590bc0567f359662 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8457098983977924, + "test_acc": 0.8459770114942529, + "test_auc": 0.8889135978678501, + "val_auc": 0.885551948051948, + "k": 1, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967 + ] + }, + { + "test_f1": 0.8745638932388586, + "test_acc": 0.8749425287356322, + "test_auc": 0.9214040261918935, + "val_auc": 0.9237012987012987, + "k": 2, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352 + ] + }, + { + "test_f1": 0.9356169137653396, + "test_acc": 0.9356321839080459, + "test_auc": 0.9744371989826288, + "val_auc": 0.9833603896103896, + "k": 5, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 1682, + 495, + 12346 + ] + }, + { + "test_f1": 0.949860128887772, + "test_acc": 0.9498850574712644, + "test_auc": 0.9820936941934088, + "val_auc": 0.9862012987012987, + "k": 10, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 1682, + 495, + 12346, + 19873, + 178, + 859, + 108, + 1168 + ] + }, + { + "test_f1": 0.9475568074308742, + "test_acc": 0.9475862068965517, + "test_auc": 0.9827760566047947, + "val_auc": 0.994724025974026, + "k": 16, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 1682, + 495, + 12346, + 19873, + 178, + 859, + 108, + 1168, + 1762, + 1037, + 960, + 1797, + 987, + 507 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ae24045c2a310b145d57a544410e65fc5732f348 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8940653999394145, + "test_acc": 0.8942528735632184, + "test_auc": 0.9445548832458466, + "val_auc": 0.9618506493506493, + "k": 1, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335 + ] + }, + { + "test_f1": 0.924445881310453, + "test_acc": 0.9245977011494253, + "test_auc": 0.9630746387791547, + "val_auc": 0.9756493506493505, + "k": 2, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443 + ] + }, + { + "test_f1": 0.9618363214314368, + "test_acc": 0.9618390804597701, + "test_auc": 0.9923460414524595, + "val_auc": 0.9983766233766234, + "k": 5, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1079, + 1281 + ] + }, + { + "test_f1": 0.9641377036083489, + "test_acc": 0.9641379310344828, + "test_auc": 0.9939686603712321, + "val_auc": 0.9987824675324676, + "k": 10, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1079, + 1281, + 1515, + 21, + 760, + 165, + 482 + ] + }, + { + "test_f1": 0.9687350373700552, + "test_acc": 0.968735632183908, + "test_auc": 0.9942282455760594, + "val_auc": 1.0, + "k": 16, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1079, + 1281, + 1515, + 21, + 760, + 165, + 482, + 559, + 1928, + 878, + 5792, + 1631, + 539 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a30c69c2060a58bde6951db3783df56cc5d34498 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6870339840139198, + "test_acc": 0.6920754716981132, + "test_auc": 0.7080314043293479, + "val_auc": 0.7348939575830331, + "k": 1, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313 + ] + }, + { + "test_f1": 0.7069012964306998, + "test_acc": 0.7094339622641509, + "test_auc": 0.7431593840850843, + "val_auc": 0.7446978791516607, + "k": 2, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 433 + ] + }, + { + "test_f1": 0.7070026536112534, + "test_acc": 0.7101886792452831, + "test_auc": 0.7577366564763519, + "val_auc": 0.7498999599839935, + "k": 5, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 433, + 25757, + 24630, + 19000 + ] + }, + { + "test_f1": 0.7046034548716898, + "test_acc": 0.7071698113207547, + "test_auc": 0.7804360131599866, + "val_auc": 0.765906362545018, + "k": 10, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 433, + 25757, + 24630, + 19000, + 1105, + 4333, + 1532, + 121, + 1928 + ] + }, + { + "test_f1": 0.7255064413526926, + "test_acc": 0.7270440251572327, + "test_auc": 0.7924794615706316, + "val_auc": 0.792717086834734, + "k": 16, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 433, + 25757, + 24630, + 19000, + 1105, + 4333, + 1532, + 121, + 1928, + 6636, + 1067, + 6718, + 1019, + 26569, + 2424 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f7eb40b8f572a00768e0f3c730342171fd68543b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6973505634174435, + "test_acc": 0.698459280997799, + "test_auc": 0.7556584029868101, + "val_auc": 0.7958733974358974, + "k": 1, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394 + ] + }, + { + "test_f1": 0.6870466814876092, + "test_acc": 0.6889214966984593, + "test_auc": 0.7556584029868101, + "val_auc": 0.7958733974358974, + "k": 2, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041 + ] + }, + { + "test_f1": 0.6870466814876092, + "test_acc": 0.6889214966984593, + "test_auc": 0.7556584029868101, + "val_auc": 0.7958733974358974, + "k": 5, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584 + ] + }, + { + "test_f1": 0.6993728962900126, + "test_acc": 0.6999266324284666, + "test_auc": 0.7609841487203999, + "val_auc": 0.8229166666666667, + "k": 10, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584, + 1737, + 1551, + 465, + 1257, + 28053 + ] + }, + { + "test_f1": 0.6935673495906371, + "test_acc": 0.6940572267057961, + "test_auc": 0.7672045163874068, + "val_auc": 0.8317307692307693, + "k": 16, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584, + 1737, + 1551, + 465, + 1257, + 28053, + 7649, + 1697, + 24358, + 31, + 28397, + 2367 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..31d982995f90e8a6a42ac62f86e57eeac8391f39 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6557187102111091, + "test_acc": 0.6581132075471698, + "test_auc": 0.6847377673185566, + "val_auc": 0.6694677871148459, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927 + ] + }, + { + "test_f1": 0.6971513961315426, + "test_acc": 0.6978616352201258, + "test_auc": 0.7691310165978255, + "val_auc": 0.7759103641456582, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927, + 24630 + ] + }, + { + "test_f1": 0.7325776685632281, + "test_acc": 0.7325786163522012, + "test_auc": 0.8103469837646918, + "val_auc": 0.7651060424169668, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927, + 24630, + 1011, + 17636, + 1520 + ] + }, + { + "test_f1": 0.7504381667092875, + "test_acc": 0.7504402515723271, + "test_auc": 0.8264405253868454, + "val_auc": 0.7903161264505802, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927, + 24630, + 1011, + 17636, + 1520, + 1848, + 1036, + 6417, + 5653, + 4520 + ] + }, + { + "test_f1": 0.7682992179372842, + "test_acc": 0.7683018867924528, + "test_auc": 0.845205986801534, + "val_auc": 0.8107242897158863, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927, + 24630, + 1011, + 17636, + 1520, + 1848, + 1036, + 6417, + 5653, + 4520, + 7614, + 3801, + 1928, + 4674, + 1500, + 1470 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f6bd4c51ba3f79431f5745b4d38ea81c49758b2e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7335291511658554, + "test_acc": 0.7335849056603774, + "test_auc": 0.7941172449898182, + "val_auc": 0.7785114045618247, + "k": 1, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503 + ] + }, + { + "test_f1": 0.7426160977870231, + "test_acc": 0.7426415094339622, + "test_auc": 0.8037105370015767, + "val_auc": 0.7955182072829132, + "k": 2, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472 + ] + }, + { + "test_f1": 0.7672534394726308, + "test_acc": 0.7672955974842768, + "test_auc": 0.8417090363013511, + "val_auc": 0.8251300520208084, + "k": 5, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472, + 1643, + 1928, + 1848 + ] + }, + { + "test_f1": 0.7761145784163014, + "test_acc": 0.7763522012578616, + "test_auc": 0.8508215878056462, + "val_auc": 0.8663465386154462, + "k": 10, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472, + 1643, + 1928, + 1848, + 7708, + 17225, + 7303, + 5891, + 7032 + ] + }, + { + "test_f1": 0.7761632211034578, + "test_acc": 0.7763522012578616, + "test_auc": 0.8542943620454484, + "val_auc": 0.8579431772709084, + "k": 16, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472, + 1643, + 1928, + 1848, + 7708, + 17225, + 7303, + 5891, + 7032, + 18220, + 446, + 6327, + 2471, + 2, + 7615 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7f8ae81a424e5e3e8414ac21c1f91599ae729a1e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8322560709795449, + "test_acc": 0.8327044025157233, + "test_auc": 0.8585977363931957, + "val_auc": 0.8473389355742297, + "k": 1, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8514888649912181, + "test_acc": 0.8525786163522012, + "test_auc": 0.9183038593918824, + "val_auc": 0.9293717486994798, + "k": 2, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8643002643629352, + "test_acc": 0.8649056603773585, + "test_auc": 0.9303027526001505, + "val_auc": 0.9349739895958382, + "k": 5, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 100, + 1438, + 4675 + ] + }, + { + "test_f1": 0.8683238791690037, + "test_acc": 0.8686792452830189, + "test_auc": 0.9371632411479445, + "val_auc": 0.9679871948779512, + "k": 10, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 100, + 1438, + 4675, + 7284, + 1132, + 777, + 1874, + 1928 + ] + }, + { + "test_f1": 0.8753552457532927, + "test_acc": 0.8754716981132076, + "test_auc": 0.9431113606652496, + "val_auc": 0.9623849539815926, + "k": 16, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 100, + 1438, + 4675, + 7284, + 1132, + 777, + 1874, + 1928, + 794, + 6745, + 1335, + 505, + 1960, + 22141 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..236152ac2f175cb9a77e37c4e6e4eda465628034 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8982430450367329, + "test_acc": 0.8988679245283019, + "test_auc": 0.9073320142293114, + "val_auc": 0.8999599839935974, + "k": 1, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480 + ] + }, + { + "test_f1": 0.8999705103380158, + "test_acc": 0.9006289308176101, + "test_auc": 0.9546559680174657, + "val_auc": 0.9671868747499001, + "k": 2, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433 + ] + }, + { + "test_f1": 0.9210765969157784, + "test_acc": 0.9212578616352202, + "test_auc": 0.9697624600142374, + "val_auc": 0.9831932773109243, + "k": 5, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 7666 + ] + }, + { + "test_f1": 0.9250134980195005, + "test_acc": 0.9250314465408805, + "test_auc": 0.9805101368148499, + "val_auc": 0.9931972789115646, + "k": 10, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 7666, + 3481, + 1064, + 6553, + 1335, + 1928 + ] + }, + { + "test_f1": 0.9302681002422619, + "test_acc": 0.930314465408805, + "test_auc": 0.982345254212745, + "val_auc": 0.995998399359744, + "k": 16, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 7666, + 3481, + 1064, + 6553, + 1335, + 1928, + 988, + 1500, + 1927, + 4110, + 15841, + 993 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..289c3792e173b4648a107251646b1f22b80f0e98 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8136249715197084, + "test_acc": 0.8181818181818182, + "test_auc": 0.8142857142857144, + "val_auc": 0.7413862179487178, + "k": 1, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629 + ] + }, + { + "test_f1": 0.7676767676767676, + "test_acc": 0.7676767676767676, + "test_auc": 0.8412244897959184, + "val_auc": 0.8108974358974359, + "k": 2, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625 + ] + }, + { + "test_f1": 0.7474747474747475, + "test_acc": 0.7474747474747475, + "test_auc": 0.8520408163265307, + "val_auc": 0.8465544871794871, + "k": 5, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 530, + 794 + ] + }, + { + "test_f1": 0.8382848088730442, + "test_acc": 0.8383838383838383, + "test_auc": 0.8879591836734694, + "val_auc": 0.8826121794871794, + "k": 10, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 530, + 794, + 2046, + 777, + 1345, + 765, + 1182 + ] + }, + { + "test_f1": 0.8382848088730442, + "test_acc": 0.8383838383838383, + "test_auc": 0.9016326530612244, + "val_auc": 0.8982371794871795, + "k": 16, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 530, + 794, + 2046, + 777, + 1345, + 765, + 1182, + 567, + 680, + 1758, + 6400, + 1335, + 613 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f8a1f437d8b82b9a14f0172ac0a33367e9537a0a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9189433817942175, + "test_acc": 0.9189765458422174, + "test_auc": 0.9407892344062556, + "val_auc": 0.9323999999999999, + "k": 1, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655 + ] + }, + { + "test_f1": 0.929550416241899, + "test_acc": 0.929637526652452, + "test_auc": 0.9818330605564648, + "val_auc": 0.9844, + "k": 2, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 12485 + ] + }, + { + "test_f1": 0.9488245002631316, + "test_acc": 0.9488272921108742, + "test_auc": 0.9877886888525187, + "val_auc": 0.9948, + "k": 5, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 12485, + 1417, + 28188, + 1927 + ] + }, + { + "test_f1": 0.9573556890165338, + "test_acc": 0.9573560767590619, + "test_auc": 0.9900163666121113, + "val_auc": 0.9936, + "k": 10, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 12485, + 1417, + 28188, + 1927, + 4080, + 1625, + 5406, + 330, + 727 + ] + }, + { + "test_f1": 0.9637523356595585, + "test_acc": 0.9637526652452025, + "test_auc": 0.9934351700309147, + "val_auc": 0.994, + "k": 16, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 12485, + 1417, + 28188, + 1927, + 4080, + 1625, + 5406, + 330, + 727, + 1118, + 6096, + 5664, + 984, + 567, + 3958 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/saebench_sparse_probing_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/saebench_sparse_probing_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..52dc9da6ff98eb5d2e3fddbf7c68c90545a8b394 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/saebench_sparse_probing_custom_sae_eval_results.json @@ -0,0 +1,782 @@ +{ + "eval_type_id": "sparse_probing", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "LabHC/bias_in_bios_class_set2", + "LabHC/bias_in_bios_class_set3", + "canrager/amazon_reviews_mcauley_1and5", + "canrager/amazon_reviews_mcauley_1and5_sentiment", + "codeparrot/github-code", + "fancyzhx/ag_news", + "Helsinki-NLP/europarl" + ], + "probe_train_set_size": 4000, + "probe_test_set_size": 1000, + "context_length": 128, + "sae_batch_size": 125, + "llm_batch_size": 128, + "llm_dtype": "float32", + "model_name": "gemma-2-2b", + "k_values": [ + 1, + 2, + 5, + 10 + ], + "lower_vram_usage": false + }, + "eval_id": "c9f9397a-1d94-4d8a-88e9-d4216504434f", + "datetime_epoch_millis": 1776879705732, + "eval_result_metrics": { + "llm": { + "llm_test_accuracy": 0.9584312926977874, + "llm_top_1_test_accuracy": 0.66844375, + "llm_top_2_test_accuracy": 0.71781875, + "llm_top_5_test_accuracy": 0.786, + "llm_top_10_test_accuracy": 0.83568125, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null + }, + "sae": { + "sae_test_accuracy": 0.9560437943786383, + "sae_top_1_test_accuracy": 0.76995, + "sae_top_2_test_accuracy": 0.8146375, + "sae_top_5_test_accuracy": 0.8768499999999999, + "sae_top_10_test_accuracy": 0.9115312499999999, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_results", + "llm_test_accuracy": 0.9674000382423401, + "llm_top_1_test_accuracy": 0.6487999999999999, + "llm_top_2_test_accuracy": 0.6956, + "llm_top_5_test_accuracy": 0.7882, + "llm_top_10_test_accuracy": 0.8400000000000001, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.96500004529953, + "sae_top_1_test_accuracy": 0.7444, + "sae_top_2_test_accuracy": 0.7849999999999999, + "sae_top_5_test_accuracy": 0.8465999999999999, + "sae_top_10_test_accuracy": 0.9048, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set2_results", + "llm_test_accuracy": 0.953600037097931, + "llm_top_1_test_accuracy": 0.6714, + "llm_top_2_test_accuracy": 0.7110000000000001, + "llm_top_5_test_accuracy": 0.7674, + "llm_top_10_test_accuracy": 0.807, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9472000598907471, + "sae_top_1_test_accuracy": 0.7026, + "sae_top_2_test_accuracy": 0.7856, + "sae_top_5_test_accuracy": 0.841, + "sae_top_10_test_accuracy": 0.9057999999999999, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set3_results", + "llm_test_accuracy": 0.9302000522613525, + "llm_top_1_test_accuracy": 0.6878, + "llm_top_2_test_accuracy": 0.7394000000000001, + "llm_top_5_test_accuracy": 0.7634, + "llm_top_10_test_accuracy": 0.8048, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9286000609397889, + "sae_top_1_test_accuracy": 0.7134, + "sae_top_2_test_accuracy": 0.8263999999999999, + "sae_top_5_test_accuracy": 0.8654, + "sae_top_10_test_accuracy": 0.8884000000000001, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_results", + "llm_test_accuracy": 0.9136000394821167, + "llm_top_1_test_accuracy": 0.5955999999999999, + "llm_top_2_test_accuracy": 0.6424, + "llm_top_5_test_accuracy": 0.7004, + "llm_top_10_test_accuracy": 0.7674, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9158000469207763, + "sae_top_1_test_accuracy": 0.663, + "sae_top_2_test_accuracy": 0.7026, + "sae_top_5_test_accuracy": 0.7996000000000001, + "sae_top_10_test_accuracy": 0.8314, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_sentiment_results", + "llm_test_accuracy": 0.9810000658035278, + "llm_top_1_test_accuracy": 0.674, + "llm_top_2_test_accuracy": 0.724, + "llm_top_5_test_accuracy": 0.764, + "llm_top_10_test_accuracy": 0.827, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9750000238418579, + "sae_top_1_test_accuracy": 0.918, + "sae_top_2_test_accuracy": 0.936, + "sae_top_5_test_accuracy": 0.936, + "sae_top_10_test_accuracy": 0.959, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "codeparrot/github-code_results", + "llm_test_accuracy": 0.9700000405311584, + "llm_top_1_test_accuracy": 0.6452, + "llm_top_2_test_accuracy": 0.6862000000000001, + "llm_top_5_test_accuracy": 0.7644, + "llm_top_10_test_accuracy": 0.8066000000000001, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9670000433921814, + "sae_top_1_test_accuracy": 0.739, + "sae_top_2_test_accuracy": 0.74, + "sae_top_5_test_accuracy": 0.8677999999999999, + "sae_top_10_test_accuracy": 0.8968, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "fancyzhx/ag_news_results", + "llm_test_accuracy": 0.9522500485181808, + "llm_top_1_test_accuracy": 0.69575, + "llm_top_2_test_accuracy": 0.76175, + "llm_top_5_test_accuracy": 0.831, + "llm_top_10_test_accuracy": 0.87025, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9507500380277634, + "sae_top_1_test_accuracy": 0.7949999999999999, + "sae_top_2_test_accuracy": 0.8285, + "sae_top_5_test_accuracy": 0.872, + "sae_top_10_test_accuracy": 0.9112500000000001, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + }, + { + "dataset_name": "Helsinki-NLP/europarl_results", + "llm_test_accuracy": 0.9994000196456909, + "llm_top_1_test_accuracy": 0.729, + "llm_top_2_test_accuracy": 0.7822, + "llm_top_5_test_accuracy": 0.9092, + "llm_top_10_test_accuracy": 0.9624, + "llm_top_20_test_accuracy": null, + "llm_top_50_test_accuracy": null, + "llm_top_100_test_accuracy": null, + "sae_test_accuracy": 0.9990000367164612, + "sae_top_1_test_accuracy": 0.8841999999999999, + "sae_top_2_test_accuracy": 0.913, + "sae_top_5_test_accuracy": 0.9863999999999999, + "sae_top_10_test_accuracy": 0.9947999999999999, + "sae_top_20_test_accuracy": null, + "sae_top_50_test_accuracy": null, + "sae_top_100_test_accuracy": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_sparse_probing", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1_results": { + "sae_test_accuracy": { + "0": 0.9440000653266907, + "1": 0.9610000252723694, + "2": 0.9600000381469727, + "6": 0.9810000658035278, + "9": 0.9790000319480896 + }, + "llm_test_accuracy": { + "0": 0.9480000734329224, + "1": 0.9650000333786011, + "2": 0.9510000348091125, + "6": 0.9930000305175781, + "9": 0.9800000190734863 + }, + "llm_top_1_test_accuracy": { + "0": 0.586, + "1": 0.633, + "2": 0.689, + "6": 0.803, + "9": 0.533 + }, + "llm_top_2_test_accuracy": { + "0": 0.587, + "1": 0.645, + "2": 0.716, + "6": 0.82, + "9": 0.71 + }, + "llm_top_5_test_accuracy": { + "0": 0.72, + "1": 0.707, + "2": 0.768, + "6": 0.891, + "9": 0.855 + }, + "llm_top_10_test_accuracy": { + "0": 0.788, + "1": 0.796, + "2": 0.809, + "6": 0.902, + "9": 0.905 + }, + "sae_top_1_test_accuracy": { + "0": 0.591, + "1": 0.605, + "2": 0.786, + "6": 0.801, + "9": 0.939 + }, + "sae_top_2_test_accuracy": { + "0": 0.573, + "1": 0.623, + "2": 0.822, + "6": 0.976, + "9": 0.931 + }, + "sae_top_5_test_accuracy": { + "0": 0.733, + "1": 0.706, + "2": 0.859, + "6": 0.981, + "9": 0.954 + }, + "sae_top_10_test_accuracy": { + "0": 0.832, + "1": 0.882, + "2": 0.864, + "6": 0.992, + "9": 0.954 + } + }, + "LabHC/bias_in_bios_class_set2_results": { + "sae_test_accuracy": { + "11": 0.9600000381469727, + "13": 0.9530000686645508, + "14": 0.9530000686645508, + "18": 0.9160000681877136, + "19": 0.9540000557899475 + }, + "llm_test_accuracy": { + "11": 0.9630000591278076, + "13": 0.9520000219345093, + "14": 0.9650000333786011, + "18": 0.9330000281333923, + "19": 0.9550000429153442 + }, + "llm_top_1_test_accuracy": { + "11": 0.567, + "13": 0.682, + "14": 0.627, + "18": 0.693, + "19": 0.788 + }, + "llm_top_2_test_accuracy": { + "11": 0.682, + "13": 0.709, + "14": 0.676, + "18": 0.718, + "19": 0.77 + }, + "llm_top_5_test_accuracy": { + "11": 0.793, + "13": 0.75, + "14": 0.719, + "18": 0.731, + "19": 0.844 + }, + "llm_top_10_test_accuracy": { + "11": 0.838, + "13": 0.79, + "14": 0.789, + "18": 0.755, + "19": 0.863 + }, + "sae_top_1_test_accuracy": { + "11": 0.637, + "13": 0.694, + "14": 0.685, + "18": 0.648, + "19": 0.849 + }, + "sae_top_2_test_accuracy": { + "11": 0.857, + "13": 0.673, + "14": 0.87, + "18": 0.697, + "19": 0.831 + }, + "sae_top_5_test_accuracy": { + "11": 0.926, + "13": 0.797, + "14": 0.89, + "18": 0.742, + "19": 0.85 + }, + "sae_top_10_test_accuracy": { + "11": 0.944, + "13": 0.904, + "14": 0.911, + "18": 0.876, + "19": 0.894 + } + }, + "LabHC/bias_in_bios_class_set3_results": { + "sae_test_accuracy": { + "20": 0.9530000686645508, + "21": 0.9250000715255737, + "22": 0.9160000681877136, + "25": 0.9630000591278076, + "26": 0.8860000371932983 + }, + "llm_test_accuracy": { + "20": 0.9600000381469727, + "21": 0.9290000200271606, + "22": 0.9110000729560852, + "25": 0.9630000591278076, + "26": 0.8880000710487366 + }, + "llm_top_1_test_accuracy": { + "20": 0.698, + "21": 0.77, + "22": 0.641, + "25": 0.699, + "26": 0.631 + }, + "llm_top_2_test_accuracy": { + "20": 0.809, + "21": 0.774, + "22": 0.673, + "25": 0.771, + "26": 0.67 + }, + "llm_top_5_test_accuracy": { + "20": 0.818, + "21": 0.797, + "22": 0.706, + "25": 0.811, + "26": 0.685 + }, + "llm_top_10_test_accuracy": { + "20": 0.88, + "21": 0.828, + "22": 0.75, + "25": 0.843, + "26": 0.723 + }, + "sae_top_1_test_accuracy": { + "20": 0.885, + "21": 0.558, + "22": 0.613, + "25": 0.872, + "26": 0.639 + }, + "sae_top_2_test_accuracy": { + "20": 0.904, + "21": 0.826, + "22": 0.897, + "25": 0.872, + "26": 0.633 + }, + "sae_top_5_test_accuracy": { + "20": 0.922, + "21": 0.843, + "22": 0.887, + "25": 0.909, + "26": 0.766 + }, + "sae_top_10_test_accuracy": { + "20": 0.935, + "21": 0.853, + "22": 0.899, + "25": 0.934, + "26": 0.821 + } + }, + "canrager/amazon_reviews_mcauley_1and5_results": { + "sae_test_accuracy": { + "1": 0.9480000734329224, + "2": 0.937000036239624, + "3": 0.9220000505447388, + "5": 0.9150000214576721, + "6": 0.8570000529289246 + }, + "llm_test_accuracy": { + "1": 0.9530000686645508, + "2": 0.9290000200271606, + "3": 0.9200000166893005, + "5": 0.9110000729560852, + "6": 0.8550000190734863 + }, + "llm_top_1_test_accuracy": { + "1": 0.661, + "2": 0.596, + "3": 0.574, + "5": 0.567, + "6": 0.58 + }, + "llm_top_2_test_accuracy": { + "1": 0.739, + "2": 0.617, + "3": 0.607, + "5": 0.628, + "6": 0.621 + }, + "llm_top_5_test_accuracy": { + "1": 0.781, + "2": 0.739, + "3": 0.681, + "5": 0.636, + "6": 0.665 + }, + "llm_top_10_test_accuracy": { + "1": 0.823, + "2": 0.79, + "3": 0.751, + "5": 0.739, + "6": 0.734 + }, + "sae_top_1_test_accuracy": { + "1": 0.822, + "2": 0.62, + "3": 0.574, + "5": 0.76, + "6": 0.539 + }, + "sae_top_2_test_accuracy": { + "1": 0.861, + "2": 0.672, + "3": 0.615, + "5": 0.762, + "6": 0.603 + }, + "sae_top_5_test_accuracy": { + "1": 0.923, + "2": 0.891, + "3": 0.622, + "5": 0.854, + "6": 0.708 + }, + "sae_top_10_test_accuracy": { + "1": 0.918, + "2": 0.908, + "3": 0.721, + "5": 0.861, + "6": 0.749 + } + }, + "canrager/amazon_reviews_mcauley_1and5_sentiment_results": { + "sae_test_accuracy": { + "1.0": 0.9750000238418579, + "5.0": 0.9750000238418579 + }, + "llm_test_accuracy": { + "1.0": 0.9810000658035278, + "5.0": 0.9810000658035278 + }, + "llm_top_1_test_accuracy": { + "1.0": 0.674, + "5.0": 0.674 + }, + "llm_top_2_test_accuracy": { + "1.0": 0.724, + "5.0": 0.724 + }, + "llm_top_5_test_accuracy": { + "1.0": 0.764, + "5.0": 0.764 + }, + "llm_top_10_test_accuracy": { + "1.0": 0.827, + "5.0": 0.827 + }, + "sae_top_1_test_accuracy": { + "1.0": 0.918, + "5.0": 0.918 + }, + "sae_top_2_test_accuracy": { + "1.0": 0.936, + "5.0": 0.936 + }, + "sae_top_5_test_accuracy": { + "1.0": 0.936, + "5.0": 0.936 + }, + "sae_top_10_test_accuracy": { + "1.0": 0.959, + "5.0": 0.959 + } + }, + "codeparrot/github-code_results": { + "sae_test_accuracy": { + "C": 0.9480000734329224, + "Python": 0.9830000400543213, + "HTML": 0.984000027179718, + "Java": 0.9640000462532043, + "PHP": 0.956000030040741 + }, + "llm_test_accuracy": { + "C": 0.9530000686645508, + "Python": 0.9890000224113464, + "HTML": 0.9900000691413879, + "Java": 0.9610000252723694, + "PHP": 0.9570000171661377 + }, + "llm_top_1_test_accuracy": { + "C": 0.668, + "Python": 0.633, + "HTML": 0.714, + "Java": 0.634, + "PHP": 0.577 + }, + "llm_top_2_test_accuracy": { + "C": 0.668, + "Python": 0.674, + "HTML": 0.785, + "Java": 0.683, + "PHP": 0.621 + }, + "llm_top_5_test_accuracy": { + "C": 0.756, + "Python": 0.732, + "HTML": 0.896, + "Java": 0.734, + "PHP": 0.704 + }, + "llm_top_10_test_accuracy": { + "C": 0.805, + "Python": 0.743, + "HTML": 0.948, + "Java": 0.802, + "PHP": 0.735 + }, + "sae_top_1_test_accuracy": { + "C": 0.617, + "Python": 0.65, + "HTML": 0.912, + "Java": 0.628, + "PHP": 0.888 + }, + "sae_top_2_test_accuracy": { + "C": 0.649, + "Python": 0.633, + "HTML": 0.915, + "Java": 0.62, + "PHP": 0.883 + }, + "sae_top_5_test_accuracy": { + "C": 0.881, + "Python": 0.911, + "HTML": 0.943, + "Java": 0.702, + "PHP": 0.902 + }, + "sae_top_10_test_accuracy": { + "C": 0.889, + "Python": 0.913, + "HTML": 0.954, + "Java": 0.81, + "PHP": 0.918 + } + }, + "fancyzhx/ag_news_results": { + "sae_test_accuracy": { + "0": 0.9350000619888306, + "1": 0.9830000400543213, + "2": 0.9330000281333923, + "3": 0.9520000219345093 + }, + "llm_test_accuracy": { + "0": 0.9360000491142273, + "1": 0.9860000610351562, + "2": 0.9320000410079956, + "3": 0.9550000429153442 + }, + "llm_top_1_test_accuracy": { + "0": 0.809, + "1": 0.671, + "2": 0.67, + "3": 0.633 + }, + "llm_top_2_test_accuracy": { + "0": 0.804, + "1": 0.804, + "2": 0.702, + "3": 0.737 + }, + "llm_top_5_test_accuracy": { + "0": 0.826, + "1": 0.883, + "2": 0.764, + "3": 0.851 + }, + "llm_top_10_test_accuracy": { + "0": 0.863, + "1": 0.933, + "2": 0.801, + "3": 0.884 + }, + "sae_top_1_test_accuracy": { + "0": 0.841, + "1": 0.832, + "2": 0.836, + "3": 0.671 + }, + "sae_top_2_test_accuracy": { + "0": 0.857, + "1": 0.949, + "2": 0.853, + "3": 0.655 + }, + "sae_top_5_test_accuracy": { + "0": 0.858, + "1": 0.974, + "2": 0.851, + "3": 0.805 + }, + "sae_top_10_test_accuracy": { + "0": 0.911, + "1": 0.979, + "2": 0.854, + "3": 0.901 + } + }, + "Helsinki-NLP/europarl_results": { + "sae_test_accuracy": { + "en": 0.999000072479248, + "fr": 1.0, + "de": 0.999000072479248, + "es": 0.9970000386238098, + "nl": 1.0 + }, + "llm_test_accuracy": { + "en": 1.0, + "fr": 1.0, + "de": 1.0, + "es": 0.999000072479248, + "nl": 0.9980000257492065 + }, + "llm_top_1_test_accuracy": { + "en": 0.755, + "fr": 0.59, + "de": 0.751, + "es": 0.907, + "nl": 0.642 + }, + "llm_top_2_test_accuracy": { + "en": 0.825, + "fr": 0.593, + "de": 0.827, + "es": 0.916, + "nl": 0.75 + }, + "llm_top_5_test_accuracy": { + "en": 0.891, + "fr": 0.899, + "de": 0.919, + "es": 0.978, + "nl": 0.859 + }, + "llm_top_10_test_accuracy": { + "en": 0.978, + "fr": 0.966, + "de": 0.963, + "es": 0.99, + "nl": 0.915 + }, + "sae_top_1_test_accuracy": { + "en": 0.999, + "fr": 0.99, + "de": 0.919, + "es": 0.86, + "nl": 0.653 + }, + "sae_top_2_test_accuracy": { + "en": 0.999, + "fr": 0.994, + "de": 0.914, + "es": 0.991, + "nl": 0.667 + }, + "sae_top_5_test_accuracy": { + "en": 0.998, + "fr": 0.997, + "de": 0.946, + "es": 0.994, + "nl": 0.997 + }, + "sae_top_10_test_accuracy": { + "en": 0.999, + "fr": 0.994, + "de": 0.985, + "es": 0.997, + "nl": 0.999 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/scr/saebench_scr_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/scr/saebench_scr_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..5dbfc605aa844b9b2355a1a55c2c1cdbb0ff79d8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/scr/saebench_scr_custom_sae_eval_results.json @@ -0,0 +1,326 @@ +{ + "eval_type_id": "scr", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": true, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "9c96b6f1-8acd-4c98-900f-49b32989b9fd", + "datetime_epoch_millis": 1776879254905, + "eval_result_metrics": { + "scr_metrics": { + "scr_dir1_threshold_2": 0.2565406098407402, + "scr_metric_threshold_2": 0.18136855287184347, + "scr_dir2_threshold_2": 0.1752704813081063, + "scr_dir1_threshold_5": 0.08430090154966209, + "scr_metric_threshold_5": 0.23282672333667337, + "scr_dir2_threshold_5": 0.22245370159286793, + "scr_dir1_threshold_10": 0.15526437488462547, + "scr_metric_threshold_10": 0.29113307196002924, + "scr_dir2_threshold_10": 0.27203912723680074, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results", + "scr_dir1_threshold_2": 0.5384614208981502, + "scr_metric_threshold_2": 0.06476677137111979, + "scr_dir2_threshold_2": 0.06476677137111979, + "scr_dir1_threshold_5": 0.6025646805840948, + "scr_metric_threshold_5": 0.21502589713474632, + "scr_dir2_threshold_5": 0.21502589713474632, + "scr_dir1_threshold_10": 0.5769232238773122, + "scr_metric_threshold_10": 0.22538861143735814, + "scr_dir2_threshold_10": 0.22538861143735814, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results", + "scr_dir1_threshold_2": 0.23484857036936357, + "scr_metric_threshold_2": 0.33027532969386547, + "scr_dir2_threshold_2": 0.33027532969386547, + "scr_dir1_threshold_5": -0.6666665161499201, + "scr_metric_threshold_5": 0.2660550659387731, + "scr_dir2_threshold_5": 0.2660550659387731, + "scr_dir1_threshold_10": -0.24999988711244006, + "scr_metric_threshold_10": 0.4556574282511415, + "scr_dir2_threshold_10": 0.4556574282511415, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results", + "scr_dir1_threshold_2": 0.3749997671693945, + "scr_metric_threshold_2": 0.05527641050700931, + "scr_dir2_threshold_2": 0.05527641050700931, + "scr_dir1_threshold_5": 0.09374970896174313, + "scr_metric_threshold_5": 0.10804029093651613, + "scr_dir2_threshold_5": 0.10804029093651613, + "scr_dir1_threshold_10": 0.20312514551912844, + "scr_metric_threshold_10": 0.18341709182393628, + "scr_dir2_threshold_10": 0.18341709182393628, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results", + "scr_dir1_threshold_2": 0.21476475165062875, + "scr_metric_threshold_2": -0.01910822342857295, + "scr_dir2_threshold_2": -0.01910822342857295, + "scr_dir1_threshold_5": -0.3020135570574505, + "scr_metric_threshold_5": -0.10509551359270591, + "scr_dir2_threshold_5": -0.10509551359270591, + "scr_dir1_threshold_10": -0.5369130792201792, + "scr_metric_threshold_10": -0.15923570143464957, + "scr_dir2_threshold_10": -0.15923570143464957, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results", + "scr_dir1_threshold_2": 0.03684225881529179, + "scr_metric_threshold_2": 0.49603167939138937, + "scr_dir2_threshold_2": 0.49603167939138937, + "scr_dir1_threshold_5": 0.052631529414422, + "scr_metric_threshold_5": 0.5873015159682478, + "scr_dir2_threshold_5": 0.5873015159682478, + "scr_dir1_threshold_10": 0.07368420392192293, + "scr_metric_threshold_10": 0.6111109665672388, + "scr_dir2_threshold_10": 0.6111109665672388, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results", + "scr_dir1_threshold_2": 0.2896173902280937, + "scr_metric_threshold_2": 0.11450392446499717, + "scr_dir2_threshold_2": 0.11450392446499717, + "scr_dir1_threshold_5": 0.3825136878996036, + "scr_metric_threshold_5": 0.1603053122520976, + "scr_dir2_threshold_5": 0.1603053122520976, + "scr_dir1_threshold_10": 0.4590164254795243, + "scr_metric_threshold_10": 0.21755716073528464, + "scr_dir2_threshold_10": 0.21755716073528464, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results", + "scr_dir1_threshold_2": 0.23255815887657183, + "scr_metric_threshold_2": 0.2789699702565123, + "scr_dir2_threshold_2": 0.2789699702565123, + "scr_dir1_threshold_5": 0.2883719395791968, + "scr_metric_threshold_5": 0.40772547889010513, + "scr_dir2_threshold_5": 0.40772547889010513, + "scr_dir1_threshold_10": 0.4232558436107435, + "scr_metric_threshold_10": 0.5021458952989327, + "scr_dir2_threshold_10": 0.5021458952989327, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results", + "scr_dir1_threshold_2": 0.13023256071842734, + "scr_metric_threshold_2": 0.13023256071842734, + "scr_dir2_threshold_2": 0.08144798820852994, + "scr_dir1_threshold_5": 0.22325573916560695, + "scr_metric_threshold_5": 0.22325573916560695, + "scr_dir2_threshold_5": 0.14027156521516332, + "scr_dir1_threshold_10": 0.29302312300099165, + "scr_metric_threshold_10": 0.29302312300099165, + "scr_dir2_threshold_10": 0.14027156521516332, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_scr", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": null +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/tpp/saebench_tpp_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/tpp/saebench_tpp_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..c7287ad2468c376eb8d86eb751465e0f775e1405 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-122070-tokens-500000000/tpp/saebench_tpp_custom_sae_eval_results.json @@ -0,0 +1,297 @@ +{ + "eval_type_id": "tpp", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": false, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "eaf20f93-9d31-4ab7-ba4f-a1f8ce28e8b0", + "datetime_epoch_millis": 1776879449297, + "eval_result_metrics": { + "tpp_metrics": { + "tpp_threshold_2_total_metric": 0.04002499580383301, + "tpp_threshold_2_intended_diff_only": 0.046400004625320436, + "tpp_threshold_2_unintended_diff_only": 0.006375008821487427, + "tpp_threshold_5_total_metric": 0.07825000733137132, + "tpp_threshold_5_intended_diff_only": 0.0946000099182129, + "tpp_threshold_5_unintended_diff_only": 0.01635000258684158, + "tpp_threshold_10_total_metric": 0.13287501484155656, + "tpp_threshold_10_intended_diff_only": 0.17720001339912414, + "tpp_threshold_10_unintended_diff_only": 0.0443249985575676, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results", + "tpp_threshold_2_total_metric": 0.06974999010562896, + "tpp_threshold_2_intended_diff_only": 0.08100000619888306, + "tpp_threshold_2_unintended_diff_only": 0.01125001609325409, + "tpp_threshold_5_total_metric": 0.13275001049041749, + "tpp_threshold_5_intended_diff_only": 0.1610000252723694, + "tpp_threshold_5_unintended_diff_only": 0.028250014781951903, + "tpp_threshold_10_total_metric": 0.21845002174377443, + "tpp_threshold_10_intended_diff_only": 0.2994000315666199, + "tpp_threshold_10_unintended_diff_only": 0.08095000982284546, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results", + "tpp_threshold_2_total_metric": 0.010300001502037049, + "tpp_threshold_2_intended_diff_only": 0.011800003051757813, + "tpp_threshold_2_unintended_diff_only": 0.001500001549720764, + "tpp_threshold_5_total_metric": 0.023750004172325135, + "tpp_threshold_5_intended_diff_only": 0.028199994564056398, + "tpp_threshold_5_unintended_diff_only": 0.004449990391731262, + "tpp_threshold_10_total_metric": 0.04730000793933868, + "tpp_threshold_10_intended_diff_only": 0.05499999523162842, + "tpp_threshold_10_unintended_diff_only": 0.007699987292289734, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_tpp", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1": { + "0": { + "tpp_threshold_2_total_metric": 0.16449995338916779, + "tpp_threshold_2_intended_diff_only": 0.20499998331069946, + "tpp_threshold_2_unintended_diff_only": 0.04050002992153168, + "tpp_threshold_5_total_metric": 0.19875000417232513, + "tpp_threshold_5_intended_diff_only": 0.24800002574920654, + "tpp_threshold_5_unintended_diff_only": 0.04925002157688141, + "tpp_threshold_10_total_metric": 0.23099996149539948, + "tpp_threshold_10_intended_diff_only": 0.30699998140335083, + "tpp_threshold_10_unintended_diff_only": 0.07600001990795135 + }, + "1": { + "tpp_threshold_2_total_metric": 0.0007499605417251587, + "tpp_threshold_2_intended_diff_only": 0.0009999871253967285, + "tpp_threshold_2_unintended_diff_only": 0.0002500265836715698, + "tpp_threshold_5_total_metric": 0.0847499817609787, + "tpp_threshold_5_intended_diff_only": 0.11100000143051147, + "tpp_threshold_5_unintended_diff_only": 0.026250019669532776, + "tpp_threshold_10_total_metric": 0.17400000989437103, + "tpp_threshold_10_intended_diff_only": 0.23000001907348633, + "tpp_threshold_10_unintended_diff_only": 0.056000009179115295 + }, + "2": { + "tpp_threshold_2_total_metric": 0.12324999272823334, + "tpp_threshold_2_intended_diff_only": 0.1340000033378601, + "tpp_threshold_2_unintended_diff_only": 0.01075001060962677, + "tpp_threshold_5_total_metric": 0.18000003695487976, + "tpp_threshold_5_intended_diff_only": 0.21400004625320435, + "tpp_threshold_5_unintended_diff_only": 0.034000009298324585, + "tpp_threshold_10_total_metric": 0.18250006437301636, + "tpp_threshold_10_intended_diff_only": 0.2820000648498535, + "tpp_threshold_10_unintended_diff_only": 0.09950000047683716 + }, + "6": { + "tpp_threshold_2_total_metric": 0.050999999046325684, + "tpp_threshold_2_intended_diff_only": 0.05500000715255737, + "tpp_threshold_2_unintended_diff_only": 0.0040000081062316895, + "tpp_threshold_5_total_metric": 0.0729999840259552, + "tpp_threshold_5_intended_diff_only": 0.08300000429153442, + "tpp_threshold_5_unintended_diff_only": 0.010000020265579224, + "tpp_threshold_10_total_metric": 0.3137500435113907, + "tpp_threshold_10_intended_diff_only": 0.36600005626678467, + "tpp_threshold_10_unintended_diff_only": 0.05225001275539398 + }, + "9": { + "tpp_threshold_2_total_metric": 0.009250044822692871, + "tpp_threshold_2_intended_diff_only": 0.010000050067901611, + "tpp_threshold_2_unintended_diff_only": 0.0007500052452087402, + "tpp_threshold_5_total_metric": 0.1272500455379486, + "tpp_threshold_5_intended_diff_only": 0.14900004863739014, + "tpp_threshold_5_unintended_diff_only": 0.02175000309944153, + "tpp_threshold_10_total_metric": 0.19100002944469452, + "tpp_threshold_10_intended_diff_only": 0.312000036239624, + "tpp_threshold_10_unintended_diff_only": 0.1210000067949295 + } + }, + "canrager/amazon_reviews_mcauley_1and5": { + "1": { + "tpp_threshold_2_total_metric": 0.011750012636184692, + "tpp_threshold_2_intended_diff_only": 0.008000016212463379, + "tpp_threshold_2_unintended_diff_only": -0.0037499964237213135, + "tpp_threshold_5_total_metric": 0.01124998927116394, + "tpp_threshold_5_intended_diff_only": 0.010999977588653564, + "tpp_threshold_5_unintended_diff_only": -0.000250011682510376, + "tpp_threshold_10_total_metric": 0.013250023126602173, + "tpp_threshold_10_intended_diff_only": 0.022000014781951904, + "tpp_threshold_10_unintended_diff_only": 0.008749991655349731 + }, + "2": { + "tpp_threshold_2_total_metric": 0.004500046372413635, + "tpp_threshold_2_intended_diff_only": 0.01100003719329834, + "tpp_threshold_2_unintended_diff_only": 0.006499990820884705, + "tpp_threshold_5_total_metric": 0.023500055074691772, + "tpp_threshold_5_intended_diff_only": 0.03400003910064697, + "tpp_threshold_5_unintended_diff_only": 0.0104999840259552, + "tpp_threshold_10_total_metric": 0.07225002348423004, + "tpp_threshold_10_intended_diff_only": 0.08300000429153442, + "tpp_threshold_10_unintended_diff_only": 0.010749980807304382 + }, + "3": { + "tpp_threshold_2_total_metric": 0.016999974846839905, + "tpp_threshold_2_intended_diff_only": 0.014999985694885254, + "tpp_threshold_2_unintended_diff_only": -0.001999989151954651, + "tpp_threshold_5_total_metric": 0.020500019192695618, + "tpp_threshold_5_intended_diff_only": 0.023000001907348633, + "tpp_threshold_5_unintended_diff_only": 0.002499982714653015, + "tpp_threshold_10_total_metric": 0.026250049471855164, + "tpp_threshold_10_intended_diff_only": 0.0350000262260437, + "tpp_threshold_10_unintended_diff_only": 0.008749976754188538 + }, + "5": { + "tpp_threshold_2_total_metric": -0.0012500286102294922, + "tpp_threshold_2_intended_diff_only": 0.006999969482421875, + "tpp_threshold_2_unintended_diff_only": 0.008249998092651367, + "tpp_threshold_5_total_metric": 0.012749984860420227, + "tpp_threshold_5_intended_diff_only": 0.014999985694885254, + "tpp_threshold_5_unintended_diff_only": 0.002250000834465027, + "tpp_threshold_10_total_metric": 0.027499958872795105, + "tpp_threshold_10_intended_diff_only": 0.031999945640563965, + "tpp_threshold_10_unintended_diff_only": 0.00449998676776886 + }, + "6": { + "tpp_threshold_2_total_metric": 0.0195000022649765, + "tpp_threshold_2_intended_diff_only": 0.018000006675720215, + "tpp_threshold_2_unintended_diff_only": -0.0014999955892562866, + "tpp_threshold_5_total_metric": 0.050749972462654114, + "tpp_threshold_5_intended_diff_only": 0.05799996852874756, + "tpp_threshold_5_unintended_diff_only": 0.007249996066093445, + "tpp_threshold_10_total_metric": 0.09724998474121094, + "tpp_threshold_10_intended_diff_only": 0.1029999852180481, + "tpp_threshold_10_unintended_diff_only": 0.005750000476837158 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cdc31228ea011d1c41a696263b7672ac2423ce2d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9751747215395078, + "test_acc": 0.9751789976133651, + "test_auc": 0.9806471780514302, + "val_auc": 0.9851994851994852, + "k": 1, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431 + ] + }, + { + "test_f1": 0.9780373449880904, + "test_acc": 0.9780429594272076, + "test_auc": 0.9935192881150798, + "val_auc": 0.9927069927069927, + "k": 2, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928 + ] + }, + { + "test_f1": 0.9804277213459535, + "test_acc": 0.9804295942720763, + "test_auc": 0.9977913085004775, + "val_auc": 0.9982839982839983, + "k": 5, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472 + ] + }, + { + "test_f1": 0.9918854378297842, + "test_acc": 0.9918854415274463, + "test_auc": 0.9994477132045757, + "val_auc": 0.9982839982839983, + "k": 10, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472, + 1411, + 5951, + 461, + 786, + 542 + ] + }, + { + "test_f1": 0.9832908835720354, + "test_acc": 0.9832935560859188, + "test_auc": 0.9987578103924699, + "val_auc": 0.9965679965679966, + "k": 16, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472, + 1411, + 5951, + 461, + 786, + 542, + 126, + 4674, + 1634, + 5020, + 1872, + 4110 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..508df11d7092550caa24f94ae79879119662113a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9274085792226037, + "test_acc": 0.9274141283214518, + "test_auc": 0.9579343158404063, + "val_auc": 0.9462587846217446, + "k": 1, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838 + ] + }, + { + "test_f1": 0.9552784996608675, + "test_acc": 0.9552819183408944, + "test_auc": 0.9651871938065764, + "val_auc": 0.9565936337329476, + "k": 2, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441 + ] + }, + { + "test_f1": 0.9585039536600967, + "test_acc": 0.9585223590408296, + "test_auc": 0.970261016242952, + "val_auc": 0.9681686647374949, + "k": 5, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 943 + ] + }, + { + "test_f1": 0.9507260538712988, + "test_acc": 0.9507453013609851, + "test_auc": 0.9728668104809715, + "val_auc": 0.9669284828441506, + "k": 10, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 943, + 1697, + 1180, + 698, + 654, + 908 + ] + }, + { + "test_f1": 0.9578493786378132, + "test_acc": 0.9578742709008425, + "test_auc": 0.9825037129627763, + "val_auc": 0.985944605208764, + "k": 16, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 943, + 1697, + 1180, + 698, + 654, + 908, + 1411, + 30737, + 976, + 302, + 1417, + 601 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9cb9fd59c09e391641e51e0c1530ba7fc17daace --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5702438051699386, + "test_acc": 0.5851572327044026, + "test_auc": 0.592469386019185, + "val_auc": 0.6254501800720288, + "k": 1, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921 + ] + }, + { + "test_f1": 0.5720784355897249, + "test_acc": 0.5871698113207547, + "test_auc": 0.6069628642514371, + "val_auc": 0.6166466586634654, + "k": 2, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777 + ] + }, + { + "test_f1": 0.5839655335453441, + "test_acc": 0.5944654088050314, + "test_auc": 0.6206059963201453, + "val_auc": 0.6106442577030813, + "k": 5, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 2046 + ] + }, + { + "test_f1": 0.5894686686356223, + "test_acc": 0.5952201257861636, + "test_auc": 0.6192194688007259, + "val_auc": 0.6302521008403361, + "k": 10, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 2046, + 505, + 1380, + 969, + 1335, + 1178 + ] + }, + { + "test_f1": 0.5965940358251131, + "test_acc": 0.6002515723270441, + "test_auc": 0.6227828723726354, + "val_auc": 0.6522609043617447, + "k": 16, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 2046, + 505, + 1380, + 969, + 1335, + 1178, + 521, + 1533, + 5586, + 559, + 6411, + 26301 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1b1af748304579e758bc533edfd62c38cc2d9341 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6718251777366863, + "test_acc": 0.6789937106918239, + "test_auc": 0.6900195840366811, + "val_auc": 0.728891556622649, + "k": 1, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567 + ] + }, + { + "test_f1": 0.6993049812307934, + "test_acc": 0.7013836477987422, + "test_auc": 0.729801557204323, + "val_auc": 0.723889555822329, + "k": 2, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470 + ] + }, + { + "test_f1": 0.6990148307868451, + "test_acc": 0.7001257861635221, + "test_auc": 0.7525100274520805, + "val_auc": 0.801920768307323, + "k": 5, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 1164, + 433, + 1345 + ] + }, + { + "test_f1": 0.7210569602234581, + "test_acc": 0.7222641509433962, + "test_auc": 0.7742092970505468, + "val_auc": 0.8015206082432974, + "k": 10, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 1164, + 433, + 1345, + 559, + 6411, + 1731, + 1629, + 4505 + ] + }, + { + "test_f1": 0.726706231048462, + "test_acc": 0.7275471698113207, + "test_auc": 0.7800537497759582, + "val_auc": 0.81312525010004, + "k": 16, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 1164, + 433, + 1345, + 559, + 6411, + 1731, + 1629, + 4505, + 4196, + 628, + 794, + 993, + 1335, + 534 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0706d4a993dc0b56fe980fbdc67301327a93f5c7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7520932534916563, + "test_acc": 0.7532075471698113, + "test_auc": 0.8157265181425746, + "val_auc": 0.8585434173669468, + "k": 1, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411 + ] + }, + { + "test_f1": 0.7310090916612276, + "test_acc": 0.7333333333333333, + "test_auc": 0.823091417149095, + "val_auc": 0.8727490996398559, + "k": 2, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786 + ] + }, + { + "test_f1": 0.7598888399966138, + "test_acc": 0.760251572327044, + "test_auc": 0.8436802242746869, + "val_auc": 0.8701480592236895, + "k": 5, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1758 + ] + }, + { + "test_f1": 0.7916549784937797, + "test_acc": 0.7919496855345912, + "test_auc": 0.887862909717996, + "val_auc": 0.926170468187275, + "k": 10, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1758, + 1417, + 1335, + 181, + 5054, + 1360 + ] + }, + { + "test_f1": 0.8019793424303054, + "test_acc": 0.8020125786163522, + "test_auc": 0.8988333625304926, + "val_auc": 0.918967587034814, + "k": 16, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1758, + 1417, + 1335, + 181, + 5054, + 1360, + 1183, + 19471, + 1108, + 12220, + 126, + 191 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b51711af96c957ec1123d8c680f8d0e4d1a697a8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8463701816989496, + "test_acc": 0.8465408805031447, + "test_auc": 0.8911744245037412, + "val_auc": 0.9393757503001201, + "k": 1, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.873424753129876, + "test_acc": 0.8734591194968554, + "test_auc": 0.9352921758026771, + "val_auc": 0.9583833533413366, + "k": 2, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.872641273420193, + "test_acc": 0.8727044025157232, + "test_auc": 0.9381510502370034, + "val_auc": 0.959983993597439, + "k": 5, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 100, + 6319, + 4675 + ] + }, + { + "test_f1": 0.8721488245012882, + "test_acc": 0.8722012578616353, + "test_auc": 0.9423139238045282, + "val_auc": 0.956782713085234, + "k": 10, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 100, + 6319, + 4675, + 22141, + 821, + 505, + 6745, + 1697 + ] + }, + { + "test_f1": 0.8855334610645773, + "test_acc": 0.8855345911949686, + "test_auc": 0.9471322145251984, + "val_auc": 0.9591836734693877, + "k": 16, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 100, + 6319, + 4675, + 22141, + 821, + 505, + 6745, + 1697, + 7143, + 1438, + 3654, + 7284, + 1108, + 5256 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9c22aa7bb9a7b5e453bae185748ff56e12cabd72 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5363417106321053, + "test_acc": 0.5579487179487179, + "test_auc": 0.5653654862490322, + "val_auc": 0.5771113831089352, + "k": 1, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945 + ] + }, + { + "test_f1": 0.5515695740257144, + "test_acc": 0.5528205128205128, + "test_auc": 0.5778478136466153, + "val_auc": 0.6062831497348021, + "k": 2, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 1264 + ] + }, + { + "test_f1": 0.5650584929089602, + "test_acc": 0.5692307692307692, + "test_auc": 0.5935869492038913, + "val_auc": 0.605875152998776, + "k": 5, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 1264, + 238, + 6892, + 1132 + ] + }, + { + "test_f1": 0.5767413476289216, + "test_acc": 0.5774358974358974, + "test_auc": 0.6183811896186084, + "val_auc": 0.6536107711138311, + "k": 10, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 1264, + 238, + 6892, + 1132, + 1851, + 7893, + 470, + 1523, + 7073 + ] + }, + { + "test_f1": 0.5883663519500254, + "test_acc": 0.5887179487179487, + "test_auc": 0.6262980947251491, + "val_auc": 0.6634026927784578, + "k": 16, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 1264, + 238, + 6892, + 1132, + 1851, + 7893, + 470, + 1523, + 7073, + 4531, + 913, + 4110, + 386, + 538, + 1588 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..65fd136d599fcee5cf270155b28cbb664cb68006 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.4646361101175241, + "test_acc": 0.49641025641025643, + "test_auc": 0.5007952671087621, + "val_auc": 0.5467156262749898, + "k": 1, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341 + ] + }, + { + "test_f1": 0.4786821111169773, + "test_acc": 0.5005128205128205, + "test_auc": 0.4965117649038947, + "val_auc": 0.5234598123215014, + "k": 2, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 239 + ] + }, + { + "test_f1": 0.47533237814567736, + "test_acc": 0.48717948717948717, + "test_auc": 0.47897381088632307, + "val_auc": 0.5271317829457365, + "k": 5, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 239, + 538, + 827, + 1264 + ] + }, + { + "test_f1": 0.5108744314981026, + "test_acc": 0.5128205128205128, + "test_auc": 0.5110790386104285, + "val_auc": 0.5463076295389637, + "k": 10, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 239, + 538, + 827, + 1264, + 1409, + 456, + 1478, + 238, + 326 + ] + }, + { + "test_f1": 0.5292277978937331, + "test_acc": 0.5292307692307693, + "test_auc": 0.529576362473491, + "val_auc": 0.5964912280701755, + "k": 16, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 239, + 538, + 827, + 1264, + 1409, + 456, + 1478, + 238, + 326, + 400, + 8141, + 2671, + 347, + 1482, + 1444 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..24da9a6ba5e359b70378e660caab9eda12038bfd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5312721898574546, + "test_acc": 0.5312820512820513, + "test_auc": 0.5482377890732824, + "val_auc": 0.616483068135455, + "k": 1, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815 + ] + }, + { + "test_f1": 0.5451357575579855, + "test_acc": 0.5456410256410257, + "test_auc": 0.5589318174167704, + "val_auc": 0.6291309669522644, + "k": 2, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156 + ] + }, + { + "test_f1": 0.5316991346501465, + "test_acc": 0.5353846153846153, + "test_auc": 0.5431968896219745, + "val_auc": 0.627906976744186, + "k": 5, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 16809, + 12594, + 1770 + ] + }, + { + "test_f1": 0.5219440883583915, + "test_acc": 0.5261538461538462, + "test_auc": 0.5415769010670886, + "val_auc": 0.6348429212566299, + "k": 10, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 16809, + 12594, + 1770, + 6990, + 698, + 360, + 1409, + 4347 + ] + }, + { + "test_f1": 0.5634667843140098, + "test_acc": 0.5661538461538461, + "test_auc": 0.5684960615343185, + "val_auc": 0.6303549571603427, + "k": 16, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 16809, + 12594, + 1770, + 6990, + 698, + 360, + 1409, + 4347, + 24861, + 4110, + 1611, + 1684, + 191, + 538 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f368f75b2e3a738402c63d1f06f53a0cca2bd4d7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8473644735326978, + "test_acc": 0.8484848484848485, + "test_auc": 0.8828571428571429, + "val_auc": 0.8673878205128205, + "k": 1, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997 + ] + }, + { + "test_f1": 0.858151895385938, + "test_acc": 0.8585858585858586, + "test_auc": 0.8916326530612245, + "val_auc": 0.8968349358974358, + "k": 2, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335 + ] + }, + { + "test_f1": 0.8374557718282082, + "test_acc": 0.8383838383838383, + "test_auc": 0.8812244897959184, + "val_auc": 0.9190705128205128, + "k": 5, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 1156, + 1499 + ] + }, + { + "test_f1": 0.848175159876744, + "test_acc": 0.8484848484848485, + "test_auc": 0.8987755102040816, + "val_auc": 0.9270833333333334, + "k": 10, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 1156, + 1499, + 3842, + 314, + 3526, + 1734, + 1279 + ] + }, + { + "test_f1": 0.8080808080808081, + "test_acc": 0.8080808080808081, + "test_auc": 0.8828571428571427, + "val_auc": 0.9391025641025641, + "k": 16, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 1156, + 1499, + 3842, + 314, + 3526, + 1734, + 1279, + 7481, + 1473, + 851, + 5888, + 1158, + 364 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9d197971a5a0a5b4be4a56b2c59a44f3b549813c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8886617839096123, + "test_acc": 0.8888888888888888, + "test_auc": 0.9106122448979591, + "val_auc": 0.9471153846153846, + "k": 1, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8686064482736934, + "test_acc": 0.8686868686868687, + "test_auc": 0.9238775510204081, + "val_auc": 0.9643429487179487, + "k": 2, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335 + ] + }, + { + "test_f1": 0.8886617839096123, + "test_acc": 0.8888888888888888, + "test_auc": 0.936734693877551, + "val_auc": 0.9887820512820512, + "k": 5, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4889, + 1116 + ] + }, + { + "test_f1": 0.8886617839096123, + "test_acc": 0.8888888888888888, + "test_auc": 0.9614285714285714, + "val_auc": 0.9867788461538461, + "k": 10, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4889, + 1116, + 2832, + 4201, + 1924, + 7997, + 2900 + ] + }, + { + "test_f1": 0.8885475337088241, + "test_acc": 0.8888888888888888, + "test_auc": 0.9636734693877551, + "val_auc": 0.9899839743589745, + "k": 16, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4889, + 1116, + 2832, + 4201, + 1924, + 7997, + 2900, + 581, + 461, + 3526, + 2000, + 10707, + 925 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bd2837509a53d38430080cf2a6aab752588b5f9b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.836887392442948, + "test_acc": 0.8383838383838383, + "test_auc": 0.8620408163265306, + "val_auc": 0.8082932692307693, + "k": 1, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388 + ] + }, + { + "test_f1": 0.8467668125435367, + "test_acc": 0.8484848484848485, + "test_auc": 0.9032653061224489, + "val_auc": 0.8770032051282051, + "k": 2, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335 + ] + }, + { + "test_f1": 0.8178098497247434, + "test_acc": 0.8181818181818182, + "test_auc": 0.9326530612244898, + "val_auc": 0.9451121794871794, + "k": 5, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 2900, + 1596 + ] + }, + { + "test_f1": 0.8383838383838383, + "test_acc": 0.8383838383838383, + "test_auc": 0.9485714285714284, + "val_auc": 0.9441105769230769, + "k": 10, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 2900, + 1596, + 5888, + 7481, + 1847, + 1499, + 2000 + ] + }, + { + "test_f1": 0.8585858585858586, + "test_acc": 0.8585858585858586, + "test_auc": 0.9534693877551019, + "val_auc": 0.9501201923076923, + "k": 16, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 2900, + 1596, + 5888, + 7481, + 1847, + 1499, + 2000, + 1769, + 3825, + 5463, + 4201, + 2832, + 6747 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6e6a18ab75bcdf4690c3d7992dc0180abb96cfa7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6419277413444837, + "test_acc": 0.6627654963155614, + "test_auc": 0.6763446573703444, + "val_auc": 0.7802, + "k": 1, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.6527948767093337, + "test_acc": 0.6766363242306025, + "test_auc": 0.7640613515191326, + "val_auc": 0.871, + "k": 2, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388 + ] + }, + { + "test_f1": 0.7314465321230659, + "test_acc": 0.7321196358907672, + "test_auc": 0.825287735558358, + "val_auc": 0.9376, + "k": 5, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 5352, + 789, + 2335 + ] + }, + { + "test_f1": 0.7602549579778558, + "test_acc": 0.7607282184655396, + "test_auc": 0.8593086981290613, + "val_auc": 0.9412, + "k": 10, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 5352, + 789, + 2335, + 337, + 2000, + 2659, + 7481, + 27426 + ] + }, + { + "test_f1": 0.7906371124827244, + "test_acc": 0.7906371911573472, + "test_auc": 0.8776592146777076, + "val_auc": 0.9548000000000001, + "k": 16, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 5352, + 789, + 2335, + 337, + 2000, + 2659, + 7481, + 27426, + 239, + 7384, + 1596, + 1158, + 4889, + 1116 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..48fb85dfeb9b58d938daae2a32fc5b8d425a63e0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6329901647279108, + "test_acc": 0.6536627654963155, + "test_auc": 0.6705317001387383, + "val_auc": 0.7166, + "k": 1, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.696054716209701, + "test_acc": 0.6996098829648895, + "test_auc": 0.7437669195422686, + "val_auc": 0.7516, + "k": 2, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384 + ] + }, + { + "test_f1": 0.7368894291709555, + "test_acc": 0.741655830082358, + "test_auc": 0.8507777164837113, + "val_auc": 0.8745999999999999, + "k": 5, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 3526, + 4201, + 3842 + ] + }, + { + "test_f1": 0.7950451980855036, + "test_acc": 0.7962722149978326, + "test_auc": 0.9009997279345119, + "val_auc": 0.9148000000000001, + "k": 10, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 3526, + 4201, + 3842, + 7481, + 5388, + 2900, + 1924, + 1091 + ] + }, + { + "test_f1": 0.8082932648870724, + "test_acc": 0.8088426527958388, + "test_auc": 0.9078727635390158, + "val_auc": 0.9248, + "k": 16, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 3526, + 4201, + 3842, + 7481, + 5388, + 2900, + 1924, + 1091, + 6244, + 2832, + 5352, + 4037, + 1101, + 3825 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c3d7a439914b8e11c69f84001bba11ff366d8a2c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8248463554413563, + "test_acc": 0.8257477243172952, + "test_auc": 0.8558556459601281, + "val_auc": 0.881, + "k": 1, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8187141002131596, + "test_acc": 0.8214130905938448, + "test_auc": 0.8558556459601281, + "val_auc": 0.881, + "k": 2, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889 + ] + }, + { + "test_f1": 0.8365881383750466, + "test_acc": 0.8378846987429562, + "test_auc": 0.9183931301209565, + "val_auc": 0.9254, + "k": 5, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 5352, + 789, + 14154 + ] + }, + { + "test_f1": 0.846809859564628, + "test_acc": 0.8478543563068921, + "test_auc": 0.9254893796756558, + "val_auc": 0.9288, + "k": 10, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 5352, + 789, + 14154, + 12946, + 1924, + 1158, + 2335, + 1091 + ] + }, + { + "test_f1": 0.8221091065804056, + "test_acc": 0.8244473342002601, + "test_auc": 0.9185847784620335, + "val_auc": 0.9528000000000001, + "k": 16, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 5352, + 789, + 14154, + 12946, + 1924, + 1158, + 2335, + 1091, + 4531, + 4201, + 1851, + 737, + 2900, + 3526 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..45f6b36921b3fc75b9fb2fada55ddbaba535fdca --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9435804851195586, + "test_acc": 0.9435897435897436, + "test_auc": 0.9809619786582287, + "val_auc": 0.985720114239086, + "k": 1, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958 + ] + }, + { + "test_f1": 0.9435648102237594, + "test_acc": 0.9435897435897436, + "test_auc": 0.9809619786582287, + "val_auc": 0.985720114239086, + "k": 2, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818 + ] + }, + { + "test_f1": 0.9733333333333334, + "test_acc": 0.9733333333333334, + "test_auc": 0.9924428585855185, + "val_auc": 0.996328029375765, + "k": 5, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 2780 + ] + }, + { + "test_f1": 0.9794869205457708, + "test_acc": 0.9794871794871794, + "test_auc": 0.9948286599118052, + "val_auc": 0.9975520195838433, + "k": 10, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 2780, + 4586, + 5463, + 1279, + 1499, + 1596 + ] + }, + { + "test_f1": 0.9794869205457708, + "test_acc": 0.9794871794871794, + "test_auc": 0.9953483185781129, + "val_auc": 0.9979600163198694, + "k": 16, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 2780, + 4586, + 5463, + 1279, + 1499, + 1596, + 730, + 386, + 1158, + 779, + 851, + 3475 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1d40fc6c580fe9939b245fb4d1c9a2f794299ee8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9341299108923368, + "test_acc": 0.9343589743589743, + "test_auc": 0.9366437203352745, + "val_auc": 0.9883720930232558, + "k": 1, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065 + ] + }, + { + "test_f1": 0.9351348940914158, + "test_acc": 0.9353846153846154, + "test_auc": 0.9797396236577237, + "val_auc": 0.9971440228478172, + "k": 2, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508 + ] + }, + { + "test_f1": 0.9351669323216011, + "test_acc": 0.9353846153846154, + "test_auc": 0.987595516208301, + "val_auc": 0.9975520195838433, + "k": 5, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 3475 + ] + }, + { + "test_f1": 0.9487178408266029, + "test_acc": 0.9487179487179487, + "test_auc": 0.9935999932675801, + "val_auc": 0.9995920032639738, + "k": 10, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 3475, + 1279, + 992, + 581, + 1499, + 2708 + ] + }, + { + "test_f1": 0.9753844600197702, + "test_acc": 0.9753846153846154, + "test_auc": 0.9973154475376175, + "val_auc": 1.0, + "k": 16, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 3475, + 1279, + 992, + 581, + 1499, + 2708, + 3958, + 778, + 5888, + 343, + 1079, + 1482 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6ad9545dc3c6ba5c5ade609aa9890a1484f3793c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.965111249707749, + "test_acc": 0.9651282051282051, + "test_auc": 0.9851444945635709, + "val_auc": 1.0, + "k": 1, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037 + ] + }, + { + "test_f1": 0.9661388861342556, + "test_acc": 0.9661538461538461, + "test_auc": 0.9874945299087757, + "val_auc": 1.0, + "k": 2, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508 + ] + }, + { + "test_f1": 0.9764084696111573, + "test_acc": 0.9764102564102564, + "test_auc": 0.9902842764331639, + "val_auc": 1.0, + "k": 5, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 26034, + 1253 + ] + }, + { + "test_f1": 0.977433761007462, + "test_acc": 0.9774358974358974, + "test_auc": 0.9956575891204094, + "val_auc": 1.0, + "k": 10, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 26034, + 1253, + 5472, + 1279, + 493, + 1079, + 778 + ] + }, + { + "test_f1": 0.983588776834472, + "test_acc": 0.9835897435897436, + "test_auc": 0.9980938835964588, + "val_auc": 1.0, + "k": 16, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 26034, + 1253, + 5472, + 1279, + 493, + 1079, + 778, + 6396, + 582, + 581, + 3958, + 343, + 12813 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..644b4fdf4fa45f27b57eb3b0adfbe0e0ab378d71 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5653122417038261, + "test_acc": 0.58864325964456, + "test_auc": 0.6011354600537216, + "val_auc": 0.5826, + "k": 1, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894 + ] + }, + { + "test_f1": 0.6363869536306507, + "test_acc": 0.6423927178153446, + "test_auc": 0.6711900685574969, + "val_auc": 0.6612, + "k": 2, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351 + ] + }, + { + "test_f1": 0.6599792067687549, + "test_acc": 0.6631989596879063, + "test_auc": 0.7160804983157492, + "val_auc": 0.7352, + "k": 5, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351, + 1441, + 559, + 654 + ] + }, + { + "test_f1": 0.6923872191223165, + "test_acc": 0.6926744690073688, + "test_auc": 0.767993524540758, + "val_auc": 0.8104, + "k": 10, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351, + 1441, + 559, + 654, + 4510, + 3078, + 1960, + 619, + 343 + ] + }, + { + "test_f1": 0.7111804424405184, + "test_acc": 0.7113133940182055, + "test_auc": 0.7885923391769794, + "val_auc": 0.8384, + "k": 16, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351, + 1441, + 559, + 654, + 4510, + 3078, + 1960, + 619, + 343, + 1205, + 1602, + 698, + 121, + 3274, + 3301 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d87ff2d40783592d9257bf8492c6bd965d9e2e2e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.610643164663129, + "test_acc": 0.635890767230169, + "test_auc": 0.6418761395560673, + "val_auc": 0.6567999999999998, + "k": 1, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274 + ] + }, + { + "test_f1": 0.6777072612411335, + "test_acc": 0.6887732986562636, + "test_auc": 0.7029676181944171, + "val_auc": 0.7492000000000001, + "k": 2, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 4510 + ] + }, + { + "test_f1": 0.7134212500273137, + "test_acc": 0.7152145643693107, + "test_auc": 0.7575806313422448, + "val_auc": 0.7856000000000001, + "k": 5, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 4510, + 121, + 567, + 698 + ] + }, + { + "test_f1": 0.7457978890542092, + "test_acc": 0.7459904638058084, + "test_auc": 0.8215227851088487, + "val_auc": 0.8608, + "k": 10, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 4510, + 121, + 567, + 698, + 1602, + 1524, + 3078, + 52, + 19018 + ] + }, + { + "test_f1": 0.7641953937792689, + "test_acc": 0.7641959254442999, + "test_auc": 0.8457287221489866, + "val_auc": 0.8752, + "k": 16, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 4510, + 121, + 567, + 698, + 1602, + 1524, + 3078, + 52, + 19018, + 1894, + 1960, + 1158, + 1895, + 2046, + 831 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2bda88e56b345909bd82faec6bb4d69203192e14 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5493346779297936, + "test_acc": 0.5799739921976593, + "test_auc": 0.5851828024549024, + "val_auc": 0.545, + "k": 1, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.5524554176031351, + "test_acc": 0.5903771131339401, + "test_auc": 0.6571550217126297, + "val_auc": 0.6676, + "k": 2, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508 + ] + }, + { + "test_f1": 0.6324751734390397, + "test_acc": 0.6397919375812744, + "test_auc": 0.7083085944134885, + "val_auc": 0.7156, + "k": 5, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121 + ] + }, + { + "test_f1": 0.6472698415537319, + "test_acc": 0.6514954486345904, + "test_auc": 0.7335268856317856, + "val_auc": 0.7696000000000001, + "k": 10, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 7893, + 567, + 51, + 4510, + 925 + ] + }, + { + "test_f1": 0.674882401244196, + "test_acc": 0.6757693974859125, + "test_auc": 0.7434925993677859, + "val_auc": 0.7771999999999999, + "k": 16, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 7893, + 567, + 51, + 4510, + 925, + 1441, + 52, + 628, + 3274, + 4654, + 1253 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5f1c39675749d9ccf65fdca224ddfa61026de227 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6247576894668192, + "test_acc": 0.6375266524520256, + "test_auc": 0.652282233133297, + "val_auc": 0.5955999999999999, + "k": 1, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4672 + ] + }, + { + "test_f1": 0.6076009779759937, + "test_acc": 0.6076759061833689, + "test_auc": 0.6855428259683579, + "val_auc": 0.6234000000000001, + "k": 2, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4672, + 1499 + ] + }, + { + "test_f1": 0.6947628796104721, + "test_acc": 0.6950959488272921, + "test_auc": 0.7651027459538097, + "val_auc": 0.7592000000000001, + "k": 5, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4672, + 1499, + 1048, + 927, + 1504 + ] + }, + { + "test_f1": 0.7137390670553936, + "test_acc": 0.7142857142857143, + "test_auc": 0.7523731587561374, + "val_auc": 0.782, + "k": 10, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4672, + 1499, + 1048, + 927, + 1504, + 1818, + 1927, + 696, + 246, + 499 + ] + }, + { + "test_f1": 0.7289152846890207, + "test_acc": 0.7292110874200426, + "test_auc": 0.7943262411347518, + "val_auc": 0.8164, + "k": 16, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4672, + 1499, + 1048, + 927, + 1504, + 1818, + 1927, + 696, + 246, + 499, + 1902, + 437, + 29085, + 7129, + 6611, + 1371 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e4e9e804ce1669e7888072e601a23a74ebf13899 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9393815708101421, + "test_acc": 0.9393939393939394, + "test_auc": 0.9718367346938775, + "val_auc": 0.9524959742351048, + "k": 1, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299 + ] + }, + { + "test_f1": 0.9393815708101421, + "test_acc": 0.9393939393939394, + "test_auc": 0.9718367346938775, + "val_auc": 0.9524959742351048, + "k": 2, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114 + ] + }, + { + "test_f1": 0.9595959595959596, + "test_acc": 0.9595959595959596, + "test_auc": 0.9746938775510203, + "val_auc": 0.980877616747182, + "k": 5, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 23633, + 6877, + 432 + ] + }, + { + "test_f1": 1.0, + "test_acc": 1.0, + "test_auc": 1.0, + "val_auc": 0.9955716586151369, + "k": 10, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 23633, + 6877, + 432, + 13324, + 6055, + 2535, + 1915, + 1610 + ] + }, + { + "test_f1": 1.0, + "test_acc": 1.0, + "test_auc": 1.0, + "val_auc": 1.0, + "k": 16, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 23633, + 6877, + 432, + 13324, + 6055, + 2535, + 1915, + 1610, + 1500, + 29410, + 1232, + 15752, + 27804, + 1744 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d562541ad6a6c679551a2a28fde2175c262ffe8c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8282477732309314, + "test_acc": 0.8282828282828283, + "test_auc": 0.8685714285714285, + "val_auc": 0.8283011272141707, + "k": 1, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791 + ] + }, + { + "test_f1": 0.8686600618824769, + "test_acc": 0.8686868686868687, + "test_auc": 0.9310204081632654, + "val_auc": 0.9078099838969405, + "k": 2, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 1657 + ] + }, + { + "test_f1": 0.9191919191919192, + "test_acc": 0.9191919191919192, + "test_auc": 0.9718367346938777, + "val_auc": 0.9516908212560387, + "k": 5, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 1657, + 429, + 114, + 25266 + ] + }, + { + "test_f1": 0.9191754277468563, + "test_acc": 0.9191919191919192, + "test_auc": 0.9795918367346937, + "val_auc": 0.9522946859903382, + "k": 10, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 1657, + 429, + 114, + 25266, + 278, + 1396, + 4580, + 19000, + 7508 + ] + }, + { + "test_f1": 0.9292784948597953, + "test_acc": 0.9292929292929293, + "test_auc": 0.9877551020408163, + "val_auc": 0.9758454106280193, + "k": 16, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 1657, + 429, + 114, + 25266, + 278, + 1396, + 4580, + 19000, + 7508, + 521, + 1299, + 1411, + 400, + 1610, + 16 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d233d2b2210dbd114c110432c330482f0edb7826 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.92920621648266, + "test_acc": 0.9292929292929293, + "test_auc": 0.9510204081632653, + "val_auc": 0.9410225442834138, + "k": 1, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.979793856936714, + "test_acc": 0.9797979797979798, + "test_auc": 0.9983673469387755, + "val_auc": 0.9822866344605475, + "k": 2, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610 + ] + }, + { + "test_f1": 0.979793856936714, + "test_acc": 0.9797979797979798, + "test_auc": 0.9983673469387755, + "val_auc": 0.9863123993558776, + "k": 5, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610, + 1244, + 7577, + 177 + ] + }, + { + "test_f1": 0.979793856936714, + "test_acc": 0.9797979797979798, + "test_auc": 0.9975510204081632, + "val_auc": 0.9917471819645733, + "k": 10, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610, + 1244, + 7577, + 177, + 1501, + 1550, + 419, + 907, + 114 + ] + }, + { + "test_f1": 0.9797979797979798, + "test_acc": 0.9797979797979798, + "test_auc": 0.9983673469387755, + "val_auc": 0.9963768115942029, + "k": 16, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610, + 1244, + 7577, + 177, + 1501, + 1550, + 419, + 907, + 114, + 23791, + 1960, + 3644, + 27804, + 635, + 15752 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..222cdb59d3aa72b10d672d73535ad05a182d73a8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9199683794466404, + "test_acc": 0.92, + "test_auc": 0.9406605562579013, + "val_auc": 0.9147727272727273, + "k": 1, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752 + ] + }, + { + "test_f1": 0.92, + "test_acc": 0.92, + "test_auc": 0.9512879266750949, + "val_auc": 0.9172077922077921, + "k": 2, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309 + ] + }, + { + "test_f1": 0.9377630261840788, + "test_acc": 0.9377777777777778, + "test_auc": 0.9648783185840709, + "val_auc": 0.9788961038961039, + "k": 5, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580 + ] + }, + { + "test_f1": 0.9333333333333333, + "test_acc": 0.9333333333333333, + "test_auc": 0.968552465233881, + "val_auc": 0.9825487012987012, + "k": 10, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 3301, + 26752, + 1915, + 969, + 1295 + ] + }, + { + "test_f1": 0.9511111111111111, + "test_acc": 0.9511111111111111, + "test_auc": 0.9715549936788875, + "val_auc": 0.9878246753246753, + "k": 16, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 3301, + 26752, + 1915, + 969, + 1295, + 1327, + 5563, + 4332, + 1541, + 1960, + 1073 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7ca3f79c567c0548aba44d08eeea89a6bed59d18 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8039333298132602, + "test_acc": 0.8044444444444444, + "test_auc": 0.8185840707964602, + "val_auc": 0.8486201298701299, + "k": 1, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406 + ] + }, + { + "test_f1": 0.8167937510064612, + "test_acc": 0.8177777777777778, + "test_auc": 0.8591577117572694, + "val_auc": 0.8575487012987013, + "k": 2, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854 + ] + }, + { + "test_f1": 0.8344374863322797, + "test_acc": 0.8355555555555556, + "test_auc": 0.8974399494310998, + "val_auc": 0.8754058441558441, + "k": 5, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1182, + 1804 + ] + }, + { + "test_f1": 0.8439637355308598, + "test_acc": 0.8444444444444444, + "test_auc": 0.9260429835651075, + "val_auc": 0.8867694805194806, + "k": 10, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1182, + 1804, + 1232, + 311, + 1570, + 6662, + 322 + ] + }, + { + "test_f1": 0.8487215719357447, + "test_acc": 0.8488888888888889, + "test_auc": 0.9305072692793932, + "val_auc": 0.8896103896103896, + "k": 16, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1182, + 1804, + 1232, + 311, + 1570, + 6662, + 322, + 15752, + 1420, + 800, + 920, + 748, + 1396 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..da5414333ee813927bdf0ff96ab5c0a3ce911688 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9199810336652442, + "test_acc": 0.92, + "test_auc": 0.9529077117572694, + "val_auc": 0.9370941558441558, + "k": 1, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9243817167104839, + "test_acc": 0.9244444444444444, + "test_auc": 0.9663005689001264, + "val_auc": 0.9573863636363635, + "k": 2, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16 + ] + }, + { + "test_f1": 0.9510917908760189, + "test_acc": 0.9511111111111111, + "test_auc": 0.9811551833122629, + "val_auc": 0.9650974025974026, + "k": 5, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 21467, + 114 + ] + }, + { + "test_f1": 0.9421742539550759, + "test_acc": 0.9422222222222222, + "test_auc": 0.9917035398230087, + "val_auc": 0.9809253246753247, + "k": 10, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 21467, + 114, + 15752, + 1835, + 1960, + 969, + 6718 + ] + }, + { + "test_f1": 0.9466350461133068, + "test_acc": 0.9466666666666667, + "test_auc": 0.9911504424778761, + "val_auc": 0.981737012987013, + "k": 16, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 21467, + 114, + 15752, + 1835, + 1960, + 969, + 6718, + 7033, + 4580, + 1232, + 1417, + 1596, + 3481 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a294b2790bbac3aaf12782a9ace6bcf473331f76 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6883350535621909, + "test_acc": 0.6931079323797139, + "test_auc": 0.7093626602894115, + "val_auc": 0.7676000000000001, + "k": 1, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394 + ] + }, + { + "test_f1": 0.6986304385327713, + "test_acc": 0.706111833550065, + "test_auc": 0.7628505849407995, + "val_auc": 0.821, + "k": 2, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 7649 + ] + }, + { + "test_f1": 0.6986304385327713, + "test_acc": 0.706111833550065, + "test_auc": 0.7628468271301901, + "val_auc": 0.821, + "k": 5, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 7649, + 7614, + 943, + 7246 + ] + }, + { + "test_f1": 0.7339607422842214, + "test_acc": 0.7351538794971825, + "test_auc": 0.7968854514107573, + "val_auc": 0.8737999999999999, + "k": 10, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 7649, + 7614, + 943, + 7246, + 1928, + 1560, + 1737, + 28053, + 17225 + ] + }, + { + "test_f1": 0.7456076089302879, + "test_acc": 0.7459904638058084, + "test_auc": 0.8196168235677856, + "val_auc": 0.8696, + "k": 16, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 7649, + 7614, + 943, + 7246, + 1928, + 1560, + 1737, + 28053, + 17225, + 811, + 584, + 28397, + 253, + 1551, + 1848 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..83f33197e27d8a244b7cfe4867ff4c1cdab182ef --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6143231932687245, + "test_acc": 0.6150845253576073, + "test_auc": 0.6367568741629477, + "val_auc": 0.6511999999999999, + "k": 1, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943 + ] + }, + { + "test_f1": 0.6298088950555678, + "test_acc": 0.6306892067620286, + "test_auc": 0.6701232261255019, + "val_auc": 0.7108, + "k": 2, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253 + ] + }, + { + "test_f1": 0.6420247165899957, + "test_acc": 0.6432596445600347, + "test_auc": 0.7048844773862473, + "val_auc": 0.7496, + "k": 5, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 5394, + 1011, + 32305 + ] + }, + { + "test_f1": 0.6341538610760423, + "test_acc": 0.635023840485479, + "test_auc": 0.7067246772416468, + "val_auc": 0.7532, + "k": 10, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 5394, + 1011, + 32305, + 7614, + 7649, + 1551, + 1928, + 1875 + ] + }, + { + "test_f1": 0.6765265590252721, + "test_acc": 0.6766363242306025, + "test_auc": 0.751401287576227, + "val_auc": 0.7912, + "k": 16, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 5394, + 1011, + 32305, + 7614, + 7649, + 1551, + 1928, + 1875, + 1411, + 1749, + 1848, + 844, + 613, + 1417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9ea74736fafb3f8e5f17d573c4ee8ae2488e6828 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6089529204980799, + "test_acc": 0.6228868660598179, + "test_auc": 0.6413538038813674, + "val_auc": 0.6924000000000001, + "k": 1, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253 + ] + }, + { + "test_f1": 0.6175877610466614, + "test_acc": 0.6332899869960988, + "test_auc": 0.7097572304033934, + "val_auc": 0.7288, + "k": 2, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394 + ] + }, + { + "test_f1": 0.7192991337629314, + "test_acc": 0.7221499783268314, + "test_auc": 0.8017533192741113, + "val_auc": 0.7938, + "k": 5, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 7649, + 1099, + 844 + ] + }, + { + "test_f1": 0.7170841409388153, + "test_acc": 0.7191157347204161, + "test_auc": 0.8070281580264581, + "val_auc": 0.7788, + "k": 10, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 7649, + 1099, + 844, + 1541, + 794, + 9605, + 7246, + 1960 + ] + }, + { + "test_f1": 0.727445343461045, + "test_acc": 0.7286519289120069, + "test_auc": 0.8212214086979788, + "val_auc": 0.7976, + "k": 16, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 7649, + 1099, + 844, + 1541, + 794, + 9605, + 7246, + 1960, + 100, + 7614, + 1737, + 1749, + 4675, + 17225 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6441580d6ea59c20be8cb34062351e3291b96202 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8106110766404607, + "test_acc": 0.8128813559322033, + "test_auc": 0.8301563137748066, + "val_auc": 0.8011272141706925, + "k": 1, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.8125266348204845, + "test_acc": 0.8149152542372882, + "test_auc": 0.8797181865984196, + "val_auc": 0.8615136876006442, + "k": 2, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 266 + ] + }, + { + "test_f1": 0.8114912331960104, + "test_acc": 0.8135593220338984, + "test_auc": 0.9075511209657551, + "val_auc": 0.891304347826087, + "k": 5, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 266, + 1847, + 5020, + 508 + ] + }, + { + "test_f1": 0.8370048006616643, + "test_acc": 0.8372881355932204, + "test_auc": 0.9182616113813785, + "val_auc": 0.9303542673107891, + "k": 10, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 266, + 1847, + 5020, + 508, + 4413, + 6838, + 32278, + 776, + 1697 + ] + }, + { + "test_f1": 0.8642660342606496, + "test_acc": 0.864406779661017, + "test_auc": 0.938940736083073, + "val_auc": 0.9367954911433172, + "k": 16, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 266, + 1847, + 5020, + 508, + 4413, + 6838, + 32278, + 776, + 1697, + 1433, + 1528, + 1585, + 8604, + 4674, + 1180 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e1e5a1c55f54cfe117e2173a2f21541c74c9d56a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7913541289724214, + "test_acc": 0.7959322033898305, + "test_auc": 0.8080458755740882, + "val_auc": 0.7286634460547504, + "k": 1, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528 + ] + }, + { + "test_f1": 0.8016106335262307, + "test_acc": 0.8054237288135593, + "test_auc": 0.8169867587414001, + "val_auc": 0.7783816425120772, + "k": 2, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791 + ] + }, + { + "test_f1": 0.8031171020950075, + "test_acc": 0.8040677966101695, + "test_auc": 0.8782684875695433, + "val_auc": 0.8433977455716586, + "k": 5, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1815, + 1628, + 20997 + ] + }, + { + "test_f1": 0.8199506969599957, + "test_acc": 0.8203389830508474, + "test_auc": 0.9029997830507477, + "val_auc": 0.8695652173913043, + "k": 10, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1815, + 1628, + 20997, + 16223, + 1103, + 1866, + 1789, + 266 + ] + }, + { + "test_f1": 0.825960243247074, + "test_acc": 0.8264406779661017, + "test_auc": 0.9104780973182867, + "val_auc": 0.85829307568438, + "k": 16, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1815, + 1628, + 20997, + 16223, + 1103, + 1866, + 1789, + 266, + 4558, + 203, + 31149, + 2023, + 1180, + 416 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3fdc72fc86469230863689cc5ecc8a93ec743e9d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6525083228285771, + "test_acc": 0.6583050847457627, + "test_auc": 0.6873228094560457, + "val_auc": 0.7232286634460547, + "k": 1, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020 + ] + }, + { + "test_f1": 0.6975580915375061, + "test_acc": 0.7016949152542373, + "test_auc": 0.7754648781223226, + "val_auc": 0.822061191626409, + "k": 2, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266 + ] + }, + { + "test_f1": 0.7368280408743263, + "test_acc": 0.7383050847457627, + "test_auc": 0.8070309943262255, + "val_auc": 0.8858695652173914, + "k": 5, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 431, + 373 + ] + }, + { + "test_f1": 0.784756923503518, + "test_acc": 0.7871186440677966, + "test_auc": 0.8644287799730102, + "val_auc": 0.9043880837359097, + "k": 10, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 431, + 373, + 27673, + 32278, + 1628, + 895, + 16223 + ] + }, + { + "test_f1": 0.8119007057165939, + "test_acc": 0.8128813559322033, + "test_auc": 0.8949579522932272, + "val_auc": 0.8977455716586151, + "k": 16, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 431, + 373, + 27673, + 32278, + 1628, + 895, + 16223, + 8604, + 461, + 5804, + 582, + 4413, + 960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..07361286709bb366e5abd8f74f0fd964b4b10a26 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5447398976201777, + "test_acc": 0.5743440233236151, + "test_auc": 0.5725554195566437, + "val_auc": 0.6250500200080032, + "k": 1, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639 + ] + }, + { + "test_f1": 0.5936170796128573, + "test_acc": 0.6034985422740525, + "test_auc": 0.604311165510676, + "val_auc": 0.5946378551420568, + "k": 2, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639, + 4381 + ] + }, + { + "test_f1": 0.5895839299779732, + "test_acc": 0.5947521865889213, + "test_auc": 0.6228920168638652, + "val_auc": 0.6226490596238494, + "k": 5, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639, + 4381, + 219, + 1122, + 799 + ] + }, + { + "test_f1": 0.6442785867443592, + "test_acc": 0.6443148688046647, + "test_auc": 0.681031551747586, + "val_auc": 0.7779111644657863, + "k": 10, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639, + 4381, + 219, + 1122, + 799, + 993, + 1697, + 1799, + 275, + 5897 + ] + }, + { + "test_f1": 0.6498779541367835, + "test_acc": 0.6501457725947521, + "test_auc": 0.7087243302053583, + "val_auc": 0.7887154861944778, + "k": 16, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26639, + 4381, + 219, + 1122, + 799, + 993, + 1697, + 1799, + 275, + 5897, + 801, + 304, + 1508, + 673, + 1411, + 7774 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7b6bb280f37014b49e64b809c153d5050ab2613c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7338643056665477, + "test_acc": 0.749271137026239, + "test_auc": 0.7503399972800219, + "val_auc": 0.7316926770708283, + "k": 1, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7842304634141368, + "test_acc": 0.7900874635568513, + "test_auc": 0.8016795865633075, + "val_auc": 0.844937975190076, + "k": 2, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275 + ] + }, + { + "test_f1": 0.8447524916488901, + "test_acc": 0.8454810495626822, + "test_auc": 0.8876308989528084, + "val_auc": 0.8743497398959584, + "k": 5, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 433, + 8006, + 191 + ] + }, + { + "test_f1": 0.8626163773475664, + "test_acc": 0.8629737609329446, + "test_auc": 0.9078607371141032, + "val_auc": 0.9109643857543017, + "k": 10, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 433, + 8006, + 191, + 534, + 26639, + 1928, + 1122, + 799 + ] + }, + { + "test_f1": 0.9182589481144858, + "test_acc": 0.9183673469387755, + "test_auc": 0.9604583163334693, + "val_auc": 0.9423769507803121, + "k": 16, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 433, + 8006, + 191, + 534, + 26639, + 1928, + 1122, + 799, + 506, + 801, + 142, + 1799, + 1239, + 1960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..708ca6fb6a271f685776178af13ec9b0921039f6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6799713641145131, + "test_acc": 0.6880466472303207, + "test_auc": 0.6875254997960016, + "val_auc": 0.6906762705082033, + "k": 1, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.6964524600353283, + "test_acc": 0.6967930029154519, + "test_auc": 0.7455290357677139, + "val_auc": 0.7470988395358142, + "k": 2, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191 + ] + }, + { + "test_f1": 0.6966486186311259, + "test_acc": 0.6967930029154519, + "test_auc": 0.8069665442676459, + "val_auc": 0.7899159663865545, + "k": 5, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 8006, + 433, + 534 + ] + }, + { + "test_f1": 0.7959183673469388, + "test_acc": 0.7959183673469388, + "test_auc": 0.8724330205358357, + "val_auc": 0.8695478191276511, + "k": 10, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 8006, + 433, + 534, + 1239, + 275, + 5897, + 142, + 635 + ] + }, + { + "test_f1": 0.8046614022742482, + "test_acc": 0.8046647230320699, + "test_auc": 0.8675880592955256, + "val_auc": 0.8647458983593437, + "k": 16, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 8006, + 433, + 534, + 1239, + 275, + 5897, + 142, + 635, + 1464, + 1195, + 1596, + 3233, + 14404, + 181 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d8803e02580ff10119a496d98a0be8512e9dd263 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320 + ] + }, + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 2, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 5, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 1073, + 6534, + 987 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 10, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 1073, + 6534, + 987, + 1030, + 1417, + 5883, + 1939, + 1224 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 16, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 1073, + 6534, + 987, + 1030, + 1417, + 5883, + 1939, + 1224, + 1703, + 13612, + 31245, + 859, + 4264, + 6804 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c8a089fdb78a1c475610a2ecfd22ea4ef9e3a611 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5754976232917409, + "test_acc": 0.5757575757575758, + "test_auc": 0.5869387755102041, + "val_auc": 0.5672727272727273, + "k": 1, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520 + ] + }, + { + "test_f1": 0.5913682277318641, + "test_acc": 0.5959595959595959, + "test_auc": 0.5575510204081633, + "val_auc": 0.5793939393939394, + "k": 2, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 5883 + ] + }, + { + "test_f1": 0.5162567896560609, + "test_acc": 0.5252525252525253, + "test_auc": 0.4722448979591837, + "val_auc": 0.5604040404040405, + "k": 5, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 5883, + 5129, + 1818, + 6534 + ] + }, + { + "test_f1": 0.6336700336700336, + "test_acc": 0.6363636363636364, + "test_auc": 0.6297959183673469, + "val_auc": 0.5636363636363636, + "k": 10, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 5883, + 5129, + 1818, + 6534, + 2046, + 1788, + 7026, + 4932, + 1030 + ] + }, + { + "test_f1": 0.5141595992659822, + "test_acc": 0.5151515151515151, + "test_auc": 0.5577551020408164, + "val_auc": 0.5927272727272727, + "k": 16, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 5883, + 5129, + 1818, + 6534, + 2046, + 1788, + 7026, + 4932, + 1030, + 320, + 1707, + 31245, + 987, + 7666, + 3648 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..56ef1a3f0ccafdef883152ab0a29438afa0b4c46 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6160832817975676, + "test_acc": 0.6161616161616161, + "test_auc": 0.6118367346938776, + "val_auc": 0.585050505050505, + "k": 1, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666 + ] + }, + { + "test_f1": 0.5857740413216579, + "test_acc": 0.5858585858585859, + "test_auc": 0.6161224489795918, + "val_auc": 0.6048484848484849, + "k": 2, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245 + ] + }, + { + "test_f1": 0.5353535353535354, + "test_acc": 0.5353535353535354, + "test_auc": 0.5983673469387755, + "val_auc": 0.6464646464646465, + "k": 5, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 7026 + ] + }, + { + "test_f1": 0.5136636413232157, + "test_acc": 0.5151515151515151, + "test_auc": 0.5751020408163265, + "val_auc": 0.682020202020202, + "k": 10, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 7026, + 27174, + 1116, + 111, + 1011, + 373 + ] + }, + { + "test_f1": 0.5352587095444239, + "test_acc": 0.5353535353535354, + "test_auc": 0.573469387755102, + "val_auc": 0.6428282828282829, + "k": 16, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 7026, + 27174, + 1116, + 111, + 1011, + 373, + 5883, + 2595, + 680, + 6027, + 1180, + 10244 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d9aeb02142f13e47998762cd7480207c62004636 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.4925985828283408, + "test_acc": 0.5457627118644067, + "test_auc": 0.5474107658308606, + "val_auc": 0.5986312399355878, + "k": 1, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217 + ] + }, + { + "test_f1": 0.5361824897469355, + "test_acc": 0.5430508474576271, + "test_auc": 0.5574373145359676, + "val_auc": 0.6694847020933977, + "k": 2, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1960 + ] + }, + { + "test_f1": 0.5736809775547959, + "test_acc": 0.5789830508474576, + "test_auc": 0.6083330575503856, + "val_auc": 0.6090982286634461, + "k": 5, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1960, + 539, + 1541, + 1319 + ] + }, + { + "test_f1": 0.5633969604030621, + "test_acc": 0.5654237288135593, + "test_auc": 0.599388497277103, + "val_auc": 0.6229871175523349, + "k": 10, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1960, + 539, + 1541, + 1319, + 311, + 186, + 20, + 373, + 836 + ] + }, + { + "test_f1": 0.5676809497773959, + "test_acc": 0.5688135593220339, + "test_auc": 0.6003390291704817, + "val_auc": 0.6215780998389694, + "k": 16, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1960, + 539, + 1541, + 1319, + 311, + 186, + 20, + 373, + 836, + 1616, + 1812, + 1255, + 443, + 26270, + 253 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..29a7316f8988813f63628f6cb6670e7b68945ee6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6189672885712528, + "test_acc": 0.6189830508474576, + "test_auc": 0.6454286218574533, + "val_auc": 0.6549919484702094, + "k": 1, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319 + ] + }, + { + "test_f1": 0.6508471366589516, + "test_acc": 0.6508474576271186, + "test_auc": 0.7031380422352391, + "val_auc": 0.6829710144927537, + "k": 2, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541 + ] + }, + { + "test_f1": 0.6657613104266966, + "test_acc": 0.6664406779661017, + "test_auc": 0.721196309656448, + "val_auc": 0.7077294685990339, + "k": 5, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1874, + 1960 + ] + }, + { + "test_f1": 0.6606500803274267, + "test_acc": 0.6610169491525424, + "test_auc": 0.7221082319371361, + "val_auc": 0.75, + "k": 10, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1874, + 1960, + 186, + 1255, + 1335, + 908, + 100 + ] + }, + { + "test_f1": 0.6695241305989862, + "test_acc": 0.6705084745762712, + "test_auc": 0.7331928678852596, + "val_auc": 0.7493961352657005, + "k": 16, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1874, + 1960, + 186, + 1255, + 1335, + 908, + 100, + 311, + 253, + 765, + 1500, + 217, + 3780 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7b98f3fd69b8e73184d3ed92d17414ea240eb41d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5894457071769821, + "test_acc": 0.5979661016949153, + "test_auc": 0.6060422205307534, + "val_auc": 0.6302334943639292, + "k": 1, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.5837865539913408, + "test_acc": 0.5945762711864406, + "test_auc": 0.6060422205307534, + "val_auc": 0.6302334943639292, + "k": 2, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255 + ] + }, + { + "test_f1": 0.6155652877281658, + "test_acc": 0.6169491525423729, + "test_auc": 0.6436884314569062, + "val_auc": 0.678743961352657, + "k": 5, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 12485, + 186 + ] + }, + { + "test_f1": 0.6175751345624335, + "test_acc": 0.6189830508474576, + "test_auc": 0.6450636690898794, + "val_auc": 0.6968599033816425, + "k": 10, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 12485, + 186, + 521, + 4511, + 1942, + 6212, + 1319 + ] + }, + { + "test_f1": 0.617887199354318, + "test_acc": 0.6196610169491525, + "test_auc": 0.6458533275970479, + "val_auc": 0.6700885668276972, + "k": 16, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 12485, + 186, + 521, + 4511, + 1942, + 6212, + 1319, + 249, + 1500, + 567, + 1596, + 311, + 908 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..adb44d1eb73bc2541344d53f659672ff19e43f24 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5780858386223925, + "test_acc": 0.5781637717121588, + "test_auc": 0.5882468843899316, + "val_auc": 0.5876350540216086, + "k": 1, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067 + ] + }, + { + "test_f1": 0.5719835845404257, + "test_acc": 0.5732009925558312, + "test_auc": 0.6323210679276883, + "val_auc": 0.611844737895158, + "k": 2, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268 + ] + }, + { + "test_f1": 0.6004224656952357, + "test_acc": 0.6004962779156328, + "test_auc": 0.6412245702182159, + "val_auc": 0.6080432172869147, + "k": 5, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 5731, + 1441, + 2 + ] + }, + { + "test_f1": 0.5904294892620233, + "test_acc": 0.5905707196029777, + "test_auc": 0.6272104822422541, + "val_auc": 0.6454581832733093, + "k": 10, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 5731, + 1441, + 2, + 124, + 8013, + 8006, + 1551, + 1619 + ] + }, + { + "test_f1": 0.5891741876053063, + "test_acc": 0.5905707196029777, + "test_auc": 0.6566917885818433, + "val_auc": 0.6990796318527411, + "k": 16, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 5731, + 1441, + 2, + 124, + 8013, + 8006, + 1551, + 1619, + 1674, + 1101, + 24630, + 1955, + 10, + 1070 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d5aa837f9732df15b0be5a5c45f587283d4f19eb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5669527254960918, + "test_acc": 0.5955334987593052, + "test_auc": 0.6061524062854047, + "val_auc": 0.6206482593037215, + "k": 1, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575 + ] + }, + { + "test_f1": 0.6138388578706881, + "test_acc": 0.6153846153846154, + "test_auc": 0.65957342002857, + "val_auc": 0.6520608243297319, + "k": 2, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711 + ] + }, + { + "test_f1": 0.6166715411532361, + "test_acc": 0.6178660049627791, + "test_auc": 0.6585882468843899, + "val_auc": 0.6598639455782312, + "k": 5, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 1625, + 583, + 1619 + ] + }, + { + "test_f1": 0.6020128136913694, + "test_acc": 0.6054590570719603, + "test_auc": 0.6542042263927885, + "val_auc": 0.6928771508603442, + "k": 10, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 1625, + 583, + 1619, + 124, + 1381, + 1551, + 1799, + 7044 + ] + }, + { + "test_f1": 0.6064864355488413, + "test_acc": 0.607940446650124, + "test_auc": 0.671543273730358, + "val_auc": 0.6866746698679471, + "k": 16, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 1625, + 583, + 1619, + 124, + 1381, + 1551, + 1799, + 7044, + 680, + 765, + 434, + 1653, + 1955, + 898 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0229aaa0cb0e1b6bc979e332920e1f92bf9c6562 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5450484233412745, + "test_acc": 0.5459057071960298, + "test_auc": 0.5746268656716417, + "val_auc": 0.5542216886754702, + "k": 1, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619 + ] + }, + { + "test_f1": 0.5679455620850131, + "test_acc": 0.5682382133995038, + "test_auc": 0.6114969705925817, + "val_auc": 0.6162464985994398, + "k": 2, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042 + ] + }, + { + "test_f1": 0.6625268615524935, + "test_acc": 0.6625310173697271, + "test_auc": 0.7262326978966553, + "val_auc": 0.6810724289715886, + "k": 5, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 32279, + 1067 + ] + }, + { + "test_f1": 0.6772762753548769, + "test_acc": 0.6774193548387096, + "test_auc": 0.740135953893897, + "val_auc": 0.7118847539015606, + "k": 10, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 32279, + 1067, + 1625, + 10, + 1662, + 772, + 2 + ] + }, + { + "test_f1": 0.687132993233975, + "test_acc": 0.6873449131513648, + "test_auc": 0.7413920496527264, + "val_auc": 0.7086834733893557, + "k": 16, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 32279, + 1067, + 1625, + 10, + 1662, + 772, + 2, + 124, + 347, + 1872, + 125, + 2311, + 517 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bde144cb829aa8ff77c9cb2adae9c03cb9b1c1db --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8384287229941335, + "test_acc": 0.8406961178045516, + "test_auc": 0.8519877851213602, + "val_auc": 0.844848484848485, + "k": 1, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731 + ] + }, + { + "test_f1": 0.8306982176127352, + "test_acc": 0.8326639892904953, + "test_auc": 0.8918151711086578, + "val_auc": 0.9026262626262626, + "k": 2, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226 + ] + }, + { + "test_f1": 0.8609110535917833, + "test_acc": 0.8621151271753681, + "test_auc": 0.9534845378560881, + "val_auc": 0.9581818181818182, + "k": 5, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 6567, + 737 + ] + }, + { + "test_f1": 0.8450654004340312, + "test_acc": 0.8473895582329317, + "test_auc": 0.9520078565181861, + "val_auc": 0.9563636363636364, + "k": 10, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 6567, + 737, + 1274, + 23697, + 1158, + 22330, + 292 + ] + }, + { + "test_f1": 0.868709744856377, + "test_acc": 0.8688085676037484, + "test_auc": 0.9562586916316611, + "val_auc": 0.9660606060606061, + "k": 16, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 6567, + 737, + 1274, + 23697, + 1158, + 22330, + 292, + 5872, + 805, + 13701, + 1753, + 2046, + 1499 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3ca2866ae6bac515df9b629e88a9bb13edfb4d72 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.879594202291041, + "test_acc": 0.8808567603748326, + "test_auc": 0.885610958982667, + "val_auc": 0.8973737373737374, + "k": 1, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226 + ] + }, + { + "test_f1": 0.8809138694564764, + "test_acc": 0.8821954484605087, + "test_auc": 0.9227251222204699, + "val_auc": 0.9325252525252524, + "k": 2, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731 + ] + }, + { + "test_f1": 0.8827132336334229, + "test_acc": 0.8835341365461847, + "test_auc": 0.9676563776863414, + "val_auc": 0.976969696969697, + "k": 5, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4505, + 4648, + 6567 + ] + }, + { + "test_f1": 0.8911445404928516, + "test_acc": 0.891566265060241, + "test_auc": 0.9665954609969749, + "val_auc": 0.972929292929293, + "k": 10, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4505, + 4648, + 6567, + 1065, + 1380, + 1425, + 32, + 1682 + ] + }, + { + "test_f1": 0.886731852686901, + "test_acc": 0.8875502008032129, + "test_auc": 0.9701653022895729, + "val_auc": 0.9765656565656565, + "k": 16, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4505, + 4648, + 6567, + 1065, + 1380, + 1425, + 32, + 1682, + 7481, + 10069, + 22775, + 805, + 22330, + 1759 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a92ac62c56b864fdb2b47dbaa3ed693cd4a06a42 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8797861903769895, + "test_acc": 0.8808567603748326, + "test_auc": 0.8896359908818512, + "val_auc": 0.868080808080808, + "k": 1, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648 + ] + }, + { + "test_f1": 0.881044004247915, + "test_acc": 0.8821954484605087, + "test_auc": 0.9200943355650815, + "val_auc": 0.8763636363636365, + "k": 2, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505 + ] + }, + { + "test_f1": 0.8968007996180875, + "test_acc": 0.8969210174029452, + "test_auc": 0.9753551920402574, + "val_auc": 0.9723232323232323, + "k": 5, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32 + ] + }, + { + "test_f1": 0.8861339680129727, + "test_acc": 0.8862115127175368, + "test_auc": 0.9731186649653768, + "val_auc": 0.9737373737373738, + "k": 10, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32, + 22848, + 1101, + 1682, + 1305, + 7481 + ] + }, + { + "test_f1": 0.8942091356584065, + "test_acc": 0.8942436412315931, + "test_auc": 0.9740828088486186, + "val_auc": 0.9717171717171716, + "k": 16, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32, + 22848, + 1101, + 1682, + 1305, + 7481, + 1380, + 1158, + 1065, + 23, + 1425, + 6828 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5ab48639c9b2d7e9e8453adcb06968b8f0a7ed4e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7613938541875879, + "test_acc": 0.7723076923076924, + "test_auc": 0.7740705052681186, + "val_auc": 0.7168502651978784, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518 + ] + }, + { + "test_f1": 0.7812346991084221, + "test_acc": 0.7876923076923077, + "test_auc": 0.7997862456660045, + "val_auc": 0.7478580171358629, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541 + ] + }, + { + "test_f1": 0.7783035937421902, + "test_acc": 0.7846153846153846, + "test_auc": 0.8696014407378733, + "val_auc": 0.8388412892696858, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 821 + ] + }, + { + "test_f1": 0.7990320209844364, + "test_acc": 0.8020512820512821, + "test_auc": 0.8892432759955566, + "val_auc": 0.8572011423908608, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 821, + 2632, + 374, + 1335, + 253, + 1030 + ] + }, + { + "test_f1": 0.80527973017663, + "test_acc": 0.8082051282051282, + "test_auc": 0.8918289460396539, + "val_auc": 0.8592411260709913, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 821, + 2632, + 374, + 1335, + 253, + 1030, + 302, + 7813, + 505, + 598, + 2581, + 2921 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3463f2229efbf302084d9b9f35192ae707630af3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8708714021257031, + "test_acc": 0.8722689075630252, + "test_auc": 0.8784524119429319, + "val_auc": 0.8647129666800482, + "k": 1, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852 + ] + }, + { + "test_f1": 0.8980738547030855, + "test_acc": 0.8984873949579832, + "test_auc": 0.9116055545914051, + "val_auc": 0.9502207948615015, + "k": 2, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782 + ] + }, + { + "test_f1": 0.9059021344056595, + "test_acc": 0.906218487394958, + "test_auc": 0.9618094724168599, + "val_auc": 0.9877559213167403, + "k": 5, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 60, + 1318, + 173 + ] + }, + { + "test_f1": 0.9122785497159637, + "test_acc": 0.9126050420168067, + "test_auc": 0.9763691238041522, + "val_auc": 0.9869530309112807, + "k": 10, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 60, + 1318, + 173, + 1275, + 8815, + 11077, + 23246, + 1283 + ] + }, + { + "test_f1": 0.9140886108405245, + "test_acc": 0.9142857142857143, + "test_auc": 0.9777412304488361, + "val_auc": 0.9855479727017262, + "k": 16, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 60, + 1318, + 173, + 1275, + 8815, + 11077, + 23246, + 1283, + 1521, + 1688, + 1152, + 6919, + 98, + 42 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..de506ad8f98bed0262b2274bdb224fc3e4949810 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7470706205369282, + "test_acc": 0.7536134453781512, + "test_auc": 0.756448810840908, + "val_auc": 0.7515054195102367, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021 + ] + }, + { + "test_f1": 0.7715468520134601, + "test_acc": 0.7721008403361345, + "test_auc": 0.8763350471108025, + "val_auc": 0.8839823364110799, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914 + ] + }, + { + "test_f1": 0.780632681949243, + "test_acc": 0.7818487394957984, + "test_auc": 0.8932735138223022, + "val_auc": 0.9273384183059012, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 1180, + 4790 + ] + }, + { + "test_f1": 0.8078435569815344, + "test_acc": 0.8080672268907563, + "test_auc": 0.9139617726388556, + "val_auc": 0.9506222400642312, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 1180, + 4790, + 6919, + 1275, + 1500, + 1091, + 1152 + ] + }, + { + "test_f1": 0.8335062187469964, + "test_acc": 0.8336134453781513, + "test_auc": 0.9289157465055572, + "val_auc": 0.9694901645925331, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 1180, + 4790, + 6919, + 1275, + 1500, + 1091, + 1152, + 2962, + 1067, + 1794, + 98, + 27957, + 347 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2f99602ec4de5b7247c73eda12ac8db40b4b7db4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7515816963040509, + "test_acc": 0.7603361344537816, + "test_auc": 0.7685849042960134, + "val_auc": 0.7442794058611, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235 + ] + }, + { + "test_f1": 0.7804902218501654, + "test_acc": 0.7865546218487395, + "test_auc": 0.8020085363472678, + "val_auc": 0.7904456041750301, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 8189 + ] + }, + { + "test_f1": 0.8136648672284388, + "test_acc": 0.8161344537815126, + "test_auc": 0.8221384164551563, + "val_auc": 0.7723805700521879, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 8189, + 692, + 5189, + 764 + ] + }, + { + "test_f1": 0.8161486347792434, + "test_acc": 0.8181512605042017, + "test_auc": 0.9187655017318552, + "val_auc": 0.9381774387796067, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 8189, + 692, + 5189, + 764, + 3494, + 42, + 1021, + 576, + 30916 + ] + }, + { + "test_f1": 0.8296441645907684, + "test_acc": 0.8315966386554622, + "test_auc": 0.9360772754553803, + "val_auc": 0.9223203532717785, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 8189, + 692, + 5189, + 764, + 3494, + 42, + 1021, + 576, + 30916, + 11265, + 964, + 2046, + 25737, + 214, + 60 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..96cd7931dfc3eab13f3eb5bc091c1a294ae914b9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9408400553949988, + "test_acc": 0.9408403361344538, + "test_auc": 0.98412812475143, + "val_auc": 0.9847450822962667, + "k": 1, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.987898831520315, + "test_acc": 0.9878991596638655, + "test_auc": 0.9993478425927934, + "val_auc": 0.9991971095945404, + "k": 2, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847 + ] + }, + { + "test_f1": 0.9885704522887694, + "test_acc": 0.9885714285714285, + "test_auc": 0.9995747192514336, + "val_auc": 0.9991971095945403, + "k": 5, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 279, + 1758 + ] + }, + { + "test_f1": 0.9959663263956778, + "test_acc": 0.9959663865546219, + "test_auc": 0.9999611326839779, + "val_auc": 1.0, + "k": 10, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 279, + 1758, + 1411, + 4674, + 786, + 461, + 431 + ] + }, + { + "test_f1": 0.9983193239332243, + "test_acc": 0.9983193277310924, + "test_auc": 0.999990057198227, + "val_auc": 1.0, + "k": 16, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 279, + 1758, + 1411, + 4674, + 786, + 461, + 431, + 1697, + 1634, + 1360, + 27673, + 4148, + 5020 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..60a1922b9bd5df947bc5ffa127b0cd3990ed14f7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.93108849723791, + "test_acc": 0.9310924369747899, + "test_auc": 0.9832721399078754, + "val_auc": 0.9851465274989964, + "k": 1, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9603284158857275, + "test_acc": 0.9603361344537815, + "test_auc": 0.9950484847170098, + "val_auc": 0.9971898835808912, + "k": 2, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9636776571656596, + "test_acc": 0.9636974789915966, + "test_auc": 0.9963586748233797, + "val_auc": 0.9971898835808912, + "k": 5, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758 + ] + }, + { + "test_f1": 0.9969747830798039, + "test_acc": 0.9969747899159664, + "test_auc": 0.9999850857973405, + "val_auc": 1.0, + "k": 10, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758, + 1847, + 431, + 279, + 461, + 1914 + ] + }, + { + "test_f1": 0.9983193273513061, + "test_acc": 0.9983193277310924, + "test_auc": 0.9999887013616215, + "val_auc": 1.0, + "k": 16, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758, + 1847, + 431, + 279, + 461, + 1914, + 1472, + 1551, + 1628, + 1634, + 1360, + 4674 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0d7a2d59e53c9ee2c1c65e7a716110a6c37d6929 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9223517831137886, + "test_acc": 0.9223529411764706, + "test_auc": 0.9757775270986544, + "val_auc": 0.9763147330389402, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9630205221869184, + "test_acc": 0.9630252100840336, + "test_auc": 0.9954719576834357, + "val_auc": 0.9967884383781614, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9727603060453417, + "test_acc": 0.9727731092436975, + "test_auc": 0.9988827906371347, + "val_auc": 1.0, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 1411 + ] + }, + { + "test_f1": 0.9963024750537636, + "test_acc": 0.9963025210084033, + "test_auc": 0.9997148223673269, + "val_auc": 1.0, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 1411, + 1758, + 279, + 431, + 461, + 1360 + ] + }, + { + "test_f1": 0.9983193239332243, + "test_acc": 0.9983193277310924, + "test_auc": 0.9999927688714378, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 1411, + 1758, + 279, + 431, + 461, + 1360, + 4148, + 27673, + 1634, + 4674, + 1914, + 1697 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..378998192e22137f2dd5ff180077b0b22c81ea41 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7475654813030507, + "test_acc": 0.7481761006289308, + "test_auc": 0.812197923322522, + "val_auc": 0.8179271708683473, + "k": 1, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.8117411909144813, + "test_acc": 0.8120754716981132, + "test_auc": 0.8790594599301901, + "val_auc": 0.9371748699479792, + "k": 2, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585 + ] + }, + { + "test_f1": 0.8519483012517318, + "test_acc": 0.8520754716981133, + "test_auc": 0.9171630183719327, + "val_auc": 0.9735894357743098, + "k": 5, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 776, + 32700, + 508 + ] + }, + { + "test_f1": 0.8587179735562935, + "test_acc": 0.8588679245283019, + "test_auc": 0.9257938167505283, + "val_auc": 0.9743897559023609, + "k": 10, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 776, + 32700, + 508, + 431, + 1866, + 1499, + 20310, + 1327 + ] + }, + { + "test_f1": 0.8624978154593468, + "test_acc": 0.8626415094339622, + "test_auc": 0.9295611363196794, + "val_auc": 0.9735894357743097, + "k": 16, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 776, + 32700, + 508, + 431, + 1866, + 1499, + 20310, + 1327, + 10363, + 1608, + 400, + 30737, + 3078, + 19688 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e19419af917aad74448c8cec94ad397481c66a2a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7152033069491607, + "test_acc": 0.7179080824088748, + "test_auc": 0.7398332328712075, + "val_auc": 0.6994797919167667, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21257 + ] + }, + { + "test_f1": 0.7179080824088748, + "test_acc": 0.7179080824088748, + "test_auc": 0.8034006429576049, + "val_auc": 0.7971188475390156, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21257, + 776 + ] + }, + { + "test_f1": 0.7606432004816883, + "test_acc": 0.7606973058637084, + "test_auc": 0.8502511553144465, + "val_auc": 0.8755502200880353, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21257, + 776, + 21219, + 508, + 1327 + ] + }, + { + "test_f1": 0.7859829287528103, + "test_acc": 0.786053882725832, + "test_auc": 0.8757032348804501, + "val_auc": 0.8883553421368549, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21257, + 776, + 21219, + 508, + 1327, + 431, + 5186, + 1866, + 913, + 1681 + ] + }, + { + "test_f1": 0.8097731180725332, + "test_acc": 0.8098256735340729, + "test_auc": 0.8961422543701024, + "val_auc": 0.8907563025210085, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21257, + 776, + 21219, + 508, + 1327, + 431, + 5186, + 1866, + 913, + 1681, + 1608, + 30737, + 11717, + 272, + 1551, + 28521 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..73e6226a779b3a25b4e0716dd01804f094092bce --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7005962806237124, + "test_acc": 0.7010804321728692, + "test_auc": 0.7098436635307139, + "val_auc": 0.7314814814814816, + "k": 1, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6913698402102405, + "test_acc": 0.6914765906362546, + "test_auc": 0.7284576185205681, + "val_auc": 0.7185990338164251, + "k": 2, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447 + ] + }, + { + "test_f1": 0.7142832437061987, + "test_acc": 0.7142857142857143, + "test_auc": 0.7946873270614278, + "val_auc": 0.7969001610305958, + "k": 5, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 347, + 1417 + ] + }, + { + "test_f1": 0.7466964891838833, + "test_acc": 0.7466986794717887, + "test_auc": 0.8302953790813502, + "val_auc": 0.8540660225442833, + "k": 10, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 347, + 1417, + 1528, + 30737, + 5186, + 1623, + 21219 + ] + }, + { + "test_f1": 0.7574393812186698, + "test_acc": 0.7575030012004802, + "test_auc": 0.840083702268954, + "val_auc": 0.8337359098228664, + "k": 16, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 347, + 1417, + 1528, + 30737, + 5186, + 1623, + 21219, + 559, + 1065, + 776, + 2837, + 32700, + 10363 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..18531bc235b25de678bc56fca753c03ff243f57a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7295965577052307, + "test_acc": 0.7300275482093664, + "test_auc": 0.7361726671118937, + "val_auc": 0.7290916366546618, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7245137250159852, + "test_acc": 0.7245179063360881, + "test_auc": 0.7593042316799223, + "val_auc": 0.7835134053621449, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186 + ] + }, + { + "test_f1": 0.7217208492583971, + "test_acc": 0.721763085399449, + "test_auc": 0.7675308117297066, + "val_auc": 0.7891156462585034, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 6731, + 3507, + 1377 + ] + }, + { + "test_f1": 0.7078648339118206, + "test_acc": 0.7079889807162535, + "test_auc": 0.7702021735170906, + "val_auc": 0.7947178871548619, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 6731, + 3507, + 1377, + 1681, + 508, + 3301, + 2247, + 1528 + ] + }, + { + "test_f1": 0.7354970420971703, + "test_acc": 0.7355371900826446, + "test_auc": 0.7980086212130413, + "val_auc": 0.7999199679871949, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 6731, + 3507, + 1377, + 1681, + 508, + 3301, + 2247, + 1528, + 12612, + 601, + 3078, + 30737, + 1688, + 1585 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bc18fa9121b1ef4bbc4d0f1a3e4938b84019bffa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6215885792082364, + "test_acc": 0.6274213836477988, + "test_auc": 0.6371285083424553, + "val_auc": 0.660264105642257, + "k": 1, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6412562269172027, + "test_acc": 0.6412578616352201, + "test_auc": 0.6886376132993228, + "val_auc": 0.7080832332933173, + "k": 2, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585 + ] + }, + { + "test_f1": 0.6696164074417074, + "test_acc": 0.669685534591195, + "test_auc": 0.7387243693666782, + "val_auc": 0.8035214085634254, + "k": 5, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 4605, + 27673, + 776 + ] + }, + { + "test_f1": 0.6779268819752257, + "test_acc": 0.6779874213836478, + "test_auc": 0.7413445190519059, + "val_auc": 0.8107242897158863, + "k": 10, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 4605, + 27673, + 776, + 2837, + 1411, + 1848, + 1866, + 32700 + ] + }, + { + "test_f1": 0.6776428992352093, + "test_acc": 0.6777358490566038, + "test_auc": 0.7417989314852376, + "val_auc": 0.8095238095238095, + "k": 16, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 4605, + 27673, + 776, + 2837, + 1411, + 1848, + 1866, + 32700, + 126, + 1499, + 373, + 1804, + 11717, + 2109 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..47e100e88b078294fbe0c32fae8794d662cc29dd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.803785958195501, + "test_acc": 0.8045283018867925, + "test_auc": 0.8453144635300479, + "val_auc": 0.8255302120848339, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293 + ] + }, + { + "test_f1": 0.8110495706391563, + "test_acc": 0.8110691823899371, + "test_auc": 0.8905653346348853, + "val_auc": 0.8855542216886754, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062 + ] + }, + { + "test_f1": 0.8259366737277322, + "test_acc": 0.8261635220125786, + "test_auc": 0.9127338262083826, + "val_auc": 0.9099639855942376, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1313 + ] + }, + { + "test_f1": 0.8381532383068143, + "test_acc": 0.8382389937106918, + "test_auc": 0.9194824204411168, + "val_auc": 0.9419767907162866, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1313, + 1441, + 409, + 1166, + 1116, + 1305 + ] + }, + { + "test_f1": 0.8402226314217519, + "test_acc": 0.8402515723270441, + "test_auc": 0.9218577696678307, + "val_auc": 0.9503801520608244, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1313, + 1441, + 409, + 1166, + 1116, + 1305, + 1928, + 14, + 524, + 765, + 396, + 1737 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b991034b364e19764c41884b08b4cb2f0e372432 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6500763448437867, + "test_acc": 0.6565656565656566, + "test_auc": 0.6657142857142857, + "val_auc": 0.6109995985547972, + "k": 1, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377 + ] + }, + { + "test_f1": 0.6500763448437867, + "test_acc": 0.6565656565656566, + "test_auc": 0.6983673469387754, + "val_auc": 0.6601766358892012, + "k": 2, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027 + ] + }, + { + "test_f1": 0.6453785163462583, + "test_acc": 0.6464646464646465, + "test_auc": 0.7306122448979591, + "val_auc": 0.6932958651144119, + "k": 5, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 681 + ] + }, + { + "test_f1": 0.6258042871226316, + "test_acc": 0.6262626262626263, + "test_auc": 0.7114285714285715, + "val_auc": 0.7244078683259735, + "k": 10, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 681, + 1339, + 1551, + 1790, + 524, + 505 + ] + }, + { + "test_f1": 0.6963497495412388, + "test_acc": 0.696969696969697, + "test_auc": 0.7187755102040816, + "val_auc": 0.7687675632276194, + "k": 16, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 681, + 1339, + 1551, + 1790, + 524, + 505, + 334, + 100, + 474, + 1221, + 957, + 6086 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5e5f7a791c10cfc45945d2de0bdeca9ed6ab7213 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6649923433462311, + "test_acc": 0.6697247706422018, + "test_auc": 0.689177796327212, + "val_auc": 0.7991143317230274, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6735480251514687, + "test_acc": 0.6788990825688074, + "test_auc": 0.7165776293823038, + "val_auc": 0.8478260869565217, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114 + ] + }, + { + "test_f1": 0.7333540614806602, + "test_acc": 0.737281067556297, + "test_auc": 0.789522815804118, + "val_auc": 0.8480273752012883, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 238 + ] + }, + { + "test_f1": 0.7696588105919772, + "test_acc": 0.7714762301918265, + "test_auc": 0.8502365052865889, + "val_auc": 0.889694041867955, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 238, + 3276, + 801, + 6086, + 4481, + 1335 + ] + }, + { + "test_f1": 0.7679075070203762, + "test_acc": 0.7698081734778982, + "test_auc": 0.851588759042849, + "val_auc": 0.890901771336554, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 238, + 3276, + 801, + 6086, + 4481, + 1335, + 363, + 30704, + 11788, + 1670, + 756, + 1202 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..45c4c6e31c74602cfe5dcc2397037175bbf2fd63 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.5690221211945878, + "test_acc": 0.5934591194968554, + "test_auc": 0.643444967743046, + "val_auc": 0.5608243297318928, + "k": 2, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727 + ] + }, + { + "test_f1": 0.6101136206794314, + "test_acc": 0.6123270440251573, + "test_auc": 0.6650876572975852, + "val_auc": 0.5694277711084434, + "k": 5, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1928 + ] + }, + { + "test_f1": 0.6233622174220271, + "test_acc": 0.6256603773584906, + "test_auc": 0.6864850400844928, + "val_auc": 0.6202480992396959, + "k": 10, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1928, + 7143, + 293, + 6400, + 1438, + 1739 + ] + }, + { + "test_f1": 0.6319144540972079, + "test_acc": 0.6334591194968553, + "test_auc": 0.6907529727939858, + "val_auc": 0.6128451380552221, + "k": 16, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1928, + 7143, + 293, + 6400, + 1438, + 1739, + 486, + 1848, + 1260, + 1825, + 765, + 6122 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..05cff71984a5cdcb331830cac0480afcab3c2790 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5479934506465526, + "test_acc": 0.5499700777977259, + "test_auc": 0.5689912041944818, + "val_auc": 0.5114411882778, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3727 + ] + }, + { + "test_f1": 0.5530698398370074, + "test_acc": 0.5541591861160982, + "test_auc": 0.5875168323639802, + "val_auc": 0.5453633079084705, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3727, + 7498 + ] + }, + { + "test_f1": 0.5529533322212796, + "test_acc": 0.5529622980251346, + "test_auc": 0.5902343638082685, + "val_auc": 0.5576073865917301, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3727, + 7498, + 293, + 19092, + 1336 + ] + }, + { + "test_f1": 0.5647665795234573, + "test_acc": 0.5649311789347696, + "test_auc": 0.5897050396814028, + "val_auc": 0.5684464070654356, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3727, + 7498, + 293, + 19092, + 1336, + 844, + 5392, + 5774, + 26677, + 672 + ] + }, + { + "test_f1": 0.5577067656157201, + "test_acc": 0.5577498503889886, + "test_auc": 0.5858994928802681, + "val_auc": 0.5656362906463269, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3727, + 7498, + 293, + 19092, + 1336, + 844, + 5392, + 5774, + 26677, + 672, + 6417, + 1130, + 100, + 4913, + 860, + 150 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c66d1c784907eae631645498c6b49b4f38e480de --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5621332960361692, + "test_acc": 0.5979874213836478, + "test_auc": 0.6045991601344352, + "val_auc": 0.6180472188875551, + "k": 1, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659 + ] + }, + { + "test_f1": 0.5865695760644412, + "test_acc": 0.6022641509433962, + "test_auc": 0.6144245948767593, + "val_auc": 0.6778711484593838, + "k": 2, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794 + ] + }, + { + "test_f1": 0.6000564135171245, + "test_acc": 0.6183647798742138, + "test_auc": 0.6198161540961926, + "val_auc": 0.6902761104441778, + "k": 5, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 6400 + ] + }, + { + "test_f1": 0.6013404381303474, + "test_acc": 0.6186163522012579, + "test_auc": 0.6245545745535114, + "val_auc": 0.6944777911164466, + "k": 10, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 6400, + 1377, + 7143, + 777, + 3053, + 27965 + ] + }, + { + "test_f1": 0.6205008249538724, + "test_acc": 0.6284276729559748, + "test_auc": 0.7046641955406319, + "val_auc": 0.8175270108043218, + "k": 16, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 6400, + 1377, + 7143, + 777, + 3053, + 27965, + 448, + 1224, + 1596, + 9390, + 121, + 598 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f8b2ec7ece3fd6d194e1fa9396ff3612b0c13b60 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9942136973190743, + "test_acc": 0.9942138364779874, + "test_auc": 0.999441541042936, + "val_auc": 1.0, + "k": 1, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9979874122127191, + "test_acc": 0.9979874213836478, + "test_auc": 0.9999832917991086, + "val_auc": 1.0, + "k": 2, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.9999908864358775, + "val_auc": 1.0, + "k": 5, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1551 + ] + }, + { + "test_f1": 0.9989937094180884, + "test_acc": 0.9989937106918239, + "test_auc": 0.999987342272052, + "val_auc": 1.0, + "k": 10, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1551, + 1525, + 347, + 1255, + 15467, + 6718 + ] + }, + { + "test_f1": 0.9992452824456872, + "test_acc": 0.999245283018868, + "test_auc": 0.9999979747635284, + "val_auc": 1.0, + "k": 16, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1551, + 1525, + 347, + 1255, + 15467, + 6718, + 1232, + 1335, + 1042, + 1758, + 1472, + 374 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..70db825fd9e7ad484f352de237329d6003de98fd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6449963968499409, + "test_acc": 0.6450314465408805, + "test_auc": 0.6913028244960453, + "val_auc": 0.764905962384954, + "k": 1, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6472175743441682, + "test_acc": 0.6475471698113208, + "test_auc": 0.7027737638716041, + "val_auc": 0.7511004401760705, + "k": 2, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928 + ] + }, + { + "test_f1": 0.6774739481242161, + "test_acc": 0.6774842767295598, + "test_auc": 0.7518477751258431, + "val_auc": 0.8371348539415766, + "k": 5, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 56 + ] + }, + { + "test_f1": 0.6950926028780438, + "test_acc": 0.6950943396226416, + "test_auc": 0.772108493943024, + "val_auc": 0.8351340536214485, + "k": 10, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 56, + 6400, + 5827, + 1135, + 152, + 1255 + ] + }, + { + "test_f1": 0.7184905304050307, + "test_acc": 0.7184905660377359, + "test_auc": 0.7954097255905842, + "val_auc": 0.86234493797519, + "k": 16, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 56, + 6400, + 5827, + 1135, + 152, + 1255, + 1327, + 1691, + 1396, + 2046, + 406, + 1616 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..187e3be0b06f642bbdcdb8d77c8e4dfca487d9f1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6704234465391401, + "test_acc": 0.6719496855345912, + "test_auc": 0.721129747787176, + "val_auc": 0.7190876350540216, + "k": 1, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.667526040993809, + "test_acc": 0.669937106918239, + "test_auc": 0.721129747787176, + "val_auc": 0.7190876350540216, + "k": 2, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1360 + ] + }, + { + "test_f1": 0.6796512461361974, + "test_acc": 0.6812578616352202, + "test_auc": 0.745372334662226, + "val_auc": 0.7194877951180472, + "k": 5, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1360, + 7143, + 6400, + 1011 + ] + }, + { + "test_f1": 0.7062785199455837, + "test_acc": 0.7066666666666667, + "test_auc": 0.7807108124337369, + "val_auc": 0.7494997999199681, + "k": 10, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1360, + 7143, + 6400, + 1011, + 1928, + 19940, + 598, + 253, + 5827 + ] + }, + { + "test_f1": 0.7141639797973762, + "test_acc": 0.7142138364779874, + "test_auc": 0.7890814945030019, + "val_auc": 0.7703081232492996, + "k": 16, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1360, + 7143, + 6400, + 1011, + 1928, + 19940, + 598, + 253, + 5827, + 1960, + 7068, + 30704, + 3780, + 1135, + 320 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3778868b07cfbe9fa87c158c432d24d295c8da5c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7950087103154354, + "test_acc": 0.7965834428383706, + "test_auc": 0.8200657769319903, + "val_auc": 0.7652000000000001, + "k": 1, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5448 + ] + }, + { + "test_f1": 0.7991469559718567, + "test_acc": 0.8010512483574245, + "test_auc": 0.8864085319930398, + "val_auc": 0.8052000000000001, + "k": 2, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5448, + 1062 + ] + }, + { + "test_f1": 0.7991469559718567, + "test_acc": 0.8010512483574245, + "test_auc": 0.8864096371162253, + "val_auc": 0.8052000000000001, + "k": 5, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5448, + 1062, + 811, + 1030, + 7319 + ] + }, + { + "test_f1": 0.8402068723710838, + "test_acc": 0.8402102496714849, + "test_auc": 0.9190064887307825, + "val_auc": 0.856, + "k": 10, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5448, + 1062, + 811, + 1030, + 7319, + 140, + 1662, + 2073, + 2046, + 6778 + ] + }, + { + "test_f1": 0.8578149277679735, + "test_acc": 0.8578186596583443, + "test_auc": 0.9270932276393519, + "val_auc": 0.89, + "k": 16, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5448, + 1062, + 811, + 1030, + 7319, + 140, + 1662, + 2073, + 2046, + 6778, + 296, + 628, + 1090, + 1726, + 1313, + 817 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..38a042a4072e0edea3dc35bfece592356c6cc6df --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5266775369927342, + "test_acc": 0.5288050314465409, + "test_auc": 0.5388582121819999, + "val_auc": 0.6450580232092837, + "k": 1, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411 + ] + }, + { + "test_f1": 0.5266453380488715, + "test_acc": 0.5288050314465409, + "test_auc": 0.5388582121819999, + "val_auc": 0.6450580232092837, + "k": 2, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 1011 + ] + }, + { + "test_f1": 0.5266453380488715, + "test_acc": 0.5288050314465409, + "test_auc": 0.5388582121819999, + "val_auc": 0.6450580232092837, + "k": 5, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 1011, + 1377, + 800, + 1928 + ] + }, + { + "test_f1": 0.557428817625828, + "test_acc": 0.5574842767295598, + "test_auc": 0.5954938235350705, + "val_auc": 0.6450580232092837, + "k": 10, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 1011, + 1377, + 800, + 1928, + 1703, + 15467, + 786, + 1202, + 1596 + ] + }, + { + "test_f1": 0.5547161356900944, + "test_acc": 0.5547169811320755, + "test_auc": 0.5961354437647526, + "val_auc": 0.6506602641056423, + "k": 16, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 1011, + 1377, + 800, + 1928, + 1703, + 15467, + 786, + 1202, + 1596, + 6718, + 18461, + 448, + 334, + 1417, + 4880 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cf0405046a7c68d077780d7548c998abb276464e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9932073485568238, + "test_acc": 0.9932075471698113, + "test_auc": 0.9968940467161297, + "val_auc": 1.0, + "k": 1, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.9916979135447588, + "test_acc": 0.9916981132075472, + "test_auc": 0.9984337327437194, + "val_auc": 1.0, + "k": 2, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 487 + ] + }, + { + "test_f1": 0.9919492116768907, + "test_acc": 0.9919496855345912, + "test_auc": 0.9992101577760473, + "val_auc": 1.0, + "k": 5, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 487, + 26988, + 152, + 1744 + ] + }, + { + "test_f1": 0.9916976560703812, + "test_acc": 0.9916981132075472, + "test_auc": 0.9992101577760473, + "val_auc": 1.0, + "k": 10, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 487, + 26988, + 152, + 1744, + 6935, + 292, + 559, + 3417, + 821 + ] + }, + { + "test_f1": 0.9916976560703812, + "test_acc": 0.9916981132075472, + "test_auc": 0.999273952724905, + "val_auc": 1.0, + "k": 16, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 487, + 26988, + 152, + 1744, + 6935, + 292, + 559, + 3417, + 821, + 347, + 6475, + 334, + 32198, + 1960, + 1101 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..22610555a07ac0bbc4380a01d1c4589498e8a672 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8555739053247428, + "test_acc": 0.8558490566037736, + "test_auc": 0.8927044906580904, + "val_auc": 0.9231692677070829, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487 + ] + }, + { + "test_f1": 0.8907682500603509, + "test_acc": 0.8908176100628931, + "test_auc": 0.9356972230970118, + "val_auc": 0.9711884753901561, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744 + ] + }, + { + "test_f1": 0.9060998621955282, + "test_acc": 0.9061635220125787, + "test_auc": 0.9643902671185645, + "val_auc": 0.9755902360944378, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 771, + 1524 + ] + }, + { + "test_f1": 0.9302535230693912, + "test_acc": 0.930314465408805, + "test_auc": 0.9783547788998712, + "val_auc": 0.9795918367346939, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 771, + 1524, + 7368, + 559, + 24947, + 6475, + 1028 + ] + }, + { + "test_f1": 0.9519323452743295, + "test_acc": 0.9519496855345913, + "test_auc": 0.9873532842753553, + "val_auc": 0.9859943977591037, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 771, + 1524, + 7368, + 559, + 24947, + 6475, + 1028, + 878, + 3417, + 152, + 1288, + 6935, + 6065 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8e3f3ddf4d739f83d6b8e92567dcaa05cd5297dc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8575176295029407, + "test_acc": 0.8587341772151899, + "test_auc": 0.8694306346881936, + "val_auc": 0.8459383753501402, + "k": 1, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104 + ] + }, + { + "test_f1": 0.8558023499616905, + "test_acc": 0.8572151898734177, + "test_auc": 0.917376809454077, + "val_auc": 0.9171668667466988, + "k": 2, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555 + ] + }, + { + "test_f1": 0.8595603875028988, + "test_acc": 0.8607594936708861, + "test_auc": 0.9383555041449778, + "val_auc": 0.9243697478991597, + "k": 5, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 292, + 1708, + 337 + ] + }, + { + "test_f1": 0.8691924804875802, + "test_acc": 0.8698734177215189, + "test_auc": 0.9459430080930642, + "val_auc": 0.9469787915166066, + "k": 10, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 292, + 1708, + 337, + 918, + 1815, + 968, + 1851, + 677 + ] + }, + { + "test_f1": 0.8713742025902225, + "test_acc": 0.8713924050632912, + "test_auc": 0.9474678923167166, + "val_auc": 0.9543817527010804, + "k": 16, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 292, + 1708, + 337, + 918, + 1815, + 968, + 1851, + 677, + 1438, + 1380, + 3825, + 1457, + 3203, + 1745 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fd3fe7b2c331b2855cf73d43914cbac299f6f7bf --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7508099062655202, + "test_acc": 0.7589743589743589, + "test_auc": 0.7677441343791025, + "val_auc": 0.7878416972664218, + "k": 1, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468 + ] + }, + { + "test_f1": 0.7712879059680313, + "test_acc": 0.7753846153846153, + "test_auc": 0.8332611000774228, + "val_auc": 0.8170134638922888, + "k": 2, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804 + ] + }, + { + "test_f1": 0.8566866906477258, + "test_acc": 0.8574358974358974, + "test_auc": 0.9022684047530884, + "val_auc": 0.8945328437372501, + "k": 5, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804, + 1351, + 1132, + 1474 + ] + }, + { + "test_f1": 0.8557188178942345, + "test_acc": 0.8564102564102564, + "test_auc": 0.9105871511764904, + "val_auc": 0.8990208078335373, + "k": 10, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804, + 1351, + 1132, + 1474, + 343, + 3417, + 26988, + 805, + 1305 + ] + }, + { + "test_f1": 0.8652503883059173, + "test_acc": 0.8656410256410256, + "test_auc": 0.9208435722220352, + "val_auc": 0.9006527947776418, + "k": 16, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804, + 1351, + 1132, + 1474, + 343, + 3417, + 26988, + 805, + 1305, + 930, + 6475, + 21810, + 1551, + 52, + 7595 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c4ea2ec67875b2c570905d5af519ba05201891ab --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.876594633925726, + "test_acc": 0.8774842767295598, + "test_auc": 0.8874358379770317, + "val_auc": 0.9303721488595438, + "k": 1, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.8716009032977746, + "test_acc": 0.8727044025157232, + "test_auc": 0.9208089503300629, + "val_auc": 0.9469787915166066, + "k": 2, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 11381 + ] + }, + { + "test_f1": 0.8955655601864492, + "test_acc": 0.8955974842767296, + "test_auc": 0.962733243952897, + "val_auc": 0.9867947178871549, + "k": 5, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 11381, + 1279, + 1447, + 7740 + ] + }, + { + "test_f1": 0.9096456656766495, + "test_acc": 0.9096855345911949, + "test_auc": 0.969483357112985, + "val_auc": 0.9895958383353342, + "k": 10, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 11381, + 1279, + 1447, + 7740, + 27068, + 238, + 8129, + 1012, + 32198 + ] + }, + { + "test_f1": 0.9262778769970862, + "test_acc": 0.9262893081761007, + "test_auc": 0.9792106944637122, + "val_auc": 0.9903961584633854, + "k": 16, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 11381, + 1279, + 1447, + 7740, + 27068, + 238, + 8129, + 1012, + 32198, + 1556, + 6828, + 7119, + 88, + 1478, + 262 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..75673fff2e6550dc514d27e0131a0f2ede1fce8b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8327625628756682, + "test_acc": 0.8358974358974359, + "test_auc": 0.8371469687279093, + "val_auc": 0.8784169726642187, + "k": 1, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628 + ] + }, + { + "test_f1": 0.9031527190757961, + "test_acc": 0.9035897435897436, + "test_auc": 0.9148033291816743, + "val_auc": 0.9077927376580989, + "k": 2, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416 + ] + }, + { + "test_f1": 0.8990857243373654, + "test_acc": 0.8994871794871795, + "test_auc": 0.9126384353855993, + "val_auc": 0.9218686250509996, + "k": 5, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 4211, + 9381 + ] + }, + { + "test_f1": 0.8970525587639774, + "test_acc": 0.8974358974358975, + "test_auc": 0.9285059076985223, + "val_auc": 0.9163606691146472, + "k": 10, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 4211, + 9381, + 238, + 805, + 737, + 1132, + 930 + ] + }, + { + "test_f1": 0.9052391316218514, + "test_acc": 0.9056410256410257, + "test_auc": 0.9369992762648535, + "val_auc": 0.920032639738882, + "k": 16, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 4211, + 9381, + 238, + 805, + 737, + 1132, + 930, + 334, + 2062, + 1274, + 1524, + 1158, + 52 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2a27fb3900412ac661a937ce9002409dbb3257fc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8706663790238413, + "test_acc": 0.8707692307692307, + "test_auc": 0.9030826067930118, + "val_auc": 0.9390044879640962, + "k": 1, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524 + ] + }, + { + "test_f1": 0.8716693293807722, + "test_acc": 0.8717948717948718, + "test_auc": 0.9030826067930118, + "val_auc": 0.9390044879640962, + "k": 2, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947 + ] + }, + { + "test_f1": 0.8726876636746232, + "test_acc": 0.8728205128205129, + "test_auc": 0.90512337159592, + "val_auc": 0.9375764993880049, + "k": 5, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 969 + ] + }, + { + "test_f1": 0.8726876636746232, + "test_acc": 0.8728205128205129, + "test_auc": 0.90512337159592, + "val_auc": 0.9375764993880049, + "k": 10, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 969, + 888, + 1474, + 3203, + 862, + 1260 + ] + }, + { + "test_f1": 0.8737903937731346, + "test_acc": 0.8738461538461538, + "test_auc": 0.931043188474097, + "val_auc": 0.9465524275805793, + "k": 16, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 969, + 888, + 1474, + 3203, + 862, + 1260, + 1214, + 1012, + 1158, + 1682, + 805, + 776 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b72824136f1d92319571b7f9a897b63f9916186c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8099545446837187, + "test_acc": 0.8123076923076923, + "test_auc": 0.8366883226175649, + "val_auc": 0.8180334557323541, + "k": 1, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.8113690977270988, + "test_acc": 0.8133333333333334, + "test_auc": 0.8664077490153835, + "val_auc": 0.832313341493268, + "k": 2, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815 + ] + }, + { + "test_f1": 0.8375120448002259, + "test_acc": 0.838974358974359, + "test_auc": 0.8976588009560038, + "val_auc": 0.8602611179110566, + "k": 5, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 1866, + 334, + 10639 + ] + }, + { + "test_f1": 0.8694717327996306, + "test_acc": 0.8697435897435898, + "test_auc": 0.9408725216278991, + "val_auc": 0.905140758873929, + "k": 10, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 1866, + 334, + 10639, + 487, + 1524, + 6935, + 930, + 2046 + ] + }, + { + "test_f1": 0.8829969230769231, + "test_acc": 0.8830769230769231, + "test_auc": 0.9495910054869222, + "val_auc": 0.9290085679314566, + "k": 16, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 1866, + 334, + 10639, + 487, + 1524, + 6935, + 930, + 2046, + 508, + 913, + 1744, + 1221, + 1158, + 821 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7506dde1956c5bdaafe919cff577277666f9c559 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8487582775530658, + "test_acc": 0.8492307692307692, + "test_auc": 0.8874086915541791, + "val_auc": 0.8947368421052632, + "k": 1, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52 + ] + }, + { + "test_f1": 0.8487582775530658, + "test_acc": 0.8492307692307692, + "test_auc": 0.8867270340323828, + "val_auc": 0.8947368421052632, + "k": 2, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230 + ] + }, + { + "test_f1": 0.8467351744853563, + "test_acc": 0.8471794871794872, + "test_auc": 0.8875244050223853, + "val_auc": 0.8996328029375766, + "k": 5, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 18545, + 2131, + 2454 + ] + }, + { + "test_f1": 0.8529388705778951, + "test_acc": 0.8533333333333334, + "test_auc": 0.9007599219039284, + "val_auc": 0.9179926560587516, + "k": 10, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 18545, + 2131, + 2454, + 1274, + 2884, + 888, + 2046, + 2484 + ] + }, + { + "test_f1": 0.8561239171743653, + "test_acc": 0.8564102564102564, + "test_auc": 0.905047631871276, + "val_auc": 0.9322725418196653, + "k": 16, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 18545, + 2131, + 2454, + 1274, + 2884, + 888, + 2046, + 2484, + 238, + 1193, + 821, + 1682, + 508, + 862 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..25ca1b18beec6e9a8c7b420f209cef6118f2d493 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8703684914691556, + "test_acc": 0.8705882352941177, + "test_auc": 0.9213107853364813, + "val_auc": 0.9204040404040403, + "k": 1, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839 + ] + }, + { + "test_f1": 0.8789599717707408, + "test_acc": 0.879144385026738, + "test_auc": 0.9375263090466517, + "val_auc": 0.9325252525252525, + "k": 2, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682 + ] + }, + { + "test_f1": 0.8876234103316484, + "test_acc": 0.8877005347593583, + "test_auc": 0.9453824191511558, + "val_auc": 0.9393939393939394, + "k": 5, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 930 + ] + }, + { + "test_f1": 0.8973196258916082, + "test_acc": 0.8973262032085562, + "test_auc": 0.9535222094108604, + "val_auc": 0.9519191919191918, + "k": 10, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 930, + 3417, + 343, + 340, + 874, + 1895 + ] + }, + { + "test_f1": 0.899401722911098, + "test_acc": 0.8994652406417112, + "test_auc": 0.9581068467578103, + "val_auc": 0.9612121212121212, + "k": 16, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 930, + 3417, + 343, + 340, + 874, + 1895, + 1441, + 238, + 334, + 888, + 456, + 1599 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b44bbca4fa745d928f34185f5c11441b3e376f07 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9309972870783607, + "test_acc": 0.9310344827586207, + "test_auc": 0.970137554791926, + "val_auc": 0.9857954545454546, + "k": 1, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554 + ] + }, + { + "test_f1": 0.9627526681676236, + "test_acc": 0.9627586206896551, + "test_auc": 0.9861210698630879, + "val_auc": 0.9995941558441559, + "k": 2, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 14949 + ] + }, + { + "test_f1": 0.9751720359113615, + "test_acc": 0.9751724137931035, + "test_auc": 0.9958254978624385, + "val_auc": 1.0, + "k": 5, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 14949, + 1072, + 192, + 760 + ] + }, + { + "test_f1": 0.9765515754358822, + "test_acc": 0.976551724137931, + "test_auc": 0.9944726108014502, + "val_auc": 1.0, + "k": 10, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 14949, + 1072, + 192, + 760, + 1682, + 1622, + 1065, + 1101, + 763 + ] + }, + { + "test_f1": 0.974712611605387, + "test_acc": 0.9747126436781609, + "test_auc": 0.9948894691271173, + "val_auc": 1.0, + "k": 16, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 14949, + 1072, + 192, + 760, + 1682, + 1622, + 1065, + 1101, + 763, + 1876, + 1684, + 4323, + 652, + 1872, + 432 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..37031fa4a7e6fcfd5904c10fba5811f8363e198e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9770662547920249, + "test_acc": 0.9770664118490205, + "test_auc": 0.9879241609917071, + "val_auc": 1.0, + "k": 1, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286 + ] + }, + { + "test_f1": 0.9751552795031055, + "test_acc": 0.9751552795031055, + "test_auc": 0.9926869266829931, + "val_auc": 1.0, + "k": 2, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1233 + ] + }, + { + "test_f1": 0.9861437970579846, + "test_acc": 0.9861442904921166, + "test_auc": 0.9985308109667792, + "val_auc": 1.0, + "k": 5, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1233, + 1905, + 304, + 581 + ] + }, + { + "test_f1": 0.9866215177530673, + "test_acc": 0.9866220735785953, + "test_auc": 0.9970807971788649, + "val_auc": 1.0, + "k": 10, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1233, + 1905, + 304, + 581, + 2874, + 1874, + 1079, + 1682, + 1684 + ] + }, + { + "test_f1": 0.9875771236278484, + "test_acc": 0.9875776397515528, + "test_auc": 0.9979135506893044, + "val_auc": 1.0, + "k": 16, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1233, + 1905, + 304, + 581, + 2874, + 1874, + 1079, + 1682, + 1684, + 948, + 760, + 239, + 1596, + 1156, + 63 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..96bb1bb89703af05652c60b55848bd0eaca9ab5e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9926434355656587, + "test_acc": 0.9926436781609196, + "test_auc": 0.9994596907300177, + "val_auc": 1.0, + "k": 1, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559 + ] + }, + { + "test_f1": 0.992643566195959, + "test_acc": 0.9926436781609196, + "test_auc": 0.9999036067969046, + "val_auc": 1.0, + "k": 2, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300 + ] + }, + { + "test_f1": 0.9935631612424094, + "test_acc": 0.9935632183908046, + "test_auc": 0.9999036067969046, + "val_auc": 1.0, + "k": 5, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 2236, + 334 + ] + }, + { + "test_f1": 0.9940229177503666, + "test_acc": 0.9940229885057471, + "test_auc": 0.9996854537583203, + "val_auc": 1.0, + "k": 10, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 2236, + 334, + 1934, + 982, + 3825, + 913, + 1872 + ] + }, + { + "test_f1": 0.9926436128484677, + "test_acc": 0.9926436781609196, + "test_auc": 0.99849744980789, + "val_auc": 1.0, + "k": 16, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 2236, + 334, + 1934, + 982, + 3825, + 913, + 1872, + 1592, + 119, + 1769, + 1158, + 1410, + 1305 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a10e81956b5967d8fb2ac168f6c31609956b1a02 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9462038034052619, + "test_acc": 0.9462068965517242, + "test_auc": 0.9791198793224741, + "val_auc": 0.9594155844155844, + "k": 1, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095 + ] + }, + { + "test_f1": 0.9627553140014299, + "test_acc": 0.9627586206896551, + "test_auc": 0.9876625155582012, + "val_auc": 0.976461038961039, + "k": 2, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023 + ] + }, + { + "test_f1": 0.9747115210855107, + "test_acc": 0.9747126436781609, + "test_auc": 0.9923502692245251, + "val_auc": 0.9768668831168831, + "k": 5, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 1682 + ] + }, + { + "test_f1": 0.9783900919778887, + "test_acc": 0.9783908045977011, + "test_auc": 0.9952293820011904, + "val_auc": 0.9841720779220778, + "k": 10, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 1682, + 494, + 1301, + 1079, + 211, + 2011 + ] + }, + { + "test_f1": 0.9806889122244278, + "test_acc": 0.9806896551724138, + "test_auc": 0.9966786622652741, + "val_auc": 0.9886363636363636, + "k": 16, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 1682, + 494, + 1301, + 1079, + 211, + 2011, + 1193, + 292, + 1701, + 1956, + 770, + 235 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cf9a423e46318df4daf563cb46612cb146eb5cc8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.959070163846223, + "test_acc": 0.9590804597701149, + "test_auc": 0.9903742085610693, + "val_auc": 0.9833603896103896, + "k": 1, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32 + ] + }, + { + "test_f1": 0.9935631421929442, + "test_acc": 0.9935632183908046, + "test_auc": 0.9992745143135451, + "val_auc": 0.9971590909090909, + "k": 2, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 12798 + ] + }, + { + "test_f1": 0.9972413781440628, + "test_acc": 0.9972413793103448, + "test_auc": 0.9996499404729693, + "val_auc": 0.9971590909090909, + "k": 5, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 12798, + 2683, + 1627, + 5094 + ] + }, + { + "test_f1": 0.9972413781440628, + "test_acc": 0.9972413793103448, + "test_auc": 0.9997556347746089, + "val_auc": 0.9971590909090909, + "k": 10, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 12798, + 2683, + 1627, + 5094, + 816, + 1047, + 1558, + 1458, + 87 + ] + }, + { + "test_f1": 0.9958620689655172, + "test_acc": 0.9958620689655172, + "test_auc": 0.9998232791276584, + "val_auc": 0.9983766233766234, + "k": 16, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 12798, + 2683, + 1627, + 5094, + 816, + 1047, + 1558, + 1458, + 87, + 1023, + 495, + 232, + 868, + 1682, + 1051 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5a961b35b96e57a2212acc6c79f0f645bad9a026 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8252535068240868, + "test_acc": 0.8254088050314465, + "test_auc": 0.874816336367475, + "val_auc": 0.921968787515006, + "k": 1, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.8968446390423828, + "test_acc": 0.8968553459119497, + "test_auc": 0.9514353357183869, + "val_auc": 0.9911964785914366, + "k": 2, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028 + ] + }, + { + "test_f1": 0.9179658232773042, + "test_acc": 0.9179874213836478, + "test_auc": 0.9656683179094699, + "val_auc": 0.9955982392957183, + "k": 5, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 778, + 1839, + 3417 + ] + }, + { + "test_f1": 0.9202360239600071, + "test_acc": 0.920251572327044, + "test_auc": 0.9676967188131304, + "val_auc": 0.9963985594237694, + "k": 10, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 778, + 1839, + 3417, + 1305, + 1602, + 1895, + 1478, + 4663 + ] + }, + { + "test_f1": 0.9192271864659971, + "test_acc": 0.9192452830188679, + "test_auc": 0.9700281457238652, + "val_auc": 0.9943977591036415, + "k": 16, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 778, + 1839, + 3417, + 1305, + 1602, + 1895, + 1478, + 4663, + 1065, + 5539, + 1279, + 654, + 343, + 8038 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fd4f32d813dd40c0c2601d7112211af1ca97752b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999847800205639, + "val_auc": 1.0, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379 + ] + }, + { + "test_f1": 0.9977011465095805, + "test_acc": 0.9977011494252873, + "test_auc": 0.9999813978029113, + "val_auc": 1.0, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9991569822501217, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 23428, + 4787 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999306645381243, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 23428, + 4787, + 547, + 6777, + 24163, + 1840, + 1162 + ] + }, + { + "test_f1": 0.9990804593813543, + "test_acc": 0.999080459770115, + "test_auc": 0.9999864711293901, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 23428, + 4787, + 547, + 6777, + 24163, + 1840, + 1162, + 1931, + 1584, + 829, + 613, + 3910, + 7651 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..82d050b37454234968d9a1af3e8f2bff98380ba9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9305745365345455, + "test_acc": 0.9305747126436782, + "test_auc": 0.9808076059310569, + "val_auc": 0.9862012987012987, + "k": 1, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586 + ] + }, + { + "test_f1": 0.9871261918582579, + "test_acc": 0.9871264367816092, + "test_auc": 0.9957493979652579, + "val_auc": 1.0, + "k": 2, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 2211 + ] + }, + { + "test_f1": 0.993563095929129, + "test_acc": 0.9935632183908046, + "test_auc": 0.9970819917203312, + "val_auc": 1.0, + "k": 5, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 2211, + 10, + 21724, + 559 + ] + }, + { + "test_f1": 0.994942483833445, + "test_acc": 0.9949425287356322, + "test_auc": 0.9971707749337085, + "val_auc": 1.0, + "k": 10, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 2211, + 10, + 21724, + 559, + 620, + 3223, + 1934, + 878, + 1874 + ] + }, + { + "test_f1": 0.9912642348581793, + "test_acc": 0.991264367816092, + "test_auc": 0.9964063937442503, + "val_auc": 1.0, + "k": 16, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 586, + 2211, + 10, + 21724, + 559, + 620, + 3223, + 1934, + 878, + 1874, + 495, + 1682, + 1872, + 1603, + 610, + 1023 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6f41ca4dc318c26d9439a3512258a034c107f4ab --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8817939536056102, + "test_acc": 0.8818390804597701, + "test_auc": 0.9296866544455868, + "val_auc": 0.9588068181818182, + "k": 1, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979 + ] + }, + { + "test_f1": 0.9158559858066625, + "test_acc": 0.9158620689655173, + "test_auc": 0.9702432490935657, + "val_auc": 0.9849837662337663, + "k": 2, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 1682 + ] + }, + { + "test_f1": 0.9158617132483948, + "test_acc": 0.9158620689655173, + "test_auc": 0.9744591833973699, + "val_auc": 0.9906655844155844, + "k": 5, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 1682, + 495, + 1708, + 1027 + ] + }, + { + "test_f1": 0.954013345284706, + "test_acc": 0.9540229885057471, + "test_auc": 0.9911292886519834, + "val_auc": 0.9983766233766234, + "k": 10, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 1682, + 495, + 1708, + 1027, + 1631, + 27304, + 117, + 1707, + 1840 + ] + }, + { + "test_f1": 0.9604524970555384, + "test_acc": 0.9604597701149425, + "test_auc": 0.993701465176687, + "val_auc": 0.9991883116883117, + "k": 16, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 1682, + 495, + 1708, + 1027, + 1631, + 27304, + 117, + 1707, + 1840, + 1841, + 1244, + 713, + 2024, + 2616, + 87 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c417c16f7607a70391987049ced55a66c0a51398 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8959564270676625, + "test_acc": 0.8960919540229885, + "test_auc": 0.9399994588451757, + "val_auc": 0.9366883116883117, + "k": 1, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961 + ] + }, + { + "test_f1": 0.9295703092748354, + "test_acc": 0.9296551724137931, + "test_auc": 0.9690501718166568, + "val_auc": 0.9833603896103896, + "k": 2, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 8734 + ] + }, + { + "test_f1": 0.9438701717911543, + "test_acc": 0.9439080459770115, + "test_auc": 0.9809039991341522, + "val_auc": 0.992288961038961, + "k": 5, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 8734, + 7182, + 248, + 732 + ] + }, + { + "test_f1": 0.955395170521722, + "test_acc": 0.9554022988505747, + "test_auc": 0.9843306083933114, + "val_auc": 0.9926948051948051, + "k": 10, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 8734, + 7182, + 248, + 732, + 1264, + 2689, + 1482, + 1441, + 1536 + ] + }, + { + "test_f1": 0.9682752585054257, + "test_acc": 0.9682758620689655, + "test_auc": 0.9912239907462524, + "val_auc": 0.9979707792207793, + "k": 16, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 8734, + 7182, + 248, + 732, + 1264, + 2689, + 1482, + 1441, + 1536, + 653, + 859, + 1682, + 306, + 1960, + 4186 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..15cd0014a2baba5ccda9a70baa689f1a86162430 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9581449407708581, + "test_acc": 0.9581609195402299, + "test_auc": 0.985187577791006, + "val_auc": 0.9898538961038961, + "k": 1, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786 + ] + }, + { + "test_f1": 0.9622837662115048, + "test_acc": 0.9622988505747126, + "test_auc": 0.9852095622057471, + "val_auc": 0.9902597402597402, + "k": 2, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 3355 + ] + }, + { + "test_f1": 0.9632085915007568, + "test_acc": 0.9632183908045977, + "test_auc": 0.9894999052979057, + "val_auc": 0.9914772727272727, + "k": 5, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 3355, + 1558, + 3930, + 271 + ] + }, + { + "test_f1": 0.972866737830642, + "test_acc": 0.9728735632183908, + "test_auc": 0.9937606539856053, + "val_auc": 0.9955357142857143, + "k": 10, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 3355, + 1558, + 3930, + 271, + 1910, + 653, + 292, + 760, + 327 + ] + }, + { + "test_f1": 0.96964188861032, + "test_acc": 0.9696551724137931, + "test_auc": 0.9928533741003301, + "val_auc": 0.9943181818181819, + "k": 16, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 3355, + 1558, + 3930, + 271, + 1910, + 653, + 292, + 760, + 327, + 67, + 276, + 1649, + 1627, + 460, + 1051 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..423abf97e40f714c9af1135f32289f2a80fd00ac --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9696551724137931, + "test_acc": 0.9696551724137931, + "test_auc": 0.9883702446019806, + "val_auc": 0.9813311688311688, + "k": 1, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831 + ] + }, + { + "test_f1": 0.9673563218390805, + "test_acc": 0.9673563218390805, + "test_auc": 0.9897992315601494, + "val_auc": 0.9902597402597402, + "k": 2, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856 + ] + }, + { + "test_f1": 0.9719530980989619, + "test_acc": 0.9719540229885058, + "test_auc": 0.9925971711131554, + "val_auc": 0.9935064935064934, + "k": 5, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 7739, + 1168, + 1167 + ] + }, + { + "test_f1": 0.9765487497574294, + "test_acc": 0.976551724137931, + "test_auc": 0.9948192881108285, + "val_auc": 0.9963474025974025, + "k": 10, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 7739, + 1168, + 1167, + 1037, + 18934, + 517, + 822, + 1252 + ] + }, + { + "test_f1": 0.9724115769990995, + "test_acc": 0.9724137931034482, + "test_auc": 0.9943643798365712, + "val_auc": 0.9979707792207793, + "k": 16, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 7739, + 1168, + 1167, + 1037, + 18934, + 517, + 822, + 1252, + 1449, + 1536, + 586, + 581, + 1366, + 1101 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dd44d7daa57402861d599074075530beb7cb67c1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9760911655883909, + "test_acc": 0.9760919540229885, + "test_auc": 0.9859295517885167, + "val_auc": 1.0, + "k": 1, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840 + ] + }, + { + "test_f1": 0.9751723088271955, + "test_acc": 0.9751724137931035, + "test_auc": 0.9952978719086529, + "val_auc": 1.0, + "k": 2, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 1765 + ] + }, + { + "test_f1": 0.9797701063897944, + "test_acc": 0.9797701149425287, + "test_auc": 0.9971538638454461, + "val_auc": 1.0, + "k": 5, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 1765, + 5495, + 789, + 1763 + ] + }, + { + "test_f1": 0.9770113970622181, + "test_acc": 0.9770114942528736, + "test_auc": 0.9971690838248823, + "val_auc": 1.0, + "k": 10, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 1765, + 5495, + 789, + 1763, + 1682, + 901, + 1797, + 434, + 10075 + ] + }, + { + "test_f1": 0.9779309411797295, + "test_acc": 0.9779310344827586, + "test_auc": 0.9973644068943124, + "val_auc": 1.0, + "k": 16, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 1765, + 5495, + 789, + 1763, + 1682, + 901, + 1797, + 434, + 10075, + 1310, + 1815, + 1801, + 27505, + 677, + 2120 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5630e8f7ad9a454994d54113048f4810e8d0009b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9268239742153075, + "test_acc": 0.926896551724138, + "test_auc": 0.9573113398993451, + "val_auc": 0.9707792207792209, + "k": 1, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1293 + ] + }, + { + "test_f1": 0.9618342398013067, + "test_acc": 0.9618390804597701, + "test_auc": 0.9809369757562639, + "val_auc": 0.9825487012987013, + "k": 2, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1293, + 1879 + ] + }, + { + "test_f1": 0.9894246032852515, + "test_acc": 0.9894252873563218, + "test_auc": 0.9953879234536502, + "val_auc": 1.0, + "k": 5, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1293, + 1879, + 4455, + 277, + 30 + ] + }, + { + "test_f1": 0.9898849676666431, + "test_acc": 0.9898850574712643, + "test_auc": 0.9973990746252503, + "val_auc": 1.0, + "k": 10, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1293, + 1879, + 4455, + 277, + 30, + 808, + 854, + 1501, + 990, + 30210 + ] + }, + { + "test_f1": 0.9848275669632323, + "test_acc": 0.9848275862068966, + "test_auc": 0.9919985185886682, + "val_auc": 1.0, + "k": 16, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1293, + 1879, + 4455, + 277, + 30, + 808, + 854, + 1501, + 990, + 30210, + 1867, + 1454, + 571, + 7367, + 419, + 1484 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..89b6f0d6aff4970c5f63a67dba39724400d0d99f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.982067047370732, + "test_acc": 0.9820689655172414, + "test_auc": 0.9894533998051843, + "val_auc": 1.0, + "k": 1, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439 + ] + }, + { + "test_f1": 0.9825270292029757, + "test_acc": 0.9825287356321839, + "test_auc": 0.9895979896098275, + "val_auc": 1.0, + "k": 2, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 26232 + ] + }, + { + "test_f1": 0.9806881038993697, + "test_acc": 0.9806896551724138, + "test_auc": 0.9917245589588182, + "val_auc": 1.0, + "k": 5, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 26232, + 1795, + 822, + 1121 + ] + }, + { + "test_f1": 0.9783892422563559, + "test_acc": 0.9783908045977011, + "test_auc": 0.9956589236430543, + "val_auc": 1.0, + "k": 10, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 26232, + 1795, + 822, + 1121, + 5535, + 1790, + 493, + 5797, + 802 + ] + }, + { + "test_f1": 0.9880456434463953, + "test_acc": 0.9880459770114942, + "test_auc": 0.9988170693760485, + "val_auc": 1.0, + "k": 16, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 26232, + 1795, + 822, + 1121, + 5535, + 1790, + 493, + 5797, + 802, + 6161, + 1441, + 1631, + 2020, + 345, + 495 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e15e6efda2b628a7d11177963f707094b2226910 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.954477446727041, + "test_acc": 0.9544827586206897, + "test_auc": 0.981894143351913, + "val_auc": 0.9947240259740261, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272 + ] + }, + { + "test_f1": 0.9507571719276735, + "test_acc": 0.9508045977011494, + "test_auc": 0.981894143351913, + "val_auc": 0.9947240259740261, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 5703 + ] + }, + { + "test_f1": 0.9659733718995844, + "test_acc": 0.9659770114942529, + "test_auc": 0.9938426727636778, + "val_auc": 0.9967532467532468, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 5703, + 1866, + 1072, + 3198 + ] + }, + { + "test_f1": 0.9788504852972406, + "test_acc": 0.9788505747126437, + "test_auc": 0.9962059973483415, + "val_auc": 0.9991883116883117, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 5703, + 1866, + 1072, + 3198, + 165, + 1682, + 918, + 1787, + 1726 + ] + }, + { + "test_f1": 0.9797699866508635, + "test_acc": 0.9797701149425287, + "test_auc": 0.9972942258780237, + "val_auc": 1.0, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 5703, + 1866, + 1072, + 3198, + 165, + 1682, + 918, + 1787, + 1726, + 1079, + 1715, + 1181, + 22196, + 495, + 410 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4e875af472ed799991638e77e63ea1ec8298ebad --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.576223117903976, + "test_acc": 0.589685534591195, + "test_auc": 0.5985762334449578, + "val_auc": 0.6054421768707483, + "k": 1, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.5905457353563918, + "test_acc": 0.5932075471698113, + "test_auc": 0.6257628812634235, + "val_auc": 0.644657863145258, + "k": 2, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556 + ] + }, + { + "test_f1": 0.6000255210886999, + "test_acc": 0.6002515723270441, + "test_auc": 0.6580208477842395, + "val_auc": 0.6916766706682673, + "k": 5, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556, + 1279, + 1132, + 6475 + ] + }, + { + "test_f1": 0.6024019252275621, + "test_acc": 0.6027672955974843, + "test_auc": 0.6550375478841848, + "val_auc": 0.6850740296118447, + "k": 10, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556, + 1279, + 1132, + 6475, + 1815, + 7740, + 238, + 1895, + 7893 + ] + }, + { + "test_f1": 0.6199660347196694, + "test_acc": 0.620125786163522, + "test_auc": 0.673764150074073, + "val_auc": 0.6894757903161264, + "k": 16, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1556, + 1279, + 1132, + 6475, + 1815, + 7740, + 238, + 1895, + 7893, + 581, + 925, + 1028, + 1051, + 17275, + 1866 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fe5188490c653c1188d704187cde0ec9f664642c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9687302647788286, + "test_acc": 0.968735632183908, + "test_auc": 0.981247294225878, + "val_auc": 1.0, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963 + ] + }, + { + "test_f1": 0.9678105666840883, + "test_acc": 0.967816091954023, + "test_auc": 0.9838042507711455, + "val_auc": 1.0, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217 + ] + }, + { + "test_f1": 0.9742524816858562, + "test_acc": 0.9742528735632184, + "test_auc": 0.9936625696736836, + "val_auc": 0.9995941558441559, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217, + 3772, + 352, + 1942 + ] + }, + { + "test_f1": 0.9839075153026831, + "test_acc": 0.9839080459770115, + "test_auc": 0.994196960062774, + "val_auc": 0.9991883116883117, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217, + 3772, + 352, + 1942, + 1682, + 111, + 581, + 1261, + 1898 + ] + }, + { + "test_f1": 0.9852872630188099, + "test_acc": 0.9852873563218391, + "test_auc": 0.995405257319119, + "val_auc": 0.9983766233766234, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217, + 3772, + 352, + 1942, + 1682, + 111, + 581, + 1261, + 1898, + 760, + 1726, + 393, + 1047, + 1441, + 8111 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..69b271d5e69e7cdf1609d47b084e979868eb74ff --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9788486968240413, + "test_acc": 0.9788505747126437, + "test_auc": 0.9880983988581633, + "val_auc": 0.9991883116883117, + "k": 1, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833 + ] + }, + { + "test_f1": 0.9774698069990138, + "test_acc": 0.9774712643678161, + "test_auc": 0.9968875142053142, + "val_auc": 0.9991883116883117, + "k": 2, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675 + ] + }, + { + "test_f1": 0.9834481289090525, + "test_acc": 0.983448275862069, + "test_auc": 0.9985887696845067, + "val_auc": 1.0, + "k": 5, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682 + ] + }, + { + "test_f1": 0.9875862016482797, + "test_acc": 0.9875862068965517, + "test_auc": 0.9990140835543049, + "val_auc": 1.0, + "k": 10, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682, + 829, + 1588, + 913, + 1281, + 361 + ] + }, + { + "test_f1": 0.9880456990405785, + "test_acc": 0.9880459770114942, + "test_auc": 0.9991688200119053, + "val_auc": 1.0, + "k": 16, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682, + 829, + 1588, + 913, + 1281, + 361, + 1181, + 491, + 113, + 859, + 1079, + 1515 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..81886d5f913ea2190691705b34085785910cd49d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.850065499613195, + "test_acc": 0.8501149425287357, + "test_auc": 0.9012130323610585, + "val_auc": 0.9318181818181818, + "k": 1, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020 + ] + }, + { + "test_f1": 0.9034471328941616, + "test_acc": 0.903448275862069, + "test_auc": 0.9598365881541209, + "val_auc": 0.9837662337662337, + "k": 2, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760 + ] + }, + { + "test_f1": 0.9415697699381094, + "test_acc": 0.9416091954022988, + "test_auc": 0.9829840629904215, + "val_auc": 0.984577922077922, + "k": 5, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 913 + ] + }, + { + "test_f1": 0.9553883231128933, + "test_acc": 0.9554022988505747, + "test_auc": 0.9933404134422857, + "val_auc": 0.994724025974026, + "k": 10, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 913, + 49, + 3842, + 170, + 18571, + 2817 + ] + }, + { + "test_f1": 0.959534669874365, + "test_acc": 0.9595402298850575, + "test_auc": 0.9958068956653499, + "val_auc": 0.9971590909090908, + "k": 16, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 913, + 49, + 3842, + 170, + 18571, + 2817, + 364, + 1960, + 2023, + 2401, + 1120, + 1928 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1c912926fb189bc725036884310a14c2e4272705 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9507997927120673, + "test_acc": 0.9508045977011494, + "test_auc": 0.9591939668001517, + "val_auc": 0.9525162337662337, + "k": 1, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123 + ] + }, + { + "test_f1": 0.9563153565435516, + "test_acc": 0.9563218390804598, + "test_auc": 0.9758716820444828, + "val_auc": 0.9671266233766233, + "k": 2, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491 + ] + }, + { + "test_f1": 0.9641330028431875, + "test_acc": 0.9641379310344828, + "test_auc": 0.9839090995183721, + "val_auc": 0.984577922077922, + "k": 5, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 866, + 1763, + 1682 + ] + }, + { + "test_f1": 0.9751690019683603, + "test_acc": 0.9751724137931035, + "test_auc": 0.9917575355809298, + "val_auc": 0.9906655844155844, + "k": 10, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 866, + 1763, + 1682, + 687, + 1111, + 204, + 167, + 1486 + ] + }, + { + "test_f1": 0.9719477610328515, + "test_acc": 0.9719540229885058, + "test_auc": 0.9934351155365551, + "val_auc": 0.9906655844155844, + "k": 16, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 866, + 1763, + 1682, + 687, + 1111, + 204, + 167, + 1486, + 523, + 679, + 1281, + 7316, + 1942, + 1724 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6efc9e8331d3060433ec1c941b449df0f360b350 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8442848041642262, + "test_acc": 0.8445977011494253, + "test_auc": 0.895586713296174, + "val_auc": 0.888392857142857, + "k": 1, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967 + ] + }, + { + "test_f1": 0.8717704942641643, + "test_acc": 0.872183908045977, + "test_auc": 0.923557653282104, + "val_auc": 0.9237012987012988, + "k": 2, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352 + ] + }, + { + "test_f1": 0.9282356312211963, + "test_acc": 0.9282758620689655, + "test_auc": 0.9734157692515829, + "val_auc": 0.9780844155844155, + "k": 5, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 1682, + 495, + 178 + ] + }, + { + "test_f1": 0.9475706243624062, + "test_acc": 0.9475862068965517, + "test_auc": 0.9821131419449104, + "val_auc": 0.9910714285714286, + "k": 10, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 1682, + 495, + 178, + 19873, + 1762, + 108, + 960, + 859 + ] + }, + { + "test_f1": 0.953538173834431, + "test_acc": 0.9535632183908046, + "test_auc": 0.9839378483684182, + "val_auc": 0.9963474025974026, + "k": 16, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 1682, + 495, + 178, + 19873, + 1762, + 108, + 960, + 859, + 182, + 1037, + 1797, + 1168, + 12346, + 507 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bec564ca117bbdb9b50e56856d5052c0b92882e5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.893608915014597, + "test_acc": 0.8937931034482759, + "test_auc": 0.9431812801017371, + "val_auc": 0.9565746753246753, + "k": 1, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335 + ] + }, + { + "test_f1": 0.9231130730376442, + "test_acc": 0.9232183908045977, + "test_auc": 0.9699578744791386, + "val_auc": 0.9699675324675325, + "k": 2, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1682 + ] + }, + { + "test_f1": 0.9682751243801947, + "test_acc": 0.9682758620689655, + "test_auc": 0.9932676957627578, + "val_auc": 0.9979707792207793, + "k": 5, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1682, + 1443, + 1281, + 1079 + ] + }, + { + "test_f1": 0.9714934574490965, + "test_acc": 0.9714942528735632, + "test_auc": 0.9944277964175551, + "val_auc": 0.9983766233766235, + "k": 10, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1682, + 1443, + 1281, + 1079, + 1515, + 878, + 21, + 165, + 760 + ] + }, + { + "test_f1": 0.9682740379056672, + "test_acc": 0.9682758620689655, + "test_auc": 0.9945529384706965, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1682, + 1443, + 1281, + 1079, + 1515, + 878, + 21, + 165, + 760, + 1928, + 571, + 559, + 335, + 5792, + 482 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..debab73b4e118d72e8423bd83907ecb75672d93f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6931390764479589, + "test_acc": 0.6961006289308176, + "test_auc": 0.71721521884199, + "val_auc": 0.7078831532613045, + "k": 1, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313 + ] + }, + { + "test_f1": 0.6931390764479589, + "test_acc": 0.6961006289308176, + "test_auc": 0.71721521884199, + "val_auc": 0.7078831532613045, + "k": 2, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757 + ] + }, + { + "test_f1": 0.6931390764479589, + "test_acc": 0.6961006289308176, + "test_auc": 0.71721521884199, + "val_auc": 0.7078831532613045, + "k": 5, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757, + 6636, + 19000, + 1105 + ] + }, + { + "test_f1": 0.7133096526647102, + "test_acc": 0.7147169811320755, + "test_auc": 0.780393103462243, + "val_auc": 0.7601040416166467, + "k": 10, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757, + 6636, + 19000, + 1105, + 4333, + 433, + 1067, + 121, + 6718 + ] + }, + { + "test_f1": 0.7282460907804996, + "test_acc": 0.7293081761006289, + "test_auc": 0.7929232415124872, + "val_auc": 0.7739095638255302, + "k": 16, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757, + 6636, + 19000, + 1105, + 4333, + 433, + 1067, + 121, + 6718, + 1928, + 1532, + 836, + 26569, + 1019, + 19189 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8897be8e4a4e9b17acac0db5ad4190cdebd54afe --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6896934703968214, + "test_acc": 0.6911225238444607, + "test_auc": 0.7551438069769746, + "val_auc": 0.7852564102564101, + "k": 1, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394 + ] + }, + { + "test_f1": 0.69440507011737, + "test_acc": 0.6955245781364637, + "test_auc": 0.7548886620934367, + "val_auc": 0.7880608974358974, + "k": 2, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041 + ] + }, + { + "test_f1": 0.6929323234673332, + "test_acc": 0.6940572267057961, + "test_auc": 0.7546507421809396, + "val_auc": 0.7916666666666666, + "k": 5, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 584, + 1928 + ] + }, + { + "test_f1": 0.7028224098631842, + "test_acc": 0.7035950110051358, + "test_auc": 0.7608129755706847, + "val_auc": 0.8127003205128204, + "k": 10, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 584, + 1928, + 1551, + 1737, + 28053, + 465, + 28397 + ] + }, + { + "test_f1": 0.6979764524742251, + "test_acc": 0.698459280997799, + "test_auc": 0.7680743774249529, + "val_auc": 0.8225160256410257, + "k": 16, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 584, + 1928, + 1551, + 1737, + 28053, + 465, + 28397, + 1257, + 31, + 1697, + 7649, + 2367, + 24358 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..19467298c29c9fe5b05da0a9d62d24b6319737c3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6514064789092876, + "test_acc": 0.6525786163522013, + "test_auc": 0.7050567116842981, + "val_auc": 0.6526610644257703, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011 + ] + }, + { + "test_f1": 0.7120973988663716, + "test_acc": 0.7122012578616352, + "test_auc": 0.7776094159319277, + "val_auc": 0.7186874749899961, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1848 + ] + }, + { + "test_f1": 0.7416209934630759, + "test_acc": 0.7416352201257862, + "test_auc": 0.8180828807773667, + "val_auc": 0.756702681072429, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1848, + 24630, + 1927, + 4520 + ] + }, + { + "test_f1": 0.7484273545220836, + "test_acc": 0.7484276729559748, + "test_auc": 0.8313398255663829, + "val_auc": 0.7759103641456583, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1848, + 24630, + 1927, + 4520, + 1520, + 7614, + 1036, + 6417, + 5653 + ] + }, + { + "test_f1": 0.763015718255043, + "test_acc": 0.7630188679245283, + "test_auc": 0.8396503834278951, + "val_auc": 0.7719087635054023, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1848, + 24630, + 1927, + 4520, + 1520, + 7614, + 1036, + 6417, + 5653, + 1928, + 17636, + 3089, + 4674, + 7387, + 1470 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bde8fbf6056675598e9ff51f02524c6c94532cb4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7332868121348757, + "test_acc": 0.7333333333333333, + "test_auc": 0.7940169957844704, + "val_auc": 0.7725090036014406, + "k": 1, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503 + ] + }, + { + "test_f1": 0.7453849236676082, + "test_acc": 0.7454088050314466, + "test_auc": 0.803146888376054, + "val_auc": 0.7879151660664265, + "k": 2, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472 + ] + }, + { + "test_f1": 0.7714786075731686, + "test_acc": 0.7715723270440251, + "test_auc": 0.8381489237387079, + "val_auc": 0.8495398159263705, + "k": 5, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472, + 1643, + 17225, + 1928 + ] + }, + { + "test_f1": 0.7803994415394697, + "test_acc": 0.7806289308176101, + "test_auc": 0.8515073835058667, + "val_auc": 0.87875150060024, + "k": 10, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472, + 1643, + 17225, + 1928, + 5891, + 7032, + 1848, + 7708, + 7303 + ] + }, + { + "test_f1": 0.7733835102906025, + "test_acc": 0.7735849056603774, + "test_auc": 0.8517521839643801, + "val_auc": 0.8679471788715487, + "k": 16, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472, + 1643, + 17225, + 1928, + 5891, + 7032, + 1848, + 7708, + 7303, + 7615, + 446, + 9605, + 18220, + 302, + 2471 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1b84dd79d94e5b90f718350fd9e3b78114fdceb4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8329108561931741, + "test_acc": 0.8332075471698113, + "test_auc": 0.8615368608227119, + "val_auc": 0.8627450980392157, + "k": 1, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8567916165736646, + "test_acc": 0.8576100628930817, + "test_auc": 0.9200436640983294, + "val_auc": 0.9323729491796718, + "k": 2, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8660826478278085, + "test_acc": 0.8666666666666667, + "test_auc": 0.931855602664806, + "val_auc": 0.9343737494997999, + "k": 5, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 100, + 1438, + 4675 + ] + }, + { + "test_f1": 0.8731502258765363, + "test_acc": 0.8732075471698113, + "test_auc": 0.9446701598620408, + "val_auc": 0.953981592637055, + "k": 10, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 100, + 1438, + 4675, + 1132, + 7284, + 1335, + 794, + 1874 + ] + }, + { + "test_f1": 0.8721128729586515, + "test_acc": 0.8722012578616353, + "test_auc": 0.944387006487845, + "val_auc": 0.9667867146858744, + "k": 16, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 100, + 1438, + 4675, + 1132, + 7284, + 1335, + 794, + 1874, + 777, + 1928, + 6745, + 22141, + 1960, + 12485 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0ceda023d64e91819bc66b0bc3f4ba6ebb3a5f6a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8696564344884344, + "test_acc": 0.8716981132075472, + "test_auc": 0.8727192040010572, + "val_auc": 0.8725490196078431, + "k": 1, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480 + ] + }, + { + "test_f1": 0.8777168882436788, + "test_acc": 0.8777358490566037, + "test_auc": 0.9508358657227715, + "val_auc": 0.9631852741096438, + "k": 2, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433 + ] + }, + { + "test_f1": 0.9287786536726239, + "test_acc": 0.9288050314465409, + "test_auc": 0.9820160267088185, + "val_auc": 0.9951980792316927, + "k": 5, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 3481, + 1116, + 444 + ] + }, + { + "test_f1": 0.9310357667964242, + "test_acc": 0.9310691823899371, + "test_auc": 0.9822199427060603, + "val_auc": 0.9955982392957183, + "k": 10, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 3481, + 1116, + 444, + 7666, + 1335, + 1928, + 6553, + 1064 + ] + }, + { + "test_f1": 0.9365757879018117, + "test_acc": 0.9366037735849057, + "test_auc": 0.9849542650973785, + "val_auc": 0.9971988795518207, + "k": 16, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 3481, + 1116, + 444, + 7666, + 1335, + 1928, + 6553, + 1064, + 1500, + 15841, + 988, + 1927, + 4110, + 993 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d95f597fe4c0302939c2dcd0a3b13858aa8aaf9a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8136249715197084, + "test_acc": 0.8181818181818182, + "test_auc": 0.8142857142857144, + "val_auc": 0.7315705128205128, + "k": 1, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629 + ] + }, + { + "test_f1": 0.7474747474747475, + "test_acc": 0.7474747474747475, + "test_auc": 0.8306122448979592, + "val_auc": 0.7936698717948717, + "k": 2, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625 + ] + }, + { + "test_f1": 0.7373737373737373, + "test_acc": 0.7373737373737373, + "test_auc": 0.8477551020408162, + "val_auc": 0.8313301282051283, + "k": 5, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 530, + 794 + ] + }, + { + "test_f1": 0.7978560110913052, + "test_acc": 0.797979797979798, + "test_auc": 0.893673469387755, + "val_auc": 0.889823717948718, + "k": 10, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 530, + 794, + 2046, + 1345, + 777, + 1182, + 567 + ] + }, + { + "test_f1": 0.8484539175567042, + "test_acc": 0.8484848484848485, + "test_auc": 0.9077551020408163, + "val_auc": 0.8878205128205129, + "k": 16, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 530, + 794, + 2046, + 1345, + 777, + 1182, + 567, + 765, + 1335, + 6400, + 7669, + 680, + 920 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a32d251ce3caffc1e1846257beaa7b2a5e096d3b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9189094539159566, + "test_acc": 0.9189765458422174, + "test_auc": 0.936442989634479, + "val_auc": 0.9339999999999998, + "k": 1, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655 + ] + }, + { + "test_f1": 0.9251690359651931, + "test_acc": 0.9253731343283582, + "test_auc": 0.9837606837606837, + "val_auc": 0.9862, + "k": 2, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 12485 + ] + }, + { + "test_f1": 0.946690248836681, + "test_acc": 0.9466950959488273, + "test_auc": 0.987152209492635, + "val_auc": 0.996, + "k": 5, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 12485, + 1417, + 1927, + 28188 + ] + }, + { + "test_f1": 0.9466936418945153, + "test_acc": 0.9466950959488273, + "test_auc": 0.9890161847608656, + "val_auc": 0.9956, + "k": 10, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 12485, + 1417, + 1927, + 28188, + 5406, + 1625, + 1118, + 330, + 727 + ] + }, + { + "test_f1": 0.9680153126089237, + "test_acc": 0.9680170575692963, + "test_auc": 0.9917803236952173, + "val_auc": 0.9956, + "k": 16, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 12485, + 1417, + 1927, + 28188, + 5406, + 1625, + 1118, + 330, + 727, + 984, + 4080, + 1472, + 5664, + 3958, + 567 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/scr/saebench_scr_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/scr/saebench_scr_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..34f4517687497a46dc61daee90c753dab436fb6c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/scr/saebench_scr_custom_sae_eval_results.json @@ -0,0 +1,326 @@ +{ + "eval_type_id": "scr", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": true, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "b63c0509-6787-40b6-bd40-237f57ede8da", + "datetime_epoch_millis": 1776879746023, + "eval_result_metrics": { + "scr_metrics": { + "scr_dir1_threshold_2": 0.2583111625287951, + "scr_metric_threshold_2": 0.17592138146269376, + "scr_dir2_threshold_2": 0.16922613756735483, + "scr_dir1_threshold_5": 0.047372645256674756, + "scr_metric_threshold_5": 0.234269763048432, + "scr_dir2_threshold_5": 0.22271814638463597, + "scr_dir1_threshold_10": -0.1472821486035981, + "scr_metric_threshold_10": 0.29005895468352294, + "scr_dir2_threshold_10": 0.2714674749685785, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results", + "scr_dir1_threshold_2": 0.48718003580863245, + "scr_metric_threshold_2": 0.06735756575889522, + "scr_dir2_threshold_2": 0.06735756575889522, + "scr_dir1_threshold_5": 0.5512825313325533, + "scr_metric_threshold_5": 0.1813471142552987, + "scr_dir2_threshold_5": 0.1813471142552987, + "scr_dir1_threshold_10": 0.5641028776049328, + "scr_metric_threshold_10": 0.22797925140897027, + "scr_dir2_threshold_10": 0.22797925140897027, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results", + "scr_dir1_threshold_2": 0.257575997034218, + "scr_metric_threshold_2": 0.25076456166892996, + "scr_dir2_threshold_2": 0.25076456166892996, + "scr_dir1_threshold_5": -0.7272726862227055, + "scr_metric_threshold_5": 0.34250758728798886, + "scr_dir2_threshold_5": 0.34250758728798886, + "scr_dir1_threshold_10": -0.48484845748180366, + "scr_metric_threshold_10": 0.39755347557029974, + "scr_dir2_threshold_10": 0.39755347557029974, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results", + "scr_dir1_threshold_2": 0.3906245634426147, + "scr_metric_threshold_2": 0.03517587036620758, + "scr_dir2_threshold_2": 0.03517587036620758, + "scr_dir1_threshold_5": -0.1250002328306055, + "scr_metric_threshold_5": 0.10301508102112025, + "scr_dir2_threshold_5": 0.10301508102112025, + "scr_dir1_threshold_10": -1.843750174622954, + "scr_metric_threshold_10": 0.12311562116192198, + "scr_dir2_threshold_10": 0.12311562116192198, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results", + "scr_dir1_threshold_2": 0.21476475165062875, + "scr_metric_threshold_2": 0.03184710232218929, + "scr_dir2_threshold_2": 0.03184710232218929, + "scr_dir1_threshold_5": -0.40939633291396726, + "scr_metric_threshold_5": -0.10509551359270591, + "scr_dir2_threshold_5": -0.10509551359270591, + "scr_dir1_threshold_10": -0.7248325369378812, + "scr_metric_threshold_10": -0.11146485812766252, + "scr_dir2_threshold_10": -0.11146485812766252, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results", + "scr_dir1_threshold_2": 0.03684225881529179, + "scr_metric_threshold_2": 0.48015863348328314, + "scr_dir2_threshold_2": 0.48015863348328314, + "scr_dir1_threshold_5": 0.07894729412163301, + "scr_metric_threshold_5": 0.5476190194611504, + "scr_dir2_threshold_5": 0.5476190194611504, + "scr_dir1_threshold_10": 0.07368420392192293, + "scr_metric_threshold_10": 0.6190476077844602, + "scr_dir2_threshold_10": 0.6190476077844602, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results", + "scr_dir1_threshold_2": 0.2841529787670631, + "scr_metric_threshold_2": 0.11068695256721958, + "scr_dir2_threshold_2": 0.11068695256721958, + "scr_dir1_threshold_5": 0.3825136878996036, + "scr_metric_threshold_5": 0.16412205665125212, + "scr_dir2_threshold_5": 0.16412205665125212, + "scr_dir1_threshold_10": 0.4699452484015855, + "scr_metric_threshold_10": 0.2290076214313713, + "scr_dir2_threshold_10": 0.2290076214313713, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results", + "scr_dir1_threshold_2": 0.25581394704114274, + "scr_metric_threshold_2": 0.29184559786405584, + "scr_dir2_threshold_2": 0.29184559786405584, + "scr_dir1_threshold_5": 0.39069757384182646, + "scr_metric_threshold_5": 0.4034334324782923, + "scr_dir2_threshold_5": 0.4034334324782923, + "scr_dir1_threshold_10": 0.45581383614879756, + "scr_metric_threshold_10": 0.5236051041022068, + "scr_dir2_threshold_10": 0.5236051041022068, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results", + "scr_dir1_threshold_2": 0.13953476767076942, + "scr_metric_threshold_2": 0.13953476767076942, + "scr_dir2_threshold_2": 0.08597281650805809, + "scr_dir1_threshold_5": 0.2372093268250598, + "scr_metric_threshold_5": 0.2372093268250598, + "scr_dir2_threshold_5": 0.14479639351469145, + "scr_dir1_threshold_10": 0.3116278141366155, + "scr_metric_threshold_10": 0.3116278141366155, + "scr_dir2_threshold_10": 0.16289597641705988, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_scr", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": null +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/tpp/saebench_tpp_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/tpp/saebench_tpp_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..310f2d6960281494b4e02390a5b21b81bd3ec44d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-134277-tokens-550000000/tpp/saebench_tpp_custom_sae_eval_results.json @@ -0,0 +1,297 @@ +{ + "eval_type_id": "tpp", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": false, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "e949b8d9-b2fa-47de-a93c-c75e41327285", + "datetime_epoch_millis": 1776879921188, + "eval_result_metrics": { + "tpp_metrics": { + "tpp_threshold_2_total_metric": 0.05415000170469284, + "tpp_threshold_2_intended_diff_only": 0.0634000062942505, + "tpp_threshold_2_unintended_diff_only": 0.009250004589557648, + "tpp_threshold_5_total_metric": 0.10479999929666518, + "tpp_threshold_5_intended_diff_only": 0.13109999895095825, + "tpp_threshold_5_unintended_diff_only": 0.02629999965429306, + "tpp_threshold_10_total_metric": 0.14532500505447388, + "tpp_threshold_10_intended_diff_only": 0.1815000116825104, + "tpp_threshold_10_unintended_diff_only": 0.0361750066280365, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results", + "tpp_threshold_2_total_metric": 0.09735000133514404, + "tpp_threshold_2_intended_diff_only": 0.11460001468658447, + "tpp_threshold_2_unintended_diff_only": 0.01725001335144043, + "tpp_threshold_5_total_metric": 0.19200000464916228, + "tpp_threshold_5_intended_diff_only": 0.2414000153541565, + "tpp_threshold_5_unintended_diff_only": 0.0494000107049942, + "tpp_threshold_10_total_metric": 0.2508500128984451, + "tpp_threshold_10_intended_diff_only": 0.3172000288963318, + "tpp_threshold_10_unintended_diff_only": 0.06635001599788666, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results", + "tpp_threshold_2_total_metric": 0.010950002074241637, + "tpp_threshold_2_intended_diff_only": 0.012199997901916504, + "tpp_threshold_2_unintended_diff_only": 0.0012499958276748658, + "tpp_threshold_5_total_metric": 0.017599993944168092, + "tpp_threshold_5_intended_diff_only": 0.02079998254776001, + "tpp_threshold_5_unintended_diff_only": 0.003199988603591919, + "tpp_threshold_10_total_metric": 0.039799997210502626, + "tpp_threshold_10_intended_diff_only": 0.045799994468688966, + "tpp_threshold_10_unintended_diff_only": 0.00599999725818634, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_tpp", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1": { + "0": { + "tpp_threshold_2_total_metric": 0.17649996280670166, + "tpp_threshold_2_intended_diff_only": 0.218999981880188, + "tpp_threshold_2_unintended_diff_only": 0.04250001907348633, + "tpp_threshold_5_total_metric": 0.21699996292591095, + "tpp_threshold_5_intended_diff_only": 0.296999990940094, + "tpp_threshold_5_unintended_diff_only": 0.08000002801418304, + "tpp_threshold_10_total_metric": 0.2474999874830246, + "tpp_threshold_10_intended_diff_only": 0.328000009059906, + "tpp_threshold_10_unintended_diff_only": 0.08050002157688141 + }, + "1": { + "tpp_threshold_2_total_metric": 0.08025002479553223, + "tpp_threshold_2_intended_diff_only": 0.1040000319480896, + "tpp_threshold_2_unintended_diff_only": 0.023750007152557373, + "tpp_threshold_5_total_metric": 0.08874998986721039, + "tpp_threshold_5_intended_diff_only": 0.12000000476837158, + "tpp_threshold_5_unintended_diff_only": 0.031250014901161194, + "tpp_threshold_10_total_metric": 0.22550000250339508, + "tpp_threshold_10_intended_diff_only": 0.2580000162124634, + "tpp_threshold_10_unintended_diff_only": 0.0325000137090683 + }, + "2": { + "tpp_threshold_2_total_metric": 0.15100005269050598, + "tpp_threshold_2_intended_diff_only": 0.16300004720687866, + "tpp_threshold_2_unintended_diff_only": 0.01199999451637268, + "tpp_threshold_5_total_metric": 0.2030000388622284, + "tpp_threshold_5_intended_diff_only": 0.23900002241134644, + "tpp_threshold_5_unintended_diff_only": 0.03599998354911804, + "tpp_threshold_10_total_metric": 0.22275005280971527, + "tpp_threshold_10_intended_diff_only": 0.30100005865097046, + "tpp_threshold_10_unintended_diff_only": 0.07825000584125519 + }, + "6": { + "tpp_threshold_2_total_metric": 0.06749998033046722, + "tpp_threshold_2_intended_diff_only": 0.07400000095367432, + "tpp_threshold_2_unintended_diff_only": 0.006500020623207092, + "tpp_threshold_5_total_metric": 0.24875003099441528, + "tpp_threshold_5_intended_diff_only": 0.27000004053115845, + "tpp_threshold_5_unintended_diff_only": 0.021250009536743164, + "tpp_threshold_10_total_metric": 0.3370000272989273, + "tpp_threshold_10_intended_diff_only": 0.38600003719329834, + "tpp_threshold_10_unintended_diff_only": 0.04900000989437103 + }, + "9": { + "tpp_threshold_2_total_metric": 0.011499986052513123, + "tpp_threshold_2_intended_diff_only": 0.013000011444091797, + "tpp_threshold_2_unintended_diff_only": 0.0015000253915786743, + "tpp_threshold_5_total_metric": 0.20250000059604645, + "tpp_threshold_5_intended_diff_only": 0.281000018119812, + "tpp_threshold_5_unintended_diff_only": 0.07850001752376556, + "tpp_threshold_10_total_metric": 0.2214999943971634, + "tpp_threshold_10_intended_diff_only": 0.31300002336502075, + "tpp_threshold_10_unintended_diff_only": 0.09150002896785736 + } + }, + "canrager/amazon_reviews_mcauley_1and5": { + "1": { + "tpp_threshold_2_total_metric": 0.011750012636184692, + "tpp_threshold_2_intended_diff_only": 0.008000016212463379, + "tpp_threshold_2_unintended_diff_only": -0.0037499964237213135, + "tpp_threshold_5_total_metric": 0.01000000536441803, + "tpp_threshold_5_intended_diff_only": 0.009999990463256836, + "tpp_threshold_5_unintended_diff_only": -1.4901161193847656e-08, + "tpp_threshold_10_total_metric": 0.0207500159740448, + "tpp_threshold_10_intended_diff_only": 0.027000010013580322, + "tpp_threshold_10_unintended_diff_only": 0.0062499940395355225 + }, + "2": { + "tpp_threshold_2_total_metric": 0.005250036716461182, + "tpp_threshold_2_intended_diff_only": 0.012000024318695068, + "tpp_threshold_2_unintended_diff_only": 0.006749987602233887, + "tpp_threshold_5_total_metric": 0.020500019192695618, + "tpp_threshold_5_intended_diff_only": 0.02799999713897705, + "tpp_threshold_5_unintended_diff_only": 0.007499977946281433, + "tpp_threshold_10_total_metric": 0.05475004017353058, + "tpp_threshold_10_intended_diff_only": 0.06700003147125244, + "tpp_threshold_10_unintended_diff_only": 0.012249991297721863 + }, + "3": { + "tpp_threshold_2_total_metric": 0.01974998414516449, + "tpp_threshold_2_intended_diff_only": 0.015999972820281982, + "tpp_threshold_2_unintended_diff_only": -0.0037500113248825073, + "tpp_threshold_5_total_metric": 0.02350001037120819, + "tpp_threshold_5_intended_diff_only": 0.02399998903274536, + "tpp_threshold_5_unintended_diff_only": 0.0004999786615371704, + "tpp_threshold_10_total_metric": 0.01799999177455902, + "tpp_threshold_10_intended_diff_only": 0.023000001907348633, + "tpp_threshold_10_unintended_diff_only": 0.005000010132789612 + }, + "5": { + "tpp_threshold_2_total_metric": -0.0022500157356262207, + "tpp_threshold_2_intended_diff_only": 0.0059999823570251465, + "tpp_threshold_2_unintended_diff_only": 0.008249998092651367, + "tpp_threshold_5_total_metric": 0.006999939680099487, + "tpp_threshold_5_intended_diff_only": 0.008999943733215332, + "tpp_threshold_5_unintended_diff_only": 0.0020000040531158447, + "tpp_threshold_10_total_metric": 0.013499945402145386, + "tpp_threshold_10_intended_diff_only": 0.01799994707107544, + "tpp_threshold_10_unintended_diff_only": 0.004500001668930054 + }, + "6": { + "tpp_threshold_2_total_metric": 0.020249992609024048, + "tpp_threshold_2_intended_diff_only": 0.018999993801116943, + "tpp_threshold_2_unintended_diff_only": -0.0012499988079071045, + "tpp_threshold_5_total_metric": 0.02699999511241913, + "tpp_threshold_5_intended_diff_only": 0.03299999237060547, + "tpp_threshold_5_unintended_diff_only": 0.00599999725818634, + "tpp_threshold_10_total_metric": 0.09199999272823334, + "tpp_threshold_10_intended_diff_only": 0.09399998188018799, + "tpp_threshold_10_unintended_diff_only": 0.001999989151954651 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b485c665efa5dff2e077094b2656077b9e73b842 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9780379956472244, + "test_acc": 0.9780429594272076, + "test_auc": 0.981233185327763, + "val_auc": 0.9856284856284856, + "k": 1, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431 + ] + }, + { + "test_f1": 0.9780379956472244, + "test_acc": 0.9780429594272076, + "test_auc": 0.9937535087527432, + "val_auc": 0.9931359931359931, + "k": 2, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928 + ] + }, + { + "test_f1": 0.9828136838952073, + "test_acc": 0.9828162291169451, + "test_auc": 0.9978911028966805, + "val_auc": 0.9982839982839983, + "k": 5, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472 + ] + }, + { + "test_f1": 0.9875886334975447, + "test_acc": 0.9875894988066826, + "test_auc": 0.9986297637014516, + "val_auc": 0.9991419991419991, + "k": 10, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472, + 461, + 1411, + 126, + 5020, + 786 + ] + }, + { + "test_f1": 0.984245690225062, + "test_acc": 0.9842482100238663, + "test_auc": 0.9987660126716099, + "val_auc": 0.9969969969969971, + "k": 16, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472, + 461, + 1411, + 126, + 5020, + 786, + 542, + 1634, + 1848, + 5951, + 1872, + 747 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..00eb76c4ec2f250f5c2cf1407a512371e53f1235 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9137859545465096, + "test_acc": 0.9138042773817239, + "test_auc": 0.9514492315343105, + "val_auc": 0.940884663083919, + "k": 1, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838 + ] + }, + { + "test_f1": 0.9475048165631439, + "test_acc": 0.9475048606610499, + "test_auc": 0.9657088566762766, + "val_auc": 0.9607275733774288, + "k": 2, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441 + ] + }, + { + "test_f1": 0.9604521718748158, + "test_acc": 0.9604666234607907, + "test_auc": 0.9696763506112108, + "val_auc": 0.9665150888797024, + "k": 5, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 943 + ] + }, + { + "test_f1": 0.9539689213624907, + "test_acc": 0.9539857420609202, + "test_auc": 0.969837637682036, + "val_auc": 0.9694088466308393, + "k": 10, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 943, + 1697, + 1411, + 1180, + 1417, + 654 + ] + }, + { + "test_f1": 0.962401412525797, + "test_acc": 0.9624108878807518, + "test_auc": 0.9870264712404992, + "val_auc": 0.9805704836709385, + "k": 16, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 943, + 1697, + 1411, + 1180, + 1417, + 654, + 698, + 601, + 908, + 1703, + 1619, + 559 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..251ea289c64d6cb1bb66239908df132930cc052b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5666526782374632, + "test_acc": 0.5783647798742139, + "test_auc": 0.5883425869763119, + "val_auc": 0.5750300120048019, + "k": 1, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777 + ] + }, + { + "test_f1": 0.5686365781944279, + "test_acc": 0.5859119496855346, + "test_auc": 0.6006879221985157, + "val_auc": 0.594437775110044, + "k": 2, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921 + ] + }, + { + "test_f1": 0.5748392286161321, + "test_acc": 0.5859119496855346, + "test_auc": 0.618511142344758, + "val_auc": 0.6058423369347739, + "k": 5, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921, + 7284, + 1438, + 2046 + ] + }, + { + "test_f1": 0.5783681777740308, + "test_acc": 0.5861635220125786, + "test_auc": 0.6201401919316605, + "val_auc": 0.615846338535414, + "k": 10, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921, + 7284, + 1438, + 2046, + 969, + 1380, + 27886, + 505, + 1178 + ] + }, + { + "test_f1": 0.5888656275625692, + "test_acc": 0.5932075471698113, + "test_auc": 0.6258679404053915, + "val_auc": 0.6162464985994398, + "k": 16, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921, + 7284, + 1438, + 2046, + 969, + 1380, + 27886, + 505, + 1178, + 1335, + 521, + 1533, + 5951, + 559, + 152 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..96e38706da96a219516079f41ea2f843dc7dfbfb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6516470066793991, + "test_acc": 0.6638993710691824, + "test_auc": 0.6705441506613917, + "val_auc": 0.7266906762705082, + "k": 1, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567 + ] + }, + { + "test_f1": 0.6692025128900353, + "test_acc": 0.6711949685534592, + "test_auc": 0.7077707817108996, + "val_auc": 0.8455382152861145, + "k": 2, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 433 + ] + }, + { + "test_f1": 0.6898275405718665, + "test_acc": 0.6910691823899371, + "test_auc": 0.7410903518747107, + "val_auc": 0.8067226890756303, + "k": 5, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 433, + 1164, + 1345, + 4470 + ] + }, + { + "test_f1": 0.7235955909307056, + "test_acc": 0.7250314465408805, + "test_auc": 0.7675643696097065, + "val_auc": 0.7839135654261704, + "k": 10, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 433, + 1164, + 1345, + 4470, + 1629, + 1731, + 27886, + 6411, + 559 + ] + }, + { + "test_f1": 0.7174734903932314, + "test_acc": 0.7184905660377359, + "test_auc": 0.7703278047753052, + "val_auc": 0.7859143657462985, + "k": 16, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 433, + 1164, + 1345, + 4470, + 1629, + 1731, + 27886, + 6411, + 559, + 4505, + 794, + 628, + 993, + 1335, + 28711 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..330078821d89090c861adcbae64fde028f89d31f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7659748058995153, + "test_acc": 0.7665408805031446, + "test_auc": 0.8147863021106002, + "val_auc": 0.8575430172068828, + "k": 1, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411 + ] + }, + { + "test_f1": 0.7360846586701358, + "test_acc": 0.7381132075471698, + "test_auc": 0.8200929279755027, + "val_auc": 0.8639455782312925, + "k": 2, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786 + ] + }, + { + "test_f1": 0.753671598839098, + "test_acc": 0.7542138364779875, + "test_auc": 0.8337901085425488, + "val_auc": 0.8883553421368546, + "k": 5, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417 + ] + }, + { + "test_f1": 0.7940488723634468, + "test_acc": 0.7942138364779874, + "test_auc": 0.8872183782108858, + "val_auc": 0.9225690276110444, + "k": 10, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417, + 1758, + 1335, + 181, + 5054, + 1360 + ] + }, + { + "test_f1": 0.7979535094936915, + "test_acc": 0.7979874213836478, + "test_auc": 0.8955998193489066, + "val_auc": 0.920968387354942, + "k": 16, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417, + 1758, + 1335, + 181, + 5054, + 1360, + 1183, + 19471, + 191, + 126, + 1108, + 1703 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8d7ec6d68dbc843d8746114519a259c65321a5bb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8461310033829663, + "test_acc": 0.8462893081761006, + "test_auc": 0.8918526255671928, + "val_auc": 0.9397759103641457, + "k": 1, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.875936918486553, + "test_acc": 0.8759748427672956, + "test_auc": 0.9354314108101048, + "val_auc": 0.9615846338535414, + "k": 2, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8739300935297301, + "test_acc": 0.8739622641509434, + "test_auc": 0.9400559117158916, + "val_auc": 0.9595838335334134, + "k": 5, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 821, + 22141 + ] + }, + { + "test_f1": 0.8840238213562301, + "test_acc": 0.8840251572327044, + "test_auc": 0.9435632415529919, + "val_auc": 0.9603841536614646, + "k": 10, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 821, + 22141, + 505, + 6745, + 7284, + 100, + 1697 + ] + }, + { + "test_f1": 0.8877986001168493, + "test_acc": 0.8877987421383647, + "test_auc": 0.9446062383359037, + "val_auc": 0.961984793917567, + "k": 16, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 821, + 22141, + 505, + 6745, + 7284, + 100, + 1697, + 4675, + 7143, + 7015, + 6567, + 373, + 18129 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6ffe95423203ec578d82d61a975461ee0b6dff3e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5256243841987526, + "test_acc": 0.56, + "test_auc": 0.5628892180294206, + "val_auc": 0.5450836393308853, + "k": 1, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945 + ] + }, + { + "test_f1": 0.5602444991814234, + "test_acc": 0.5620512820512821, + "test_auc": 0.5792069209277275, + "val_auc": 0.5885352917176663, + "k": 2, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 1264 + ] + }, + { + "test_f1": 0.548458919764304, + "test_acc": 0.556923076923077, + "test_auc": 0.5821481469014037, + "val_auc": 0.6119951040391678, + "k": 5, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 1264, + 238, + 6892, + 375 + ] + }, + { + "test_f1": 0.5438731866227128, + "test_acc": 0.5517948717948717, + "test_auc": 0.5943043727067694, + "val_auc": 0.7076703386372909, + "k": 10, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 1264, + 238, + 6892, + 375, + 5888, + 1523, + 1447, + 551, + 913 + ] + }, + { + "test_f1": 0.5746952043151257, + "test_acc": 0.5794871794871795, + "test_auc": 0.6300430033325479, + "val_auc": 0.6931864545083639, + "k": 16, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1945, + 1264, + 238, + 6892, + 375, + 5888, + 1523, + 1447, + 551, + 913, + 7073, + 1132, + 1851, + 1588, + 470, + 386 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3d3fddd9f834b1c07acccd939da8c623e17f3b75 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.4774287678095039, + "test_acc": 0.5158974358974359, + "test_auc": 0.5148702326051098, + "val_auc": 0.5377396980824154, + "k": 1, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341 + ] + }, + { + "test_f1": 0.52237570709382, + "test_acc": 0.5476923076923077, + "test_auc": 0.5548271451173123, + "val_auc": 0.540187678498572, + "k": 2, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 1409 + ] + }, + { + "test_f1": 0.5192352244782751, + "test_acc": 0.521025641025641, + "test_auc": 0.5378340963409298, + "val_auc": 0.5556915544675642, + "k": 5, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 1409, + 239, + 456, + 8141 + ] + }, + { + "test_f1": 0.5039992130656488, + "test_acc": 0.5066666666666667, + "test_auc": 0.5167511024337699, + "val_auc": 0.532843737250102, + "k": 10, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 1409, + 239, + 456, + 8141, + 1264, + 400, + 538, + 238, + 326 + ] + }, + { + "test_f1": 0.5105839148252755, + "test_acc": 0.5117948717948718, + "test_auc": 0.5198669505503754, + "val_auc": 0.5516115871073031, + "k": 16, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 1409, + 239, + 456, + 8141, + 1264, + 400, + 538, + 238, + 326, + 11202, + 1478, + 347, + 4201, + 2769, + 827 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..68bdb73b90f8cea0296f761d1aeb1fe8471b8ca7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5374154599047375, + "test_acc": 0.5374358974358975, + "test_auc": 0.5500892045645807, + "val_auc": 0.6054671562627498, + "k": 1, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815 + ] + }, + { + "test_f1": 0.5508594270349493, + "test_acc": 0.5517948717948717, + "test_auc": 0.5572886861682431, + "val_auc": 0.6138310893512853, + "k": 2, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156 + ] + }, + { + "test_f1": 0.5283458935111163, + "test_acc": 0.5364102564102564, + "test_auc": 0.5453176019120073, + "val_auc": 0.6177070583435333, + "k": 5, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 6990, + 12594, + 16809 + ] + }, + { + "test_f1": 0.5405696433574152, + "test_acc": 0.5446153846153846, + "test_auc": 0.5525044602282291, + "val_auc": 0.6254589963280294, + "k": 10, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 6990, + 12594, + 16809, + 698, + 1770, + 1409, + 1611, + 191 + ] + }, + { + "test_f1": 0.5400161362453546, + "test_acc": 0.5446153846153846, + "test_auc": 0.5522793449355371, + "val_auc": 0.6132190942472461, + "k": 16, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 6990, + 12594, + 16809, + 698, + 1770, + 1409, + 1611, + 191, + 1684, + 14220, + 360, + 848, + 24861, + 400 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cecdd1d6cc73973c6365a3510d595393845a1b10 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8478331704138156, + "test_acc": 0.8484848484848485, + "test_auc": 0.8814285714285713, + "val_auc": 0.87900641025641, + "k": 1, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997 + ] + }, + { + "test_f1": 0.83818577936225, + "test_acc": 0.8383838383838383, + "test_auc": 0.8873469387755102, + "val_auc": 0.8950320512820512, + "k": 2, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335 + ] + }, + { + "test_f1": 0.858151895385938, + "test_acc": 0.8585858585858586, + "test_auc": 0.9138775510204081, + "val_auc": 0.924278846153846, + "k": 5, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 1156, + 12946 + ] + }, + { + "test_f1": 0.8473644735326978, + "test_acc": 0.8484848484848485, + "test_auc": 0.8802040816326531, + "val_auc": 0.9475160256410255, + "k": 10, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 1156, + 12946, + 3526, + 314, + 1279, + 12605, + 851 + ] + }, + { + "test_f1": 0.8473644735326978, + "test_acc": 0.8484848484848485, + "test_auc": 0.8791836734693877, + "val_auc": 0.9491185897435898, + "k": 16, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 2335, + 6369, + 1156, + 12946, + 3526, + 314, + 1279, + 12605, + 851, + 5836, + 1499, + 771, + 4201, + 1684, + 2900 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..119053a0e64c9235a2ff8d3a37a1b2e7d17b475d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.877665544332211, + "test_acc": 0.8787878787878788, + "test_auc": 0.9020408163265305, + "val_auc": 0.939903846153846, + "k": 1, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.858151895385938, + "test_acc": 0.8585858585858586, + "test_auc": 0.9048979591836734, + "val_auc": 0.9607371794871794, + "k": 2, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335 + ] + }, + { + "test_f1": 0.8681220810253067, + "test_acc": 0.8686868686868687, + "test_auc": 0.9342857142857143, + "val_auc": 0.9837740384615385, + "k": 5, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 4201, + 4889, + 789 + ] + }, + { + "test_f1": 0.8074911945879687, + "test_acc": 0.8080808080808081, + "test_auc": 0.9481632653061225, + "val_auc": 0.9879807692307692, + "k": 10, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 4201, + 4889, + 789, + 2832, + 2900, + 1116, + 3526, + 1924 + ] + }, + { + "test_f1": 0.8984098573926301, + "test_acc": 0.898989898989899, + "test_auc": 0.9655102040816326, + "val_auc": 0.9895833333333334, + "k": 16, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 4201, + 4889, + 789, + 2832, + 2900, + 1116, + 3526, + 1924, + 581, + 7997, + 2000, + 12605, + 472, + 4531 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a2a2d01308c7592ae104d7f165f4837d6156db78 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8671979042043984, + "test_acc": 0.8686868686868687, + "test_auc": 0.8775510204081632, + "val_auc": 0.826923076923077, + "k": 1, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388 + ] + }, + { + "test_f1": 0.8671979042043984, + "test_acc": 0.8686868686868687, + "test_auc": 0.907142857142857, + "val_auc": 0.8886217948717948, + "k": 2, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335 + ] + }, + { + "test_f1": 0.8671979042043984, + "test_acc": 0.8686868686868687, + "test_auc": 0.9481632653061225, + "val_auc": 0.9425080128205128, + "k": 5, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 4201, + 337, + 5888 + ] + }, + { + "test_f1": 0.8784159103308039, + "test_acc": 0.8787878787878788, + "test_auc": 0.950612244897959, + "val_auc": 0.9493189102564104, + "k": 10, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 4201, + 337, + 5888, + 2900, + 1596, + 1769, + 7481, + 7384 + ] + }, + { + "test_f1": 0.8484848484848485, + "test_acc": 0.8484848484848485, + "test_auc": 0.9571428571428571, + "val_auc": 0.9551282051282051, + "k": 16, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 4201, + 337, + 5888, + 2900, + 1596, + 1769, + 7481, + 7384, + 5463, + 2000, + 2832, + 3526, + 1499, + 1473 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b8b1e66b4968684428c5574d746d0889ba539cf1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6366052590364633, + "test_acc": 0.6549631556133507, + "test_auc": 0.672278706291026, + "val_auc": 0.7712, + "k": 1, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.6568493890266051, + "test_acc": 0.6788036410923277, + "test_auc": 0.773700887294241, + "val_auc": 0.8582000000000001, + "k": 2, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388 + ] + }, + { + "test_f1": 0.7119112954011251, + "test_acc": 0.7121803207628955, + "test_auc": 0.8148353853484468, + "val_auc": 0.9079999999999999, + "k": 5, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 5352, + 789, + 337 + ] + }, + { + "test_f1": 0.7727867781389582, + "test_acc": 0.7732986562635458, + "test_auc": 0.8710108961476429, + "val_auc": 0.9348000000000001, + "k": 10, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 5352, + 789, + 337, + 2335, + 4201, + 7384, + 2000, + 2659 + ] + }, + { + "test_f1": 0.7967686616838353, + "test_acc": 0.7971391417425228, + "test_auc": 0.8958537820860659, + "val_auc": 0.9432, + "k": 16, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 5352, + 789, + 337, + 2335, + 4201, + 7384, + 2000, + 2659, + 7481, + 239, + 4889, + 3526, + 12946, + 2126 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..05aeb69d8715e3432a435f4de8e1e69e4fbc777b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6235120173365226, + "test_acc": 0.6410923276983095, + "test_auc": 0.6632118608527826, + "val_auc": 0.6864, + "k": 1, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.6444904977616862, + "test_acc": 0.6662332032943217, + "test_auc": 0.7737940809973531, + "val_auc": 0.8192, + "k": 2, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4201 + ] + }, + { + "test_f1": 0.7657999877906988, + "test_acc": 0.7667967056783702, + "test_auc": 0.8637030818556369, + "val_auc": 0.9208000000000001, + "k": 5, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4201, + 7384, + 3526, + 5388 + ] + }, + { + "test_f1": 0.7927895802757855, + "test_acc": 0.7936714347637625, + "test_auc": 0.8990967726419362, + "val_auc": 0.9396, + "k": 10, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4201, + 7384, + 3526, + 5388, + 7481, + 1924, + 3842, + 2900, + 1091 + ] + }, + { + "test_f1": 0.8067197798945919, + "test_acc": 0.8071087993064586, + "test_auc": 0.9076901339434014, + "val_auc": 0.9468, + "k": 16, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4201, + 7384, + 3526, + 5388, + 7481, + 1924, + 3842, + 2900, + 1091, + 2832, + 2050, + 6244, + 3825, + 1614, + 5352 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..872de083ee72291719b95874a04b56ff618795c2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8060494799220459, + "test_acc": 0.8071087993064586, + "test_auc": 0.8466091771747578, + "val_auc": 0.8736, + "k": 1, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8042977598726911, + "test_acc": 0.8071087993064586, + "test_auc": 0.8466091771747578, + "val_auc": 0.8736, + "k": 2, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889 + ] + }, + { + "test_f1": 0.833914344636456, + "test_acc": 0.835283918508886, + "test_auc": 0.9249557705691278, + "val_auc": 0.9339999999999999, + "k": 5, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 4201, + 789, + 12946 + ] + }, + { + "test_f1": 0.836800188628597, + "test_acc": 0.8383181621153013, + "test_auc": 0.9333311788552505, + "val_auc": 0.9319999999999999, + "k": 10, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 4201, + 789, + 12946, + 14154, + 5352, + 4531, + 3526, + 559 + ] + }, + { + "test_f1": 0.8235148464636332, + "test_acc": 0.8257477243172952, + "test_auc": 0.9169689199000122, + "val_auc": 0.958, + "k": 16, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 4201, + 789, + 12946, + 14154, + 5352, + 4531, + 3526, + 559, + 1091, + 2335, + 1851, + 1158, + 1924, + 10707 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ee3ef1f9855b9cf6772386685083609d4507bae0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9261351981351982, + "test_acc": 0.9261538461538461, + "test_auc": 0.9743368566331168, + "val_auc": 0.97796817625459, + "k": 1, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958 + ] + }, + { + "test_f1": 0.9579473908363049, + "test_acc": 0.9579487179487179, + "test_auc": 0.989964486484667, + "val_auc": 0.993064055487556, + "k": 2, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 5888 + ] + }, + { + "test_f1": 0.9651274714485318, + "test_acc": 0.9651282051282051, + "test_auc": 0.9922955868987107, + "val_auc": 0.9959200326397388, + "k": 5, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 5888, + 3818, + 508, + 2780 + ] + }, + { + "test_f1": 0.9712820512820513, + "test_acc": 0.9712820512820513, + "test_auc": 0.9941175480526474, + "val_auc": 0.9969400244798041, + "k": 10, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 5888, + 3818, + 508, + 2780, + 1279, + 386, + 5463, + 12605, + 4586 + ] + }, + { + "test_f1": 0.9743589204131313, + "test_acc": 0.9743589743589743, + "test_auc": 0.9948286599118054, + "val_auc": 0.9971440228478171, + "k": 16, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 5888, + 3818, + 508, + 2780, + 1279, + 386, + 5463, + 12605, + 4586, + 1596, + 1499, + 1158, + 730, + 3475, + 779 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..86516205a34fbbb63fc190177ff7b1e7852d08a2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9320373245537359, + "test_acc": 0.9323076923076923, + "test_auc": 0.9338161039485643, + "val_auc": 0.9767441860465116, + "k": 1, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065 + ] + }, + { + "test_f1": 0.9320193770330673, + "test_acc": 0.9323076923076923, + "test_auc": 0.9795797286834753, + "val_auc": 0.9942880456956344, + "k": 2, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508 + ] + }, + { + "test_f1": 0.934096799567259, + "test_acc": 0.9343589743589743, + "test_auc": 0.983659154408052, + "val_auc": 0.9979600163198694, + "k": 5, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 1279 + ] + }, + { + "test_f1": 0.9671676750773928, + "test_acc": 0.9671794871794872, + "test_auc": 0.9970966438886458, + "val_auc": 1.0, + "k": 10, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 1279, + 3958, + 3475, + 5406, + 992, + 1499 + ] + }, + { + "test_f1": 0.9723076340461818, + "test_acc": 0.9723076923076923, + "test_auc": 0.9976310297236342, + "val_auc": 1.0, + "k": 16, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 1279, + 3958, + 3475, + 5406, + 992, + 1499, + 581, + 2708, + 1079, + 5888, + 1482, + 19129 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..42b5609f2240b3b37ea08247944bc8a4862eeef6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9630650341644348, + "test_acc": 0.963076923076923, + "test_auc": 0.9825630322819537, + "val_auc": 0.9987760097919216, + "k": 1, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037 + ] + }, + { + "test_f1": 0.9651142602854548, + "test_acc": 0.9651282051282051, + "test_auc": 0.9857567240044435, + "val_auc": 0.9987760097919216, + "k": 2, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508 + ] + }, + { + "test_f1": 0.9753799018576346, + "test_acc": 0.9753846153846154, + "test_auc": 0.9882603426801765, + "val_auc": 1.0, + "k": 5, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 5472, + 1028, + 1253 + ] + }, + { + "test_f1": 0.9784590459118746, + "test_acc": 0.9784615384615385, + "test_auc": 0.9970377352139226, + "val_auc": 1.0, + "k": 10, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 5472, + 1028, + 1253, + 26034, + 1079, + 1279, + 778, + 3958 + ] + }, + { + "test_f1": 0.980510565348839, + "test_acc": 0.9805128205128205, + "test_auc": 0.9962340525801999, + "val_auc": 1.0, + "k": 16, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 5472, + 1028, + 1253, + 26034, + 1079, + 1279, + 778, + 3958, + 3537, + 493, + 3475, + 1132, + 581, + 305 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..eccccd708c50a92ade3b2c19a59b3830a62a7fba --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5748243776960462, + "test_acc": 0.5942782834850455, + "test_auc": 0.6070622789467909, + "val_auc": 0.5903999999999999, + "k": 1, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894 + ] + }, + { + "test_f1": 0.6411453720130356, + "test_acc": 0.6480277416558301, + "test_auc": 0.6805274763596135, + "val_auc": 0.6788, + "k": 2, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351 + ] + }, + { + "test_f1": 0.6806781087243194, + "test_acc": 0.683138274815778, + "test_auc": 0.7334303098991254, + "val_auc": 0.7934, + "k": 5, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351, + 1441, + 559, + 3078 + ] + }, + { + "test_f1": 0.703990221428529, + "test_acc": 0.7043779800606849, + "test_auc": 0.7829210514053461, + "val_auc": 0.8148000000000001, + "k": 10, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351, + 1441, + 559, + 3078, + 654, + 619, + 4510, + 1602, + 19018 + ] + }, + { + "test_f1": 0.71053675943087, + "test_acc": 0.7113133940182055, + "test_auc": 0.7899639400493927, + "val_auc": 0.8228000000000001, + "k": 16, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351, + 1441, + 559, + 3078, + 654, + 619, + 4510, + 1602, + 19018, + 52, + 1960, + 343, + 1158, + 3301, + 7595 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7823e88814c6ead12f94b9e7691b1cfd1372402d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5982168315902333, + "test_acc": 0.6276549631556133, + "test_auc": 0.6328987300103265, + "val_auc": 0.6168, + "k": 1, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274 + ] + }, + { + "test_f1": 0.600944716848428, + "test_acc": 0.6285218899003034, + "test_auc": 0.6434893676506618, + "val_auc": 0.6496, + "k": 2, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121 + ] + }, + { + "test_f1": 0.7015135906440662, + "test_acc": 0.7048114434330299, + "test_auc": 0.7484540367153127, + "val_auc": 0.7569999999999999, + "k": 5, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 4510, + 698 + ] + }, + { + "test_f1": 0.7422888841630148, + "test_acc": 0.7425227568270482, + "test_auc": 0.8207141042657163, + "val_auc": 0.8248, + "k": 10, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 4510, + 698, + 52, + 1602, + 3078, + 19018, + 1524 + ] + }, + { + "test_f1": 0.7685044615583553, + "test_acc": 0.7685305591677504, + "test_auc": 0.845616739392828, + "val_auc": 0.8492, + "k": 16, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 4510, + 698, + 52, + 1602, + 3078, + 19018, + 1524, + 1895, + 1894, + 649, + 1499, + 1158, + 2046 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bf4d6740f601977f510eca69cc316caf15a15aa7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.550815178122151, + "test_acc": 0.5786736020806242, + "test_auc": 0.5839431007348774, + "val_auc": 0.5364, + "k": 1, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.5555799878173495, + "test_acc": 0.5903771131339401, + "test_auc": 0.6545429675580696, + "val_auc": 0.6596, + "k": 2, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508 + ] + }, + { + "test_f1": 0.6438282326744897, + "test_acc": 0.6480277416558301, + "test_auc": 0.7091819095991017, + "val_auc": 0.7064, + "k": 5, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121 + ] + }, + { + "test_f1": 0.651306102885968, + "test_acc": 0.6519289120069354, + "test_auc": 0.7208596818487227, + "val_auc": 0.7816, + "k": 10, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 925, + 1441, + 52, + 957, + 628 + ] + }, + { + "test_f1": 0.6616510304222183, + "test_acc": 0.6623320329432163, + "test_auc": 0.7491642629204802, + "val_auc": 0.7687999999999999, + "k": 16, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 925, + 1441, + 52, + 957, + 628, + 7893, + 51, + 4654, + 4510, + 3274, + 567 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f826b0c9d07feb696aefc250e4d30862145f34ce --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6117072932281822, + "test_acc": 0.6119402985074627, + "test_auc": 0.6454446262956901, + "val_auc": 0.6684, + "k": 1, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048 + ] + }, + { + "test_f1": 0.6161313914982547, + "test_acc": 0.6162046908315565, + "test_auc": 0.688034188034188, + "val_auc": 0.6892, + "k": 2, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 1504 + ] + }, + { + "test_f1": 0.6565102762888453, + "test_acc": 0.6567164179104478, + "test_auc": 0.7465448263320604, + "val_auc": 0.7076, + "k": 5, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 1504, + 4672, + 927, + 1818 + ] + }, + { + "test_f1": 0.73381391366343, + "test_acc": 0.7356076759061834, + "test_auc": 0.8099108928896163, + "val_auc": 0.7888000000000001, + "k": 10, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 1504, + 4672, + 927, + 1818, + 696, + 741, + 988, + 1902, + 1927 + ] + }, + { + "test_f1": 0.719854355109232, + "test_acc": 0.720682302771855, + "test_auc": 0.8116930350972904, + "val_auc": 0.8151999999999999, + "k": 16, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1048, + 1504, + 4672, + 927, + 1818, + 696, + 741, + 988, + 1902, + 1927, + 437, + 7129, + 29085, + 499, + 2034, + 1974 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c604c086bdb706d6426d8da0c14ce30a944bc9c8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9393815708101421, + "test_acc": 0.9393939393939394, + "test_auc": 0.9604081632653062, + "val_auc": 0.9520933977455718, + "k": 1, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299 + ] + }, + { + "test_f1": 0.9494949494949495, + "test_acc": 0.9494949494949495, + "test_auc": 0.9681632653061223, + "val_auc": 0.9530998389694042, + "k": 2, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114 + ] + }, + { + "test_f1": 0.9494949494949495, + "test_acc": 0.9494949494949495, + "test_auc": 0.9946938775510203, + "val_auc": 0.967793880837359, + "k": 5, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 2535, + 432, + 23633 + ] + }, + { + "test_f1": 0.9696907835113407, + "test_acc": 0.9696969696969697, + "test_auc": 0.9967346938775511, + "val_auc": 0.9818840579710145, + "k": 10, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 2535, + 432, + 23633, + 15752, + 6877, + 13324, + 1500, + 373 + ] + }, + { + "test_f1": 0.9696907835113407, + "test_acc": 0.9696969696969697, + "test_auc": 0.9979591836734694, + "val_auc": 1.0, + "k": 16, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 2535, + 432, + 23633, + 15752, + 6877, + 13324, + 1500, + 373, + 27804, + 1232, + 19940, + 6055, + 7544, + 1610 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a76042c3bb9f61acf4ab6229555f4b056f1aa541 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7971097355838948, + "test_acc": 0.797979797979798, + "test_auc": 0.9424489795918367, + "val_auc": 0.9291465378421899, + "k": 1, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791 + ] + }, + { + "test_f1": 0.92920621648266, + "test_acc": 0.9292929292929293, + "test_auc": 0.9791836734693877, + "val_auc": 0.9681964573268921, + "k": 2, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 114 + ] + }, + { + "test_f1": 0.9090723505340225, + "test_acc": 0.9090909090909091, + "test_auc": 0.9861224489795919, + "val_auc": 0.9826892109500804, + "k": 5, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 114, + 23557, + 1657, + 429 + ] + }, + { + "test_f1": 0.9292784948597953, + "test_acc": 0.9292929292929293, + "test_auc": 0.9918367346938776, + "val_auc": 0.9895330112721418, + "k": 10, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 114, + 23557, + 1657, + 429, + 1396, + 253, + 278, + 521, + 4580 + ] + }, + { + "test_f1": 0.9797979797979798, + "test_acc": 0.9797979797979798, + "test_auc": 0.9983673469387756, + "val_auc": 0.9959742351046699, + "k": 16, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 114, + 23557, + 1657, + 429, + 1396, + 253, + 278, + 521, + 4580, + 1299, + 5387, + 1067, + 19000, + 4314, + 6088 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..79150cffdafa2ea662feb79606d2ea84eef5e827 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9292784948597953, + "test_acc": 0.9292929292929293, + "test_auc": 0.9551020408163265, + "val_auc": 0.9223027375201288, + "k": 1, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9595132999388319, + "test_acc": 0.9595959595959596, + "test_auc": 0.9975510204081632, + "val_auc": 0.9770531400966184, + "k": 2, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610 + ] + }, + { + "test_f1": 0.9696784111400831, + "test_acc": 0.9696969696969697, + "test_auc": 0.9975510204081632, + "val_auc": 0.9915458937198067, + "k": 5, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610, + 1244, + 23791, + 1501 + ] + }, + { + "test_f1": 0.9696784111400831, + "test_acc": 0.9696969696969697, + "test_auc": 0.9963265306122449, + "val_auc": 0.9975845410628019, + "k": 10, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610, + 1244, + 23791, + 1501, + 15752, + 177, + 114, + 1550, + 7577 + ] + }, + { + "test_f1": 0.979793856936714, + "test_acc": 0.9797979797979798, + "test_auc": 0.9983673469387755, + "val_auc": 1.0, + "k": 16, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1610, + 1244, + 23791, + 1501, + 15752, + 177, + 114, + 1550, + 7577, + 4580, + 1360, + 373, + 1960, + 907, + 7891 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ee4d71328682303322e8c006140e0bba802dcf4f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9154620271663052, + "test_acc": 0.9155555555555556, + "test_auc": 0.939396333754741, + "val_auc": 0.8969155844155844, + "k": 1, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752 + ] + }, + { + "test_f1": 0.9155555555555556, + "test_acc": 0.9155555555555556, + "test_auc": 0.9529867256637168, + "val_auc": 0.9249188311688311, + "k": 2, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309 + ] + }, + { + "test_f1": 0.9377777777777778, + "test_acc": 0.9377777777777778, + "test_auc": 0.962310366624526, + "val_auc": 0.9776785714285714, + "k": 5, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580 + ] + }, + { + "test_f1": 0.9377777777777778, + "test_acc": 0.9377777777777778, + "test_auc": 0.9655894437420985, + "val_auc": 0.9764610389610389, + "k": 10, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 3301, + 12400, + 969, + 5563, + 1915 + ] + }, + { + "test_f1": 0.951109179545599, + "test_acc": 0.9511111111111111, + "test_auc": 0.9736093552465233, + "val_auc": 0.9862012987012987, + "k": 16, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 3301, + 12400, + 969, + 5563, + 1915, + 2529, + 635, + 1541, + 943, + 26752, + 1073 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..89ca2aa838a8c1f79795bc56d3e1616d5aa31bcd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8, + "test_acc": 0.8, + "test_auc": 0.8427623261694059, + "val_auc": 0.8591720779220781, + "k": 1, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406 + ] + }, + { + "test_f1": 0.8261310195915296, + "test_acc": 0.8266666666666667, + "test_auc": 0.8790297092288243, + "val_auc": 0.8719561688311688, + "k": 2, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854 + ] + }, + { + "test_f1": 0.8397150997150998, + "test_acc": 0.84, + "test_auc": 0.9147835018963336, + "val_auc": 0.9025974025974026, + "k": 5, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1232, + 1570 + ] + }, + { + "test_f1": 0.844105778217604, + "test_acc": 0.8444444444444444, + "test_auc": 0.9264775600505689, + "val_auc": 0.890625, + "k": 10, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1232, + 1570, + 1182, + 311, + 6662, + 1804, + 322 + ] + }, + { + "test_f1": 0.8397150997150998, + "test_acc": 0.84, + "test_auc": 0.9255689001264222, + "val_auc": 0.8961038961038961, + "k": 16, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1232, + 1570, + 1182, + 311, + 6662, + 1804, + 322, + 15752, + 920, + 3481, + 800, + 748, + 629 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7b8b1a6be9461e457820c5880bdf03e013bee663 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9288888888888889, + "test_acc": 0.9288888888888889, + "test_auc": 0.9542114412136536, + "val_auc": 0.9362824675324676, + "k": 1, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9065337579029951, + "test_acc": 0.9066666666666666, + "test_auc": 0.9648388116308471, + "val_auc": 0.9549512987012987, + "k": 2, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16 + ] + }, + { + "test_f1": 0.9510917908760189, + "test_acc": 0.9511111111111111, + "test_auc": 0.9830910240202275, + "val_auc": 0.9646915584415585, + "k": 5, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 1835 + ] + }, + { + "test_f1": 0.9377630261840788, + "test_acc": 0.9377777777777778, + "test_auc": 0.9889380530973451, + "val_auc": 0.9772727272727272, + "k": 10, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 1835, + 15752, + 21467, + 1960, + 4114, + 969 + ] + }, + { + "test_f1": 0.955553799690968, + "test_acc": 0.9555555555555556, + "test_auc": 0.9913874841972187, + "val_auc": 0.979301948051948, + "k": 16, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 1835, + 15752, + 21467, + 1960, + 4114, + 969, + 7033, + 4580, + 1232, + 79, + 7095, + 1417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8df0c969b376faa6d5d1581e6562c01ba63871f2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6536151066024786, + "test_acc": 0.659297789336801, + "test_auc": 0.6787500319413902, + "val_auc": 0.7212000000000002, + "k": 1, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649 + ] + }, + { + "test_f1": 0.6895842674570253, + "test_acc": 0.6909406155179887, + "test_auc": 0.7383564238269242, + "val_auc": 0.7936, + "k": 2, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 943 + ] + }, + { + "test_f1": 0.7317719562130672, + "test_acc": 0.7325530992631123, + "test_auc": 0.7790899634891122, + "val_auc": 0.8248, + "k": 5, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 943, + 7614, + 5394, + 7246 + ] + }, + { + "test_f1": 0.7382215280193412, + "test_acc": 0.7394885132206328, + "test_auc": 0.8018070559658249, + "val_auc": 0.8714, + "k": 10, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 943, + 7614, + 5394, + 7246, + 1928, + 28053, + 1737, + 1560, + 15741 + ] + }, + { + "test_f1": 0.7455591199065063, + "test_acc": 0.7464239271781534, + "test_auc": 0.8205337293564674, + "val_auc": 0.87, + "k": 16, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 943, + 7614, + 5394, + 7246, + 1928, + 28053, + 1737, + 1560, + 15741, + 17225, + 253, + 584, + 2041, + 24358, + 1551 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c6dfa50c5910d020c9710e7880886d3ec22646aa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6128037168607927, + "test_acc": 0.612917208495882, + "test_auc": 0.6360462721767193, + "val_auc": 0.6567999999999999, + "k": 1, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943 + ] + }, + { + "test_f1": 0.6242075776103229, + "test_acc": 0.6250541829215431, + "test_auc": 0.6708105296859523, + "val_auc": 0.7102, + "k": 2, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253 + ] + }, + { + "test_f1": 0.6464072745712143, + "test_acc": 0.64716081491114, + "test_auc": 0.6991962794668719, + "val_auc": 0.752, + "k": 5, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 1011, + 7649, + 1551 + ] + }, + { + "test_f1": 0.6470325322562476, + "test_acc": 0.6480277416558301, + "test_auc": 0.7023742598992004, + "val_auc": 0.7612, + "k": 10, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 1011, + 7649, + 1551, + 5394, + 7614, + 28053, + 1928, + 319 + ] + }, + { + "test_f1": 0.6800405469046167, + "test_acc": 0.6801040312093628, + "test_auc": 0.7424080952259271, + "val_auc": 0.7812, + "k": 16, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 1011, + 7649, + 1551, + 5394, + 7614, + 28053, + 1928, + 319, + 844, + 1095, + 24358, + 32305, + 1411, + 17225 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7c57812968569e2a774fe12de2589f38f967d2f1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6040916496026179, + "test_acc": 0.6176853055916776, + "test_auc": 0.6411786899069718, + "val_auc": 0.6990000000000001, + "k": 1, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253 + ] + }, + { + "test_f1": 0.6138315941796622, + "test_acc": 0.6276549631556133, + "test_auc": 0.6976676021109878, + "val_auc": 0.7476, + "k": 2, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394 + ] + }, + { + "test_f1": 0.7241887112849789, + "test_acc": 0.7269180754226268, + "test_auc": 0.7978455720214466, + "val_auc": 0.7984, + "k": 5, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 844, + 1099, + 7649 + ] + }, + { + "test_f1": 0.7265070376972994, + "test_acc": 0.7282184655396619, + "test_auc": 0.8151841101729945, + "val_auc": 0.8180000000000001, + "k": 10, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 844, + 1099, + 7649, + 1541, + 7614, + 7246, + 794, + 100 + ] + }, + { + "test_f1": 0.7374440713541498, + "test_acc": 0.7386215864759428, + "test_auc": 0.8257435579852723, + "val_auc": 0.8227999999999999, + "k": 16, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 844, + 1099, + 7649, + 1541, + 7614, + 7246, + 794, + 100, + 1737, + 1960, + 28053, + 1857, + 7143, + 4481 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5bd1f0fe9703b024e0bb5b6ae5e9a290dada37da --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8222582601844816, + "test_acc": 0.8230508474576271, + "test_auc": 0.8517372487157707, + "val_auc": 0.8162238325281802, + "k": 1, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.8261429764161946, + "test_acc": 0.8271186440677966, + "test_auc": 0.8919252591440433, + "val_auc": 0.8675523349436394, + "k": 2, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 266 + ] + }, + { + "test_f1": 0.8254789300554494, + "test_acc": 0.8264406779661017, + "test_auc": 0.9107566380955533, + "val_auc": 0.8776167471819645, + "k": 5, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 266, + 5020, + 1847, + 6838 + ] + }, + { + "test_f1": 0.8634837265757646, + "test_acc": 0.8637288135593221, + "test_auc": 0.9368273194265185, + "val_auc": 0.9259259259259259, + "k": 10, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 266, + 5020, + 1847, + 6838, + 508, + 1528, + 32278, + 1697, + 776 + ] + }, + { + "test_f1": 0.8765386687619262, + "test_acc": 0.8766101694915254, + "test_auc": 0.9520358297205768, + "val_auc": 0.9363929146537843, + "k": 16, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 266, + 5020, + 1847, + 6838, + 508, + 1528, + 32278, + 1697, + 776, + 1103, + 1628, + 431, + 4413, + 1180, + 1758 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c754f3befb11d3cfdafed5f825a023982d258d53 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8299751867116448, + "test_acc": 0.8311864406779661, + "test_auc": 0.8507453493802238, + "val_auc": 0.806964573268921, + "k": 1, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528 + ] + }, + { + "test_f1": 0.8306774552334435, + "test_acc": 0.831864406779661, + "test_auc": 0.8514338874732031, + "val_auc": 0.822463768115942, + "k": 2, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791 + ] + }, + { + "test_f1": 0.819467238945249, + "test_acc": 0.8203389830508474, + "test_auc": 0.8946545910506595, + "val_auc": 0.8663446054750402, + "k": 5, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1628, + 1815, + 203 + ] + }, + { + "test_f1": 0.8383054719522601, + "test_acc": 0.8386440677966102, + "test_auc": 0.919767386276305, + "val_auc": 0.877818035426731, + "k": 10, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1628, + 1815, + 203, + 1866, + 1103, + 2023, + 6362, + 416 + ] + }, + { + "test_f1": 0.8395100559504005, + "test_acc": 0.84, + "test_auc": 0.9205818652487747, + "val_auc": 0.8824476650563606, + "k": 16, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1628, + 1815, + 203, + 1866, + 1103, + 2023, + 6362, + 416, + 31149, + 4185, + 1180, + 1544, + 461, + 4558 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..02a7c12f16204abe261276ac8050e0f24996e3a9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.660219148848853, + "test_acc": 0.6650847457627118, + "test_auc": 0.6985526910900045, + "val_auc": 0.7419484702093397, + "k": 1, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020 + ] + }, + { + "test_f1": 0.7121022425725392, + "test_acc": 0.7138983050847457, + "test_auc": 0.7853820329248069, + "val_auc": 0.8478260869565217, + "k": 2, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266 + ] + }, + { + "test_f1": 0.7319595247180118, + "test_acc": 0.7328813559322034, + "test_auc": 0.8093420554286953, + "val_auc": 0.8919082125603864, + "k": 5, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 431, + 895 + ] + }, + { + "test_f1": 0.8002951317326523, + "test_acc": 0.8013559322033899, + "test_auc": 0.8679937342114261, + "val_auc": 0.9210950080515299, + "k": 10, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 431, + 895, + 1628, + 373, + 32278, + 27673, + 1103 + ] + }, + { + "test_f1": 0.8281560953720253, + "test_acc": 0.8284745762711865, + "test_auc": 0.9077800208124198, + "val_auc": 0.9545088566827697, + "k": 16, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 431, + 895, + 1628, + 373, + 32278, + 27673, + 1103, + 1528, + 5460, + 461, + 2893, + 18827, + 582 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..371327b05b38c058e4b7d7f1310cf7a02e159b78 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.58152537452829, + "test_acc": 0.5860058309037901, + "test_auc": 0.5925132598939209, + "val_auc": 0.5372148859543818, + "k": 1, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381 + ] + }, + { + "test_f1": 0.5448401657747003, + "test_acc": 0.5539358600583091, + "test_auc": 0.6090881272949816, + "val_auc": 0.5512204881952781, + "k": 2, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219 + ] + }, + { + "test_f1": 0.6032787640136109, + "test_acc": 0.6064139941690962, + "test_auc": 0.6374269005847953, + "val_auc": 0.5942376950780311, + "k": 5, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219, + 1122, + 275, + 1799 + ] + }, + { + "test_f1": 0.6260594285015119, + "test_acc": 0.6268221574344023, + "test_auc": 0.6888854889160887, + "val_auc": 0.6688675470188075, + "k": 10, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219, + 1122, + 275, + 1799, + 799, + 993, + 304, + 5897, + 1411 + ] + }, + { + "test_f1": 0.6849812578092461, + "test_acc": 0.685131195335277, + "test_auc": 0.734326125390997, + "val_auc": 0.7583033213285314, + "k": 16, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219, + 1122, + 275, + 1799, + 799, + 993, + 304, + 5897, + 1411, + 673, + 1697, + 1039, + 801, + 1508, + 1152 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..87276313d61f6519fdb74fb7d13f97fb10077ea0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7055714675080922, + "test_acc": 0.7230320699708455, + "test_auc": 0.729158166734666, + "val_auc": 0.6766706682673069, + "k": 1, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7822616647561319, + "test_acc": 0.7871720116618076, + "test_auc": 0.8008975928192575, + "val_auc": 0.7911164465786314, + "k": 2, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275 + ] + }, + { + "test_f1": 0.827721796491067, + "test_acc": 0.8279883381924198, + "test_auc": 0.8826159390724875, + "val_auc": 0.875750300120048, + "k": 5, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 534, + 191, + 8006 + ] + }, + { + "test_f1": 0.8539667068820793, + "test_acc": 0.8542274052478134, + "test_auc": 0.8951788385692915, + "val_auc": 0.8981592637054823, + "k": 10, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 534, + 191, + 8006, + 1928, + 1122, + 799, + 635, + 433 + ] + }, + { + "test_f1": 0.894800479350373, + "test_acc": 0.8950437317784257, + "test_auc": 0.947436420508636, + "val_auc": 0.9319727891156462, + "k": 16, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 534, + 191, + 8006, + 1928, + 1122, + 799, + 635, + 433, + 506, + 801, + 786, + 1411, + 1799, + 1960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7cc1c210ad92fce4a6787bf9902e14ed9132cfa0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6719077629760518, + "test_acc": 0.6793002915451894, + "test_auc": 0.6823405412756699, + "val_auc": 0.6882753101240495, + "k": 1, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.6724114650976003, + "test_acc": 0.673469387755102, + "test_auc": 0.7450020399836801, + "val_auc": 0.7394957983193278, + "k": 2, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191 + ] + }, + { + "test_f1": 0.7667043615184145, + "test_acc": 0.7667638483965015, + "test_auc": 0.8423432612539099, + "val_auc": 0.8175270108043218, + "k": 5, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 1239, + 534, + 5897 + ] + }, + { + "test_f1": 0.8016987072906524, + "test_acc": 0.8017492711370262, + "test_auc": 0.8824799401604787, + "val_auc": 0.8887555022008803, + "k": 10, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 1239, + 534, + 5897, + 433, + 275, + 8006, + 635, + 1195 + ] + }, + { + "test_f1": 0.8163265306122449, + "test_acc": 0.8163265306122449, + "test_auc": 0.8857609139126886, + "val_auc": 0.8879551820728291, + "k": 16, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 1239, + 534, + 5897, + 433, + 275, + 8006, + 635, + 1195, + 1464, + 3233, + 4003, + 1596, + 167, + 1411 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6e58a0c37412dd0be6aedf66e41e79f8890c6a7a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320 + ] + }, + { + "test_f1": 0.5148544266191325, + "test_acc": 0.5151515151515151, + "test_auc": 0.5285714285714286, + "val_auc": 0.5365656565656566, + "k": 2, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 5, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 987, + 1073, + 1600 + ] + }, + { + "test_f1": 0.45450233727172823, + "test_acc": 0.46464646464646464, + "test_auc": 0.47346938775510206, + "val_auc": 0.5733333333333333, + "k": 10, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 987, + 1073, + 1600, + 5883, + 6804, + 859, + 20078, + 4264 + ] + }, + { + "test_f1": 0.4736728992048141, + "test_acc": 0.47474747474747475, + "test_auc": 0.5087755102040816, + "val_auc": 0.5701010101010101, + "k": 16, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 987, + 1073, + 1600, + 5883, + 6804, + 859, + 20078, + 4264, + 1030, + 1408, + 473, + 1417, + 1848, + 7666 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ef349fe0fb1fae076be27160396fd2a41f2d1334 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5651239012506258, + "test_acc": 0.5656565656565656, + "test_auc": 0.5575510204081633, + "val_auc": 0.5868686868686869, + "k": 1, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520 + ] + }, + { + "test_f1": 0.5307778641111974, + "test_acc": 0.5454545454545454, + "test_auc": 0.5448979591836735, + "val_auc": 0.6173737373737374, + "k": 2, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534 + ] + }, + { + "test_f1": 0.572615039281706, + "test_acc": 0.5757575757575758, + "test_auc": 0.5995918367346939, + "val_auc": 0.6533333333333333, + "k": 5, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534, + 5883, + 5653, + 987 + ] + }, + { + "test_f1": 0.5616269904487614, + "test_acc": 0.5656565656565656, + "test_auc": 0.5828571428571429, + "val_auc": 0.6816161616161616, + "k": 10, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534, + 5883, + 5653, + 987, + 5129, + 2046, + 1788, + 320, + 1116 + ] + }, + { + "test_f1": 0.572615039281706, + "test_acc": 0.5757575757575758, + "test_auc": 0.596734693877551, + "val_auc": 0.687070707070707, + "k": 16, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534, + 5883, + 5653, + 987, + 5129, + 2046, + 1788, + 320, + 1116, + 1030, + 2489, + 7666, + 1818, + 1707, + 7026 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..92b0703258869b37cbd8b71f600aa5ffa10fb9c5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5959595959595959, + "test_acc": 0.5959595959595959, + "test_auc": 0.5906122448979592, + "val_auc": 0.5985858585858586, + "k": 1, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666 + ] + }, + { + "test_f1": 0.6059801856474307, + "test_acc": 0.6060606060606061, + "test_auc": 0.5926530612244898, + "val_auc": 0.6084848484848485, + "k": 2, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245 + ] + }, + { + "test_f1": 0.5347841156664687, + "test_acc": 0.5353535353535354, + "test_auc": 0.5681632653061225, + "val_auc": 0.6464646464646465, + "k": 5, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 7026 + ] + }, + { + "test_f1": 0.5752376708259062, + "test_acc": 0.5757575757575758, + "test_auc": 0.5957142857142858, + "val_auc": 0.6731313131313131, + "k": 10, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 7026, + 27174, + 111, + 4985, + 1116, + 2595 + ] + }, + { + "test_f1": 0.5954644484056248, + "test_acc": 0.5959595959595959, + "test_auc": 0.6191836734693876, + "val_auc": 0.6505050505050505, + "k": 16, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 7026, + 27174, + 111, + 4985, + 1116, + 2595, + 373, + 5883, + 680, + 1717, + 1804, + 4520 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2d1265b23251d3aa92082e8a9cd74eb196eaba33 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5087612590799031, + "test_acc": 0.5545762711864407, + "test_auc": 0.5562174346302486, + "val_auc": 0.623389694041868, + "k": 1, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217 + ] + }, + { + "test_f1": 0.5372619847566739, + "test_acc": 0.5647457627118644, + "test_auc": 0.5699863579368494, + "val_auc": 0.6376811594202898, + "k": 2, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 539 + ] + }, + { + "test_f1": 0.5822957522109541, + "test_acc": 0.5871186440677966, + "test_auc": 0.6144039595077091, + "val_auc": 0.5927938808373591, + "k": 5, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 539, + 1541, + 1960, + 1319 + ] + }, + { + "test_f1": 0.5670311310577746, + "test_acc": 0.568135593220339, + "test_auc": 0.6078256169264542, + "val_auc": 0.605475040257649, + "k": 10, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 539, + 1541, + 1960, + 1319, + 186, + 311, + 20, + 490, + 26270 + ] + }, + { + "test_f1": 0.5571454004982404, + "test_acc": 0.5579661016949152, + "test_auc": 0.5957316153894239, + "val_auc": 0.6276167471819646, + "k": 16, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 539, + 1541, + 1960, + 1319, + 186, + 311, + 20, + 490, + 26270, + 443, + 1812, + 253, + 836, + 373, + 3780 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6f84351c2cbcd34996432b98b3071c6658880318 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6155037963289954, + "test_acc": 0.615593220338983, + "test_auc": 0.6376303993704795, + "val_auc": 0.6260064412238325, + "k": 1, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319 + ] + }, + { + "test_f1": 0.6514292928112215, + "test_acc": 0.6515254237288136, + "test_auc": 0.699952749188279, + "val_auc": 0.6646537842190016, + "k": 2, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541 + ] + }, + { + "test_f1": 0.6742101720738735, + "test_acc": 0.6752542372881356, + "test_auc": 0.7183796464830319, + "val_auc": 0.6912238325281804, + "k": 5, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1960, + 1874 + ] + }, + { + "test_f1": 0.6660492528041217, + "test_acc": 0.6664406779661017, + "test_auc": 0.7260611208554419, + "val_auc": 0.7306763285024154, + "k": 10, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1960, + 1874, + 186, + 1255, + 1335, + 908, + 253 + ] + }, + { + "test_f1": 0.6707320049835476, + "test_acc": 0.671864406779661, + "test_auc": 0.7343713803488103, + "val_auc": 0.7411433172302737, + "k": 16, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1960, + 1874, + 186, + 1255, + 1335, + 908, + 253, + 217, + 1791, + 765, + 311, + 100, + 821 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f28e4814e608ae5e95e15c6c58d1d7e6a2c46d75 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5945427716849451, + "test_acc": 0.6020338983050847, + "test_auc": 0.6124100487951962, + "val_auc": 0.6268115942028987, + "k": 1, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.5911101339136382, + "test_acc": 0.6, + "test_auc": 0.6124100487951962, + "val_auc": 0.6268115942028987, + "k": 2, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255 + ] + }, + { + "test_f1": 0.604996268532953, + "test_acc": 0.6081355932203389, + "test_auc": 0.6336536092633654, + "val_auc": 0.6572061191626408, + "k": 5, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 521, + 12485 + ] + }, + { + "test_f1": 0.6013114829737325, + "test_acc": 0.6033898305084746, + "test_auc": 0.6369391034480223, + "val_auc": 0.7037037037037037, + "k": 10, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 521, + 12485, + 186, + 567, + 4511, + 1942, + 1791 + ] + }, + { + "test_f1": 0.6175112614599376, + "test_acc": 0.6189830508474576, + "test_auc": 0.6484539607946962, + "val_auc": 0.680354267310789, + "k": 16, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 521, + 12485, + 186, + 567, + 4511, + 1942, + 1791, + 1319, + 311, + 2277, + 6212, + 1500, + 253 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a226a1e000b7f7ccb4142c1b2201913384d1681e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6128679388276006, + "test_acc": 0.6153846153846154, + "test_auc": 0.6227279444362348, + "val_auc": 0.5990396158463385, + "k": 1, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1268 + ] + }, + { + "test_f1": 0.5797603208989309, + "test_acc": 0.5806451612903226, + "test_auc": 0.6366927737549875, + "val_auc": 0.5884353741496599, + "k": 2, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1268, + 1067 + ] + }, + { + "test_f1": 0.6600370684354273, + "test_acc": 0.6600496277915633, + "test_auc": 0.6696837594207181, + "val_auc": 0.6814725890356143, + "k": 5, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1268, + 1067, + 5731, + 1674, + 24630 + ] + }, + { + "test_f1": 0.5897977587503294, + "test_acc": 0.5905707196029777, + "test_auc": 0.6511501896458303, + "val_auc": 0.7182873149259703, + "k": 10, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1268, + 1067, + 5731, + 1674, + 24630, + 1441, + 8006, + 1101, + 1619, + 1551 + ] + }, + { + "test_f1": 0.6128030899077319, + "test_acc": 0.6129032258064516, + "test_auc": 0.6619008915816954, + "val_auc": 0.7214885954381752, + "k": 16, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1268, + 1067, + 5731, + 1674, + 24630, + 1441, + 8006, + 1101, + 1619, + 1551, + 124, + 2047, + 10, + 8013, + 1070, + 2480 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5c7cd332e013307cf6ccf884023183dafb692164 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5829827659882735, + "test_acc": 0.5831265508684863, + "test_auc": 0.6067435101719126, + "val_auc": 0.595438175270108, + "k": 1, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28711 + ] + }, + { + "test_f1": 0.612531017369727, + "test_acc": 0.6129032258064516, + "test_auc": 0.6509777843455987, + "val_auc": 0.6482593037214885, + "k": 2, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28711, + 575 + ] + }, + { + "test_f1": 0.6072826610683872, + "test_acc": 0.607940446650124, + "test_auc": 0.6431949165065761, + "val_auc": 0.7184873949579832, + "k": 5, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28711, + 575, + 1625, + 583, + 213 + ] + }, + { + "test_f1": 0.5902376553869582, + "test_acc": 0.5905707196029777, + "test_auc": 0.6417171567903059, + "val_auc": 0.7070828331332534, + "k": 10, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28711, + 575, + 1625, + 583, + 213, + 1619, + 1551, + 1799, + 765, + 822 + ] + }, + { + "test_f1": 0.6227312158075337, + "test_acc": 0.6228287841191067, + "test_auc": 0.6643268804492389, + "val_auc": 0.7531012404961984, + "k": 16, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28711, + 575, + 1625, + 583, + 213, + 1619, + 1551, + 1799, + 765, + 822, + 1381, + 680, + 898, + 10, + 802, + 1653 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..02dcf3e9f2969746f33fb5ba13ed1ac98c33af83 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5498150265125576, + "test_acc": 0.5508684863523573, + "test_auc": 0.5749839909364072, + "val_auc": 0.5600240096038416, + "k": 1, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619 + ] + }, + { + "test_f1": 0.608323296812643, + "test_acc": 0.6104218362282878, + "test_auc": 0.6367912910694055, + "val_auc": 0.6222488995598239, + "k": 2, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 32279 + ] + }, + { + "test_f1": 0.6435015343162936, + "test_acc": 0.6451612903225806, + "test_auc": 0.6981060046303138, + "val_auc": 0.6730692276910765, + "k": 5, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 32279, + 1042, + 125, + 575 + ] + }, + { + "test_f1": 0.6549678727722499, + "test_acc": 0.6550868486352357, + "test_auc": 0.7366385892320575, + "val_auc": 0.7006802721088434, + "k": 10, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 32279, + 1042, + 125, + 575, + 1067, + 1662, + 1625, + 2536, + 772 + ] + }, + { + "test_f1": 0.6767385529735515, + "test_acc": 0.6774193548387096, + "test_auc": 0.7308876410029063, + "val_auc": 0.7086834733893557, + "k": 16, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 32279, + 1042, + 125, + 575, + 1067, + 1662, + 1625, + 2536, + 772, + 1547, + 7319, + 1872, + 10, + 2480, + 898 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e534cb21eed3d3f733be21c12c6c64033de29272 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8456415445895746, + "test_acc": 0.8473895582329317, + "test_auc": 0.8591274677065562, + "val_auc": 0.8428282828282829, + "k": 1, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731 + ] + }, + { + "test_f1": 0.844332095788144, + "test_acc": 0.8460508701472557, + "test_auc": 0.8957864403377729, + "val_auc": 0.8997979797979798, + "k": 2, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226 + ] + }, + { + "test_f1": 0.8590571119820648, + "test_acc": 0.8607764390896921, + "test_auc": 0.9519361729580939, + "val_auc": 0.958989898989899, + "k": 5, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 6567, + 4148 + ] + }, + { + "test_f1": 0.865654450661144, + "test_acc": 0.8661311914323963, + "test_auc": 0.9513411994093275, + "val_auc": 0.958989898989899, + "k": 10, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 6567, + 4148, + 1158, + 5872, + 737, + 23697, + 805 + ] + }, + { + "test_f1": 0.8603243430184104, + "test_acc": 0.8607764390896921, + "test_auc": 0.9515777551576321, + "val_auc": 0.9618181818181819, + "k": 16, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 6567, + 4148, + 1158, + 5872, + 737, + 23697, + 805, + 1274, + 13701, + 1629, + 23003, + 508, + 292 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4191360b19587689e45c23cbfa4c031851144f8f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.868382752793685, + "test_acc": 0.8701472556894244, + "test_auc": 0.8733208126048371, + "val_auc": 0.8888888888888888, + "k": 1, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226 + ] + }, + { + "test_f1": 0.868382752793685, + "test_acc": 0.8701472556894244, + "test_auc": 0.9125890668234148, + "val_auc": 0.9236363636363636, + "k": 2, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731 + ] + }, + { + "test_f1": 0.8854094811361028, + "test_acc": 0.8862115127175368, + "test_auc": 0.9657854367679316, + "val_auc": 0.9668686868686869, + "k": 5, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4648, + 4505, + 32 + ] + }, + { + "test_f1": 0.8801713689052828, + "test_acc": 0.8808567603748326, + "test_auc": 0.9660363292282548, + "val_auc": 0.9658585858585857, + "k": 10, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4648, + 4505, + 32, + 1065, + 1425, + 6567, + 1447, + 22775 + ] + }, + { + "test_f1": 0.8842793556550892, + "test_acc": 0.8848728246318608, + "test_auc": 0.9674054852260182, + "val_auc": 0.9654545454545455, + "k": 16, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4648, + 4505, + 32, + 1065, + 1425, + 6567, + 1447, + 22775, + 7481, + 805, + 1682, + 1627, + 2046, + 1274 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8e9a70f28acbd2649a9e52eb0285ed35a09fa8cb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8702298228159072, + "test_acc": 0.8714859437751004, + "test_auc": 0.8845321214032773, + "val_auc": 0.8680808080808081, + "k": 1, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648 + ] + }, + { + "test_f1": 0.8743183396998658, + "test_acc": 0.8755020080321285, + "test_auc": 0.9173775286375823, + "val_auc": 0.8848484848484849, + "k": 2, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505 + ] + }, + { + "test_f1": 0.9022091725287882, + "test_acc": 0.9022757697456493, + "test_auc": 0.9765308024257717, + "val_auc": 0.9757575757575758, + "k": 5, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32 + ] + }, + { + "test_f1": 0.8914588881706655, + "test_acc": 0.891566265060241, + "test_auc": 0.9732692004415706, + "val_auc": 0.9793939393939395, + "k": 10, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32, + 1682, + 1305, + 6828, + 2832, + 22848 + ] + }, + { + "test_f1": 0.8981973091728039, + "test_acc": 0.8982597054886211, + "test_auc": 0.9761688004473054, + "val_auc": 0.9785858585858587, + "k": 16, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32, + 1682, + 1305, + 6828, + 2832, + 22848, + 1101, + 1065, + 6369, + 7481, + 2302, + 5872 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3f8b31235859c7de43692dad5f2b3649a99f19e0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.75539064325505, + "test_acc": 0.7661538461538462, + "test_auc": 0.7701635978052311, + "val_auc": 0.755609955120359, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518 + ] + }, + { + "test_f1": 0.7648187557380826, + "test_acc": 0.7733333333333333, + "test_auc": 0.7904702595347898, + "val_auc": 0.7690738474092208, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541 + ] + }, + { + "test_f1": 0.790259031333786, + "test_acc": 0.7938461538461539, + "test_auc": 0.8873918605042582, + "val_auc": 0.8600571195430436, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 1335 + ] + }, + { + "test_f1": 0.804850881831384, + "test_acc": 0.8071794871794872, + "test_auc": 0.9059333658733632, + "val_auc": 0.879232966136271, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 1335, + 821, + 253, + 1030, + 11644, + 7813 + ] + }, + { + "test_f1": 0.824111913335561, + "test_acc": 0.8266666666666667, + "test_auc": 0.9145319285017, + "val_auc": 0.9020807833537331, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 186, + 1791, + 1335, + 821, + 253, + 1030, + 11644, + 7813, + 549, + 505, + 2581, + 1417, + 1438, + 2203 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..be934205fbd3556eb2f134557400c973e068250a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8520074843862572, + "test_acc": 0.8544537815126051, + "test_auc": 0.8580940733670304, + "val_auc": 0.8454435969490165, + "k": 1, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852 + ] + }, + { + "test_f1": 0.8919212984560353, + "test_acc": 0.892436974789916, + "test_auc": 0.9055998763477016, + "val_auc": 0.9321557607386591, + "k": 2, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782 + ] + }, + { + "test_f1": 0.9051491031892003, + "test_acc": 0.9055462184873949, + "test_auc": 0.9620519411964624, + "val_auc": 0.9787234042553191, + "k": 5, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 1275 + ] + }, + { + "test_f1": 0.9072290591601063, + "test_acc": 0.907563025210084, + "test_auc": 0.9670536224338534, + "val_auc": 0.9747089522280209, + "k": 10, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 1275, + 173, + 8815, + 11077, + 1688, + 1152 + ] + }, + { + "test_f1": 0.914963514519732, + "test_acc": 0.9152941176470588, + "test_auc": 0.9762398673811022, + "val_auc": 0.9831393014853472, + "k": 16, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 1275, + 173, + 8815, + 11077, + 1688, + 1152, + 6522, + 1283, + 259, + 6919, + 42, + 98 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e0070a635bec0dfe2dacd86d4ee077d2e8ee7ae4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7485620248953516, + "test_acc": 0.7549579831932773, + "test_auc": 0.7613451887686111, + "val_auc": 0.7482938578883982, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021 + ] + }, + { + "test_f1": 0.7705129159515133, + "test_acc": 0.7710924369747899, + "test_auc": 0.8771806372070489, + "val_auc": 0.8779606583701324, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914 + ] + }, + { + "test_f1": 0.7764378927494832, + "test_acc": 0.7778151260504201, + "test_auc": 0.891509118453117, + "val_auc": 0.9169008430349257, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1180, + 1493, + 4790 + ] + }, + { + "test_f1": 0.8055327032639958, + "test_acc": 0.8060504201680673, + "test_auc": 0.9029975739563675, + "val_auc": 0.9289441991168206, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1180, + 1493, + 4790, + 6919, + 1794, + 1067, + 1275, + 1091 + ] + }, + { + "test_f1": 0.8305394108334245, + "test_acc": 0.8305882352941176, + "test_auc": 0.9278035085435784, + "val_auc": 0.9706945002007226, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1180, + 1493, + 4790, + 6919, + 1794, + 1067, + 1275, + 1091, + 98, + 1500, + 1152, + 8815, + 2962, + 984 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1ff1e46164202a20602c0367fef8ce4c6239badb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7559531538029478, + "test_acc": 0.7643697478991597, + "test_auc": 0.7676421459097122, + "val_auc": 0.735046166198314, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235 + ] + }, + { + "test_f1": 0.7516353709421998, + "test_acc": 0.760672268907563, + "test_auc": 0.7942508912365953, + "val_auc": 0.7782015254917702, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 8189 + ] + }, + { + "test_f1": 0.8158229741086848, + "test_acc": 0.8181512605042017, + "test_auc": 0.8247025294487711, + "val_auc": 0.7916499397832194, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 8189, + 5189, + 692, + 764 + ] + }, + { + "test_f1": 0.8175337820100255, + "test_acc": 0.8194957983193277, + "test_auc": 0.908387476408443, + "val_auc": 0.9090726615816941, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 8189, + 5189, + 692, + 764, + 42, + 3494, + 2046, + 1521, + 30916 + ] + }, + { + "test_f1": 0.8247547943865424, + "test_acc": 0.8265546218487395, + "test_auc": 0.9264580666854675, + "val_auc": 0.9193095142513048, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 8189, + 5189, + 692, + 764, + 42, + 3494, + 2046, + 1521, + 30916, + 24178, + 964, + 1980, + 1021, + 4153, + 576 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..55e33a5bf0dc01c472033367715e9a2f9c822d96 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9391594576408436, + "test_acc": 0.9391596638655462, + "test_auc": 0.982403500589337, + "val_auc": 0.9827378562826175, + "k": 1, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9835286301295431, + "test_acc": 0.9835294117647059, + "test_auc": 0.9992181342242084, + "val_auc": 0.9991971095945404, + "k": 2, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847 + ] + }, + { + "test_f1": 0.9983193239332243, + "test_acc": 0.9983193277310924, + "test_auc": 0.9998924369626367, + "val_auc": 1.0, + "k": 5, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 461, + 279 + ] + }, + { + "test_f1": 0.9983193239332243, + "test_acc": 0.9983193277310924, + "test_auc": 0.9998097309297062, + "val_auc": 1.0, + "k": 10, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 461, + 279, + 1411, + 786, + 1697, + 431, + 1758 + ] + }, + { + "test_f1": 0.9986554603618987, + "test_acc": 0.998655462184874, + "test_auc": 0.9999764988321728, + "val_auc": 1.0, + "k": 16, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 461, + 279, + 1411, + 786, + 1697, + 431, + 1758, + 5020, + 27673, + 1634, + 1360, + 126, + 4148 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6fe4a514ed319ce37f87a92373df0141aa83b581 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9334425205239285, + "test_acc": 0.9334453781512605, + "test_auc": 0.9837225036336421, + "val_auc": 0.9847450822962666, + "k": 1, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9603276537290333, + "test_acc": 0.9603361344537815, + "test_auc": 0.9951113051463942, + "val_auc": 0.9967884383781613, + "k": 2, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9619929736902902, + "test_acc": 0.9620168067226891, + "test_auc": 0.9957508080786168, + "val_auc": 0.9971898835808912, + "k": 5, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758 + ] + }, + { + "test_f1": 0.9956301750789618, + "test_acc": 0.9956302521008403, + "test_auc": 0.9998752630323015, + "val_auc": 1.0, + "k": 10, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758, + 431, + 1847, + 1628, + 279, + 461 + ] + }, + { + "test_f1": 0.9989915964107838, + "test_acc": 0.9989915966386554, + "test_auc": 0.999996836381254, + "val_auc": 1.0, + "k": 16, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758, + 431, + 1847, + 1628, + 279, + 461, + 126, + 1551, + 1634, + 27673, + 1914, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..96213905315ab62a8c6d35f83654768153aae6b4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9179806726304758, + "test_acc": 0.9179831932773109, + "test_auc": 0.973802525110094, + "val_auc": 0.9747089522280209, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9613373737537765, + "test_acc": 0.9613445378151261, + "test_auc": 0.9954963627423331, + "val_auc": 0.9967884383781613, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9959663263956778, + "test_acc": 0.9959663865546219, + "test_auc": 0.999599124310331, + "val_auc": 1.0, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 461 + ] + }, + { + "test_f1": 0.9966386212811212, + "test_acc": 0.9966386554621849, + "test_auc": 0.9997197937682134, + "val_auc": 1.0, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 461, + 1411, + 279, + 1758, + 27673, + 431 + ] + }, + { + "test_f1": 0.9979831864411416, + "test_acc": 0.9979831932773109, + "test_auc": 0.9999909610892972, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 461, + 1411, + 279, + 1758, + 27673, + 431, + 4148, + 1360, + 1634, + 1697, + 1914, + 1528 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8158e6d700b5163006b2aacaac0ae2080aa253c4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.763613326009654, + "test_acc": 0.7637735849056604, + "test_auc": 0.8254117052592354, + "val_auc": 0.8923569427771109, + "k": 1, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585 + ] + }, + { + "test_f1": 0.8206751939585916, + "test_acc": 0.8208805031446541, + "test_auc": 0.8857895738801203, + "val_auc": 0.9351740696278511, + "k": 2, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585, + 559 + ] + }, + { + "test_f1": 0.8379209507457275, + "test_acc": 0.8379874213836478, + "test_auc": 0.9138536553999386, + "val_auc": 0.9719887955182074, + "k": 5, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585, + 559, + 20310, + 776, + 32700 + ] + }, + { + "test_f1": 0.8618578885647352, + "test_acc": 0.8618867924528302, + "test_auc": 0.9315006799731455, + "val_auc": 0.969187675070028, + "k": 10, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585, + 559, + 20310, + 776, + 32700, + 1866, + 1327, + 1499, + 508, + 1608 + ] + }, + { + "test_f1": 0.8668622725129405, + "test_acc": 0.8669182389937107, + "test_auc": 0.9346805543882318, + "val_auc": 0.9635854341736695, + "k": 16, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1585, + 559, + 20310, + 776, + 32700, + 1866, + 1327, + 1499, + 508, + 1608, + 30737, + 7976, + 400, + 322, + 3078, + 272 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..28264d79e54b5c17e07a09cc52dc4eee91e3c45c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6392368644696259, + "test_acc": 0.6402535657686212, + "test_auc": 0.688949166164356, + "val_auc": 0.7531012404961984, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21219 + ] + }, + { + "test_f1": 0.6814243987391373, + "test_acc": 0.6814580031695721, + "test_auc": 0.7445850914205344, + "val_auc": 0.7999199679871949, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21219, + 776 + ] + }, + { + "test_f1": 0.7399383353763104, + "test_acc": 0.7400950871632329, + "test_auc": 0.8134116937914407, + "val_auc": 0.8399359743897559, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21219, + 776, + 913, + 1327, + 431 + ] + }, + { + "test_f1": 0.7875554300018741, + "test_acc": 0.7876386687797148, + "test_auc": 0.8630751456700824, + "val_auc": 0.8877551020408163, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21219, + 776, + 913, + 1327, + 431, + 508, + 21257, + 1866, + 28521, + 5186 + ] + }, + { + "test_f1": 0.804953742373064, + "test_acc": 0.8050713153724247, + "test_auc": 0.8855887080570625, + "val_auc": 0.9147659063625451, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 21219, + 776, + 913, + 1327, + 431, + 508, + 21257, + 1866, + 28521, + 5186, + 1042, + 1585, + 1608, + 272, + 1681, + 20310 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a815c73098bdb11538358ad49ff07648c632f7ca --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7082782683095418, + "test_acc": 0.7082833133253301, + "test_auc": 0.7042404537908137, + "val_auc": 0.716183574879227, + "k": 1, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6930015920333388, + "test_acc": 0.6938775510204082, + "test_auc": 0.7524614923445858, + "val_auc": 0.7351046698872785, + "k": 2, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447 + ] + }, + { + "test_f1": 0.7574988074436555, + "test_acc": 0.7575030012004802, + "test_auc": 0.8380776148312119, + "val_auc": 0.817230273752013, + "k": 5, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 1528, + 21219 + ] + }, + { + "test_f1": 0.7790696172612332, + "test_acc": 0.7791116446578632, + "test_auc": 0.8511719470577384, + "val_auc": 0.788647342995169, + "k": 10, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 1528, + 21219, + 347, + 5186, + 32700, + 1417, + 1866 + ] + }, + { + "test_f1": 0.7778439320510695, + "test_acc": 0.7779111644657863, + "test_auc": 0.8579943737317839, + "val_auc": 0.8144122383252816, + "k": 16, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1447, + 6501, + 1528, + 21219, + 347, + 5186, + 32700, + 1417, + 1866, + 776, + 559, + 30737, + 8557, + 957, + 1377 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..84a5dc6b26ee1620179e57baf416ffcad627ca16 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7161327965715183, + "test_acc": 0.7162534435261708, + "test_auc": 0.7457652844393176, + "val_auc": 0.7150860344137655, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7189186679482901, + "test_acc": 0.71900826446281, + "test_auc": 0.7552516544229251, + "val_auc": 0.771108443377351, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186 + ] + }, + { + "test_f1": 0.7355131042165318, + "test_acc": 0.7355371900826446, + "test_auc": 0.7924078683747193, + "val_auc": 0.7867146858743497, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 3507, + 6501, + 1377 + ] + }, + { + "test_f1": 0.7382920110192838, + "test_acc": 0.7382920110192838, + "test_auc": 0.7925141157185356, + "val_auc": 0.7827130852340936, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 3507, + 6501, + 1377, + 1681, + 1447, + 508, + 1528, + 601 + ] + }, + { + "test_f1": 0.7602541654974159, + "test_acc": 0.7603305785123967, + "test_auc": 0.8310970797158642, + "val_auc": 0.8007202881152461, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 3507, + 6501, + 1377, + 1681, + 1447, + 508, + 1528, + 601, + 1417, + 3301, + 12612, + 1585, + 3078, + 1688 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e929b77457da8426e7bff974734d970dc2f1678e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6145588750359393, + "test_acc": 0.6266666666666667, + "test_auc": 0.6321661726777372, + "val_auc": 0.656062424969988, + "k": 1, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6358848318806377, + "test_acc": 0.6359748427672955, + "test_auc": 0.6828156913296589, + "val_auc": 0.7062825130052022, + "k": 2, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585 + ] + }, + { + "test_f1": 0.6695142335819075, + "test_acc": 0.669685534591195, + "test_auc": 0.7231638446684131, + "val_auc": 0.7687074829931972, + "k": 5, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 27673, + 776, + 4605 + ] + }, + { + "test_f1": 0.6684456960304164, + "test_acc": 0.6686792452830189, + "test_auc": 0.7297401672237754, + "val_auc": 0.7875150060024009, + "k": 10, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 27673, + 776, + 4605, + 1866, + 32700, + 1499, + 12749, + 11717 + ] + }, + { + "test_f1": 0.6629813549073433, + "test_acc": 0.6631446540880503, + "test_auc": 0.7285857824349216, + "val_auc": 0.7923169267707083, + "k": 16, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 27673, + 776, + 4605, + 1866, + 32700, + 1499, + 12749, + 11717, + 126, + 1848, + 1411, + 20310, + 4674, + 3832 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d0eaf4616dc76ebcd3fd142d94e49c15ab1e6932 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8013211383216383, + "test_acc": 0.8020125786163522, + "test_auc": 0.8426584924747275, + "val_auc": 0.820328131252501, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293 + ] + }, + { + "test_f1": 0.8085200606320208, + "test_acc": 0.8085534591194968, + "test_auc": 0.8905474872384787, + "val_auc": 0.8879551820728291, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062 + ] + }, + { + "test_f1": 0.8196191481916991, + "test_acc": 0.819622641509434, + "test_auc": 0.908934988896641, + "val_auc": 0.9323729491796718, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1166 + ] + }, + { + "test_f1": 0.8356283958300388, + "test_acc": 0.8357232704402515, + "test_auc": 0.9204788367851802, + "val_auc": 0.948779511804722, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1166, + 1441, + 409, + 1305, + 1313, + 1116 + ] + }, + { + "test_f1": 0.8445011791503332, + "test_acc": 0.8445283018867924, + "test_auc": 0.9234569470167762, + "val_auc": 0.9519807923169268, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1166, + 1441, + 409, + 1305, + 1313, + 1116, + 1928, + 14, + 765, + 524, + 4110, + 14876 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..12e0990d7342cccb04405a1be3fb944e09da9c5a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6424558959349187, + "test_acc": 0.6464646464646465, + "test_auc": 0.6683673469387755, + "val_auc": 0.6114010437575271, + "k": 1, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377 + ] + }, + { + "test_f1": 0.6500763448437867, + "test_acc": 0.6565656565656566, + "test_auc": 0.6983673469387754, + "val_auc": 0.6471296668004818, + "k": 2, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027 + ] + }, + { + "test_f1": 0.6767676767676768, + "test_acc": 0.6767676767676768, + "test_auc": 0.7306122448979593, + "val_auc": 0.6832597350461662, + "k": 5, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339 + ] + }, + { + "test_f1": 0.6662578777039687, + "test_acc": 0.6666666666666666, + "test_auc": 0.7306122448979592, + "val_auc": 0.6936973103171417, + "k": 10, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339, + 681, + 56, + 474, + 1551, + 524 + ] + }, + { + "test_f1": 0.6059801856474307, + "test_acc": 0.6060606060606061, + "test_auc": 0.7277551020408162, + "val_auc": 0.739462063428342, + "k": 16, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339, + 681, + 56, + 474, + 1551, + 524, + 100, + 2043, + 1790, + 505, + 272, + 957 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9c42e842057eb53f4ef48fdf3b820db6f1c0644c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6652130428440404, + "test_acc": 0.6680567139282736, + "test_auc": 0.6910823594880355, + "val_auc": 0.7966988727858293, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6733198081621736, + "test_acc": 0.6780650542118432, + "test_auc": 0.7184752365052866, + "val_auc": 0.8540660225442834, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114 + ] + }, + { + "test_f1": 0.7267440808362006, + "test_acc": 0.7289407839866555, + "test_auc": 0.8202615470228158, + "val_auc": 0.8784219001610306, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 1062, + 486, + 6086 + ] + }, + { + "test_f1": 0.7787174250660878, + "test_acc": 0.7798165137614679, + "test_auc": 0.8670534223706178, + "val_auc": 0.9146537842190017, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 1062, + 486, + 6086, + 30704, + 363, + 4481, + 6719, + 238 + ] + }, + { + "test_f1": 0.7814520236666413, + "test_acc": 0.7823185988323603, + "test_auc": 0.8724554813578186, + "val_auc": 0.927133655394525, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 1062, + 486, + 6086, + 30704, + 363, + 4481, + 6719, + 238, + 3276, + 1335, + 2744, + 1670, + 801, + 11788 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..53f46e6c6986a56ad4a57857f6af310eb30dfe3f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.5551201355328141, + "test_acc": 0.5849056603773585, + "test_auc": 0.6373247031256488, + "val_auc": 0.5426170468187275, + "k": 2, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727 + ] + }, + { + "test_f1": 0.6076910322339856, + "test_acc": 0.6138364779874214, + "test_auc": 0.6688406736341551, + "val_auc": 0.5812324929971989, + "k": 5, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1739 + ] + }, + { + "test_f1": 0.6245226927913186, + "test_acc": 0.6269182389937107, + "test_auc": 0.6936629844492218, + "val_auc": 0.6042416966786714, + "k": 10, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1739, + 1928, + 30704, + 293, + 4913, + 1438 + ] + }, + { + "test_f1": 0.640045455747909, + "test_acc": 0.6415094339622641, + "test_auc": 0.7037808127071438, + "val_auc": 0.6034413765506202, + "k": 16, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1739, + 1928, + 30704, + 293, + 4913, + 1438, + 765, + 3589, + 1260, + 6400, + 486, + 7143 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1d15d12e7fbae0c48f56a123c016d18696aeea0e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.495205220603222, + "test_acc": 0.538001196888091, + "test_auc": 0.5419283442684011, + "val_auc": 0.5666399036531513, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498 + ] + }, + { + "test_f1": 0.5450847718442882, + "test_acc": 0.5451825254338719, + "test_auc": 0.5777590751511331, + "val_auc": 0.5258932155760738, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727 + ] + }, + { + "test_f1": 0.5466260909132845, + "test_acc": 0.547576301615799, + "test_auc": 0.5739549608916139, + "val_auc": 0.5660377358490567, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 293, + 1409, + 5392 + ] + }, + { + "test_f1": 0.5551773299249134, + "test_acc": 0.5553560742070617, + "test_auc": 0.5779839841847405, + "val_auc": 0.5764753111200321, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 293, + 1409, + 5392, + 860, + 26677, + 1336, + 4913, + 19092 + ] + }, + { + "test_f1": 0.5595338247426213, + "test_acc": 0.5595451825254338, + "test_auc": 0.5924791565194969, + "val_auc": 0.5949417904456042, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 293, + 1409, + 5392, + 860, + 26677, + 1336, + 4913, + 19092, + 24099, + 844, + 2797, + 7163, + 1641, + 27404 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7161f82d6424951c24b4679cc3f8d46346d50cd3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5611618821807237, + "test_acc": 0.5954716981132075, + "test_auc": 0.6022346965537562, + "val_auc": 0.6032412965186074, + "k": 1, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659 + ] + }, + { + "test_f1": 0.5887451444276572, + "test_acc": 0.6035220125786164, + "test_auc": 0.6126036541341658, + "val_auc": 0.6698679471788714, + "k": 2, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794 + ] + }, + { + "test_f1": 0.6004697238546245, + "test_acc": 0.6155974842767296, + "test_auc": 0.6267132234777564, + "val_auc": 0.6856742697078833, + "k": 5, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 6400 + ] + }, + { + "test_f1": 0.6074820042041074, + "test_acc": 0.6203773584905661, + "test_auc": 0.6895013766544915, + "val_auc": 0.7735094037615047, + "k": 10, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 6400, + 7143, + 1377, + 1433, + 777, + 1596 + ] + }, + { + "test_f1": 0.6296664949615219, + "test_acc": 0.6344654088050314, + "test_auc": 0.7105869996020411, + "val_auc": 0.8041216486594638, + "k": 16, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6659, + 794, + 1345, + 13203, + 6400, + 7143, + 1377, + 1433, + 777, + 1596, + 1327, + 1550, + 1224, + 598, + 9390, + 6718 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b476a31a1bdd783a8c9413a8f579fa172ba0eb48 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.994968466843654, + "test_acc": 0.9949685534591195, + "test_auc": 0.9994524266889712, + "val_auc": 1.0, + "k": 1, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9979874122127191, + "test_acc": 0.9979874213836478, + "test_auc": 0.9999746845441041, + "val_auc": 1.0, + "k": 2, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.9999832917991087, + "val_auc": 1.0, + "k": 5, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1551 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.9999853170355804, + "val_auc": 1.0, + "k": 10, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1551, + 1525, + 347, + 6718, + 1232, + 1255 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.999993671136026, + "val_auc": 1.0, + "k": 16, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1625, + 1500, + 1703, + 1551, + 1525, + 347, + 6718, + 1232, + 1255, + 1042, + 1335, + 15467, + 549, + 1758, + 567 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1e4658575aeb6c6e7a5dcdf300e748ae51eeb704 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6457679593721823, + "test_acc": 0.6457861635220126, + "test_auc": 0.6881538349371519, + "val_auc": 0.7511004401760704, + "k": 1, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6516664932800128, + "test_acc": 0.6518238993710692, + "test_auc": 0.7014011598529274, + "val_auc": 0.7354941976790715, + "k": 2, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928 + ] + }, + { + "test_f1": 0.681684814825097, + "test_acc": 0.6817610062893081, + "test_auc": 0.7506409873432847, + "val_auc": 0.8323329331732693, + "k": 5, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 56 + ] + }, + { + "test_f1": 0.6910193170659633, + "test_acc": 0.6910691823899371, + "test_auc": 0.7626192484549978, + "val_auc": 0.8423369347739096, + "k": 10, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 56, + 5827, + 6400, + 152, + 4460, + 30704 + ] + }, + { + "test_f1": 0.7111916419079672, + "test_acc": 0.7111949685534591, + "test_auc": 0.7938757355405711, + "val_auc": 0.8467386954781913, + "k": 16, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 56, + 5827, + 6400, + 152, + 4460, + 30704, + 406, + 1255, + 1327, + 1691, + 121, + 1135 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fbe3729bded745d77321f4aa0e525d1b0676f40f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6719399030734693, + "test_acc": 0.6734591194968553, + "test_auc": 0.7190671710180561, + "val_auc": 0.7348939575830332, + "k": 1, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6678787557665665, + "test_acc": 0.670188679245283, + "test_auc": 0.7190671710180561, + "val_auc": 0.7348939575830332, + "k": 2, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143 + ] + }, + { + "test_f1": 0.6926832510698193, + "test_acc": 0.6938364779874214, + "test_auc": 0.7660663275070656, + "val_auc": 0.7521008403361343, + "k": 5, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 6400, + 1928 + ] + }, + { + "test_f1": 0.7085185128383352, + "test_acc": 0.7089308176100629, + "test_auc": 0.782158476779145, + "val_auc": 0.7663065226090436, + "k": 10, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 6400, + 1928, + 1011, + 30704, + 19940, + 598, + 253 + ] + }, + { + "test_f1": 0.7241787877414438, + "test_acc": 0.7245283018867924, + "test_auc": 0.7960878000767565, + "val_auc": 0.7911164465786314, + "k": 16, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 6400, + 1928, + 1011, + 30704, + 19940, + 598, + 253, + 9390, + 1960, + 7068, + 5827, + 16883, + 836 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3a88343c4bf830f83d68ac932cafc271c60ad36c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7852149332964958, + "test_acc": 0.7852825229960578, + "test_auc": 0.8165288301773778, + "val_auc": 0.7867999999999999, + "k": 1, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062 + ] + }, + { + "test_f1": 0.8041156470329448, + "test_acc": 0.8063074901445466, + "test_auc": 0.895068277273197, + "val_auc": 0.8195999999999999, + "k": 2, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448 + ] + }, + { + "test_f1": 0.8373026627028878, + "test_acc": 0.8373193166885676, + "test_auc": 0.9105930477805535, + "val_auc": 0.8436, + "k": 5, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 140 + ] + }, + { + "test_f1": 0.8486194837069319, + "test_acc": 0.8486202365308804, + "test_auc": 0.9226383379389342, + "val_auc": 0.8872, + "k": 10, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 140, + 1662, + 628, + 1030, + 296, + 2831 + ] + }, + { + "test_f1": 0.8499295081564934, + "test_acc": 0.849934296977661, + "test_auc": 0.9267916671501581, + "val_auc": 0.894, + "k": 16, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 140, + 1662, + 628, + 1030, + 296, + 2831, + 1313, + 2046, + 510, + 7451, + 1942, + 4334 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c9ece15cc78c07077b54bdacfd8d5d7288f72f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5210167881888842, + "test_acc": 0.5267924528301887, + "test_auc": 0.5349729225883738, + "val_auc": 0.6096438575430172, + "k": 1, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011 + ] + }, + { + "test_f1": 0.5221309134991221, + "test_acc": 0.5391194968553459, + "test_auc": 0.5811864139036534, + "val_auc": 0.5786314525810323, + "k": 2, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1377 + ] + }, + { + "test_f1": 0.5515485368740852, + "test_acc": 0.5524528301886793, + "test_auc": 0.5908608419515584, + "val_auc": 0.6678671468587435, + "k": 5, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1377, + 1411, + 800, + 1928 + ] + }, + { + "test_f1": 0.5531915140246981, + "test_acc": 0.5537106918238994, + "test_auc": 0.5971119874759375, + "val_auc": 0.6560624249699879, + "k": 10, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1377, + 1411, + 800, + 1928, + 1703, + 15467, + 786, + 4880, + 6718 + ] + }, + { + "test_f1": 0.5547154029713137, + "test_acc": 0.5547169811320755, + "test_auc": 0.595674196158329, + "val_auc": 0.6706682673069227, + "k": 16, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1377, + 1411, + 800, + 1928, + 1703, + 15467, + 786, + 4880, + 6718, + 1596, + 334, + 448, + 1202, + 1417, + 217 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..273b741942394ca60113f359db23e16b698b27b0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9916978231678448, + "test_acc": 0.9916981132075472, + "test_auc": 0.9971032789591094, + "val_auc": 1.0, + "k": 1, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.9927043184801099, + "test_acc": 0.9927044025157232, + "test_auc": 0.9986158774488907, + "val_auc": 1.0, + "k": 2, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988 + ] + }, + { + "test_f1": 0.9911947980288836, + "test_acc": 0.9911949685534591, + "test_auc": 0.9991428186633642, + "val_auc": 1.0, + "k": 5, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 292, + 152, + 487 + ] + }, + { + "test_f1": 0.9909431554840551, + "test_acc": 0.9909433962264151, + "test_auc": 0.9993066096630109, + "val_auc": 1.0, + "k": 10, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 292, + 152, + 487, + 6475, + 6935, + 559, + 1744, + 334 + ] + }, + { + "test_f1": 0.9924526295700459, + "test_acc": 0.9924528301886792, + "test_auc": 0.9993883785855545, + "val_auc": 1.0, + "k": 16, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 292, + 152, + 487, + 6475, + 6935, + 559, + 1744, + 334, + 347, + 821, + 7368, + 771, + 3417, + 1960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f0806f2b012f0c64160cfbf3097b318486f16da1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8523800528089518, + "test_acc": 0.8528301886792453, + "test_auc": 0.8840253397587337, + "val_auc": 0.9069627851140456, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487 + ] + }, + { + "test_f1": 0.8950056953882362, + "test_acc": 0.8950943396226415, + "test_auc": 0.9371381788466077, + "val_auc": 0.9627851140456183, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744 + ] + }, + { + "test_f1": 0.9143943801055264, + "test_acc": 0.9144654088050315, + "test_auc": 0.9692591887510266, + "val_auc": 0.9823929571828731, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 6475, + 24947 + ] + }, + { + "test_f1": 0.9400725282768482, + "test_acc": 0.940125786163522, + "test_auc": 0.9817789474643532, + "val_auc": 0.9811924769907964, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 6475, + 24947, + 7368, + 771, + 1028, + 559, + 1524 + ] + }, + { + "test_f1": 0.9501593057620207, + "test_acc": 0.950188679245283, + "test_auc": 0.9886422207122958, + "val_auc": 0.985594237695078, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 6475, + 24947, + 7368, + 771, + 1028, + 559, + 1524, + 152, + 878, + 3417, + 52, + 6935, + 1288 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e222a28e34c5c4f9f4ac3de292c908fae371cf57 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8411895448593, + "test_acc": 0.8435443037974684, + "test_auc": 0.8489210957016107, + "val_auc": 0.8175270108043217, + "k": 1, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104 + ] + }, + { + "test_f1": 0.8410277245721981, + "test_acc": 0.8435443037974684, + "test_auc": 0.9012614391953697, + "val_auc": 0.8887555022008804, + "k": 2, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555 + ] + }, + { + "test_f1": 0.8498412819208038, + "test_acc": 0.8516455696202532, + "test_auc": 0.9259779973665753, + "val_auc": 0.8997599039615847, + "k": 5, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 292, + 1708, + 337 + ] + }, + { + "test_f1": 0.8511322944830365, + "test_acc": 0.8511392405063292, + "test_auc": 0.9350329588291515, + "val_auc": 0.9191676670668267, + "k": 10, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 292, + 1708, + 337, + 1815, + 918, + 1851, + 1380, + 3385 + ] + }, + { + "test_f1": 0.8663132798122839, + "test_acc": 0.8663291139240507, + "test_auc": 0.9385431664267051, + "val_auc": 0.9375750300120049, + "k": 16, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 292, + 1708, + 337, + 1815, + 918, + 1851, + 1380, + 3385, + 677, + 968, + 3825, + 3203, + 777, + 4649 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9a1c0a823b0bda6f97d89df49d42f60351a1030d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7601483504188588, + "test_acc": 0.7651282051282051, + "test_auc": 0.780289578213889, + "val_auc": 0.7731538147694818, + "k": 1, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468 + ] + }, + { + "test_f1": 0.7805638789230808, + "test_acc": 0.7835897435897435, + "test_auc": 0.8444810987309387, + "val_auc": 0.8074255405956752, + "k": 2, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804 + ] + }, + { + "test_f1": 0.8555441595441595, + "test_acc": 0.8564102564102564, + "test_auc": 0.8997984481771973, + "val_auc": 0.898000815993472, + "k": 5, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804, + 1351, + 1132, + 1474 + ] + }, + { + "test_f1": 0.8599187163105629, + "test_acc": 0.8605128205128205, + "test_auc": 0.9035160063284747, + "val_auc": 0.8990208078335374, + "k": 10, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804, + 1351, + 1132, + 1474, + 6475, + 343, + 3417, + 805, + 21810 + ] + }, + { + "test_f1": 0.8684314698584149, + "test_acc": 0.8687179487179487, + "test_auc": 0.9262000538593597, + "val_auc": 0.901264789881681, + "k": 16, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804, + 1351, + 1132, + 1474, + 6475, + 343, + 3417, + 805, + 21810, + 930, + 1305, + 7595, + 24947, + 52, + 1473 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7a948a47094990babc3c9acc3112b7d63a3f2e8d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8826608690801097, + "test_acc": 0.8832704402515723, + "test_auc": 0.8959824624647735, + "val_auc": 0.9407763105242096, + "k": 1, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.8983285135903255, + "test_acc": 0.8986163522012579, + "test_auc": 0.9400778095852418, + "val_auc": 0.992797118847539, + "k": 2, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1012 + ] + }, + { + "test_f1": 0.905593877001827, + "test_acc": 0.9056603773584906, + "test_auc": 0.9692108362302654, + "val_auc": 0.9947979191676671, + "k": 5, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1012, + 1279, + 11381, + 1447 + ] + }, + { + "test_f1": 0.923763597337894, + "test_acc": 0.9237735849056604, + "test_auc": 0.980279766166197, + "val_auc": 0.9927971188475391, + "k": 10, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1012, + 1279, + 11381, + 1447, + 1556, + 238, + 4877, + 6828, + 7119 + ] + }, + { + "test_f1": 0.9393624113719857, + "test_acc": 0.9393710691823899, + "test_auc": 0.9821893110044261, + "val_auc": 0.9959983993597439, + "k": 16, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1012, + 1279, + 11381, + 1447, + 1556, + 238, + 4877, + 6828, + 7119, + 1478, + 1101, + 2860, + 7740, + 3301, + 6475 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5f01fc26300b6d8b215e3443b0915cbf142e374d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8402148384160218, + "test_acc": 0.8430769230769231, + "test_auc": 0.8427811626889287, + "val_auc": 0.8926968584251326, + "k": 1, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628 + ] + }, + { + "test_f1": 0.894880853037228, + "test_acc": 0.8953846153846153, + "test_auc": 0.9085358669673813, + "val_auc": 0.9135046919624644, + "k": 2, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416 + ] + }, + { + "test_f1": 0.8649000750289885, + "test_acc": 0.8666666666666667, + "test_auc": 0.9074208099101222, + "val_auc": 0.9192166462668299, + "k": 5, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 238 + ] + }, + { + "test_f1": 0.8649000750289885, + "test_acc": 0.8666666666666667, + "test_auc": 0.9074166021476421, + "val_auc": 0.9192166462668299, + "k": 10, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 238, + 4211, + 930, + 805, + 737, + 1132 + ] + }, + { + "test_f1": 0.8989739165624071, + "test_acc": 0.8994871794871795, + "test_auc": 0.9262505470091225, + "val_auc": 0.927376580987352, + "k": 16, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 238, + 4211, + 930, + 805, + 737, + 1132, + 2062, + 334, + 52, + 1524, + 1288, + 878 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2af2afe2524f3cc7149c1fa177600757a1dfc9c5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8706941484779513, + "test_acc": 0.8707692307692307, + "test_auc": 0.9125816305921163, + "val_auc": 0.9408404732762137, + "k": 1, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524 + ] + }, + { + "test_f1": 0.8695803352000449, + "test_acc": 0.8697435897435898, + "test_auc": 0.9261053792035548, + "val_auc": 0.9383924928600571, + "k": 2, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947 + ] + }, + { + "test_f1": 0.8664897128010673, + "test_acc": 0.8666666666666667, + "test_auc": 0.9325874373043391, + "val_auc": 0.9412484700122399, + "k": 5, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 969 + ] + }, + { + "test_f1": 0.8675066672597983, + "test_acc": 0.8676923076923077, + "test_auc": 0.9322634395933619, + "val_auc": 0.9392084863321093, + "k": 10, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 969, + 888, + 1474, + 3203, + 1682, + 7309 + ] + }, + { + "test_f1": 0.8707175528225678, + "test_acc": 0.8707692307692307, + "test_auc": 0.938987444036759, + "val_auc": 0.9396164830681355, + "k": 16, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 969, + 888, + 1474, + 3203, + 1682, + 7309, + 6631, + 805, + 1214, + 1012, + 483, + 1260 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1da9f118c01bdc0d900d4eec357d1579a1a8963c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7949197854477767, + "test_acc": 0.7979487179487179, + "test_auc": 0.8186265021712055, + "val_auc": 0.7982456140350876, + "k": 1, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.8008141828356498, + "test_acc": 0.8051282051282052, + "test_auc": 0.8542178611101762, + "val_auc": 0.8131374949000408, + "k": 2, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815 + ] + }, + { + "test_f1": 0.8443143186173138, + "test_acc": 0.8451282051282051, + "test_auc": 0.9173932069882519, + "val_auc": 0.9075887392900857, + "k": 5, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 334, + 1866, + 1524 + ] + }, + { + "test_f1": 0.8798316284449333, + "test_acc": 0.88, + "test_auc": 0.946443599151715, + "val_auc": 0.9106487148102815, + "k": 10, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 334, + 1866, + 1524, + 487, + 930, + 24947, + 6935, + 913 + ] + }, + { + "test_f1": 0.8840937854809031, + "test_acc": 0.884102564102564, + "test_auc": 0.9520588581815733, + "val_auc": 0.9257445940432476, + "k": 16, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 334, + 1866, + 1524, + 487, + 930, + 24947, + 6935, + 913, + 2046, + 10639, + 1221, + 6758, + 52, + 508 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5adfd7aa16836578cb2223290788dc4abd90ced7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8403697830752239, + "test_acc": 0.841025641025641, + "test_auc": 0.880743595785505, + "val_auc": 0.8875968992248062, + "k": 1, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52 + ] + }, + { + "test_f1": 0.8393609489451941, + "test_acc": 0.84, + "test_auc": 0.8805605581176152, + "val_auc": 0.8894328845369238, + "k": 2, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230 + ] + }, + { + "test_f1": 0.8527665056360708, + "test_acc": 0.8533333333333334, + "test_auc": 0.8948606389066549, + "val_auc": 0.8924928600571196, + "k": 5, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 18545, + 2884, + 1274 + ] + }, + { + "test_f1": 0.8496186219216338, + "test_acc": 0.8502564102564103, + "test_auc": 0.8948543272629346, + "val_auc": 0.9147286821705426, + "k": 10, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 18545, + 2884, + 1274, + 2131, + 2046, + 888, + 2484, + 2454 + ] + }, + { + "test_f1": 0.8487228420837359, + "test_acc": 0.8492307692307692, + "test_auc": 0.9003980543306291, + "val_auc": 0.9375764993880049, + "k": 16, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 18545, + 2884, + 1274, + 2131, + 2046, + 888, + 2484, + 2454, + 238, + 821, + 1682, + 17546, + 1193, + 1815 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6a591053a9f902adc97ff2ae9fc90a60742386aa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8768743711223518, + "test_acc": 0.8770053475935828, + "test_auc": 0.9209150057651129, + "val_auc": 0.9230303030303029, + "k": 1, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839 + ] + }, + { + "test_f1": 0.8800778305655748, + "test_acc": 0.8802139037433155, + "test_auc": 0.9357738977653325, + "val_auc": 0.9329292929292929, + "k": 2, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1274 + ] + }, + { + "test_f1": 0.8876655844870426, + "test_acc": 0.8877005347593583, + "test_auc": 0.9462769267373121, + "val_auc": 0.9426262626262627, + "k": 5, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1274, + 1682, + 1305, + 343 + ] + }, + { + "test_f1": 0.8983829360299947, + "test_acc": 0.8983957219251337, + "test_auc": 0.9544052782810813, + "val_auc": 0.9498989898989899, + "k": 10, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1274, + 1682, + 1305, + 343, + 930, + 3417, + 340, + 874, + 888 + ] + }, + { + "test_f1": 0.8887112423334863, + "test_acc": 0.8887700534759359, + "test_auc": 0.9523829133036843, + "val_auc": 0.9430303030303031, + "k": 16, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1274, + 1682, + 1305, + 343, + 930, + 3417, + 340, + 874, + 888, + 1895, + 238, + 1441, + 456, + 334, + 1776 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..98e1c1181bc629e7841b527581f1a641682fb583 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9484981044709709, + "test_acc": 0.9485057471264368, + "test_auc": 0.9839327750419395, + "val_auc": 0.9955357142857143, + "k": 1, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554 + ] + }, + { + "test_f1": 0.954477446727041, + "test_acc": 0.9544827586206897, + "test_auc": 0.9896292751231126, + "val_auc": 0.9983766233766234, + "k": 2, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072 + ] + }, + { + "test_f1": 0.9682758620689655, + "test_acc": 0.9682758620689655, + "test_auc": 0.9947245860165594, + "val_auc": 1.0, + "k": 5, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 1065, + 1101, + 760 + ] + }, + { + "test_f1": 0.9705747002033374, + "test_acc": 0.9705747126436781, + "test_auc": 0.9943322487688727, + "val_auc": 1.0, + "k": 10, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 1065, + 1101, + 760, + 1682, + 1622, + 192, + 1876, + 763 + ] + }, + { + "test_f1": 0.9825286248348366, + "test_acc": 0.9825287356321839, + "test_auc": 0.9959954542994751, + "val_auc": 1.0, + "k": 16, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 1065, + 1101, + 760, + 1682, + 1622, + 192, + 1876, + 763, + 4323, + 1684, + 1872, + 652, + 14949, + 770 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..54d45f2ee8121b1eb8dde52f600520d753bad88e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.969420793501061, + "test_acc": 0.9694218824653608, + "test_auc": 0.98603402966867, + "val_auc": 1.0, + "k": 1, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286 + ] + }, + { + "test_f1": 0.961289655175752, + "test_acc": 0.9612995699952221, + "test_auc": 0.98603402966867, + "val_auc": 1.0, + "k": 2, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905 + ] + }, + { + "test_f1": 0.98040995434776, + "test_acc": 0.9804108934543717, + "test_auc": 0.9956453931016599, + "val_auc": 1.0, + "k": 5, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581 + ] + }, + { + "test_f1": 0.9847106271146736, + "test_acc": 0.9847109412326803, + "test_auc": 0.9967776456816435, + "val_auc": 1.0, + "k": 10, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581, + 2874, + 304, + 948, + 3877, + 1682 + ] + }, + { + "test_f1": 0.989488196195524, + "test_acc": 0.9894887720974678, + "test_auc": 0.9982687492809282, + "val_auc": 1.0, + "k": 16, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581, + 2874, + 304, + 948, + 3877, + 1682, + 1079, + 1684, + 1156, + 7003, + 239, + 1596 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..da1099c4f33b94493f9e82afd3581110b453c767 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9935631204214642, + "test_acc": 0.9935632183908046, + "test_auc": 0.9984915309269982, + "val_auc": 1.0, + "k": 1, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559 + ] + }, + { + "test_f1": 0.9935631204214642, + "test_acc": 0.9935632183908046, + "test_auc": 0.9998376535526814, + "val_auc": 1.0, + "k": 2, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2236 + ] + }, + { + "test_f1": 0.9940229859788012, + "test_acc": 0.9940229885057471, + "test_auc": 0.9998156691379403, + "val_auc": 1.0, + "k": 5, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2236, + 300, + 1789, + 334 + ] + }, + { + "test_f1": 0.9958620689655172, + "test_acc": 0.9958620689655172, + "test_auc": 0.9998478002056389, + "val_auc": 1.0, + "k": 10, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2236, + 300, + 1789, + 334, + 982, + 3825, + 1934, + 1872, + 1592 + ] + }, + { + "test_f1": 0.9954022969067715, + "test_acc": 0.9954022988505747, + "test_auc": 0.9994258685534931, + "val_auc": 1.0, + "k": 16, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 2236, + 300, + 1789, + 334, + 982, + 3825, + 1934, + 1872, + 1592, + 724, + 1158, + 913, + 1268, + 1769, + 896 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fb19a2419730f19aad061036f41bf6deed44441a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9485047674330317, + "test_acc": 0.9485057471264368, + "test_auc": 0.9791537014989988, + "val_auc": 0.9655032467532468, + "k": 1, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095 + ] + }, + { + "test_f1": 0.9622977985617085, + "test_acc": 0.9622988505747126, + "test_auc": 0.9883837734725904, + "val_auc": 0.974025974025974, + "k": 2, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023 + ] + }, + { + "test_f1": 0.9719538451291926, + "test_acc": 0.9719540229885058, + "test_auc": 0.9925625033822176, + "val_auc": 0.976461038961039, + "k": 5, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1682, + 494 + ] + }, + { + "test_f1": 0.9774705214237565, + "test_acc": 0.9774712643678161, + "test_auc": 0.9957012013637101, + "val_auc": 0.9837662337662337, + "k": 10, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1682, + 494, + 1789, + 1301, + 211, + 1079, + 1701 + ] + }, + { + "test_f1": 0.9825278492159298, + "test_acc": 0.9825287356321839, + "test_auc": 0.9961696385085773, + "val_auc": 0.9837662337662336, + "k": 16, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1682, + 494, + 1789, + 1301, + 211, + 1079, + 1701, + 1193, + 235, + 2011, + 6551, + 888, + 292 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ebf75ca4b7a56a4c1f51e1c6b8649f3d34c8aef3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9576933689431343, + "test_acc": 0.9577011494252874, + "test_auc": 0.9894517086963581, + "val_auc": 0.981737012987013, + "k": 1, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32 + ] + }, + { + "test_f1": 0.9848270858568157, + "test_acc": 0.9848275862068966, + "test_auc": 0.9973770902105092, + "val_auc": 0.9955357142857143, + "k": 2, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683 + ] + }, + { + "test_f1": 0.994942483833445, + "test_acc": 0.9949425287356322, + "test_auc": 0.9997928391687861, + "val_auc": 0.9979707792207793, + "k": 5, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 816, + 1558 + ] + }, + { + "test_f1": 0.9963218297501885, + "test_acc": 0.9963218390804598, + "test_auc": 0.9997623992099139, + "val_auc": 0.997564935064935, + "k": 10, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 816, + 1558, + 1047, + 1458, + 1627, + 87, + 495 + ] + }, + { + "test_f1": 0.9967816078347392, + "test_acc": 0.9967816091954023, + "test_auc": 0.999906143460144, + "val_auc": 0.9971590909090909, + "k": 16, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 816, + 1558, + 1047, + 1458, + 1627, + 87, + 495, + 232, + 1023, + 5094, + 868, + 314, + 1682 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0f966974b58f332aa4b500b9a05c397732c588c1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8269379370926538, + "test_acc": 0.8271698113207547, + "test_auc": 0.8712258452577569, + "val_auc": 0.8903561424569828, + "k": 1, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.8958457211571182, + "test_acc": 0.8958490566037736, + "test_auc": 0.9488746773545145, + "val_auc": 0.981592637054822, + "k": 2, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028 + ] + }, + { + "test_f1": 0.9061612652180308, + "test_acc": 0.9061635220125787, + "test_auc": 0.9628166583800741, + "val_auc": 0.9925970388155262, + "k": 5, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 1305, + 1839, + 3417 + ] + }, + { + "test_f1": 0.9132006254524038, + "test_acc": 0.9132075471698113, + "test_auc": 0.9668334617670796, + "val_auc": 0.9939975990396158, + "k": 10, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 1305, + 1839, + 3417, + 778, + 1602, + 1895, + 5539, + 4663 + ] + }, + { + "test_f1": 0.9124375571914498, + "test_acc": 0.9124528301886793, + "test_auc": 0.9669309262722791, + "val_auc": 0.9943977591036415, + "k": 16, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 1305, + 1839, + 3417, + 778, + 1602, + 1895, + 5539, + 4663, + 1478, + 3203, + 343, + 30649, + 22848, + 988 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3dbf765572402f33337f85eab266cd4e5d8f99ee --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999813978029114, + "val_auc": 1.0, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379 + ] + }, + { + "test_f1": 0.9990804593813543, + "test_acc": 0.999080459770115, + "test_auc": 0.9999915444558688, + "val_auc": 1.0, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999957722279345, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 4787, + 23428 + ] + }, + { + "test_f1": 0.9990804593813543, + "test_acc": 0.999080459770115, + "test_auc": 0.9999932355646951, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 4787, + 23428, + 25696, + 547, + 1162, + 1840, + 24163 + ] + }, + { + "test_f1": 0.9990804593813543, + "test_acc": 0.999080459770115, + "test_auc": 0.999993235564695, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 4787, + 23428, + 25696, + 547, + 1162, + 1840, + 24163, + 1584, + 1931, + 613, + 6777, + 1273, + 3818 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..69fbe3042c0a507eb61d8839b160a88a8e9efb4e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9816082623325577, + "test_acc": 0.9816091954022989, + "test_auc": 0.9943381676497645, + "val_auc": 1.0, + "k": 1, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211 + ] + }, + { + "test_f1": 0.9843673798914402, + "test_acc": 0.984367816091954, + "test_auc": 0.9953773540234861, + "val_auc": 1.0, + "k": 2, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586 + ] + }, + { + "test_f1": 0.989885031813018, + "test_acc": 0.9898850574712643, + "test_auc": 0.9970143473672819, + "val_auc": 1.0, + "k": 5, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 10, + 21724, + 559 + ] + }, + { + "test_f1": 0.9912643308837131, + "test_acc": 0.991264367816092, + "test_auc": 0.9968207154066777, + "val_auc": 1.0, + "k": 10, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 10, + 21724, + 559, + 620, + 3223, + 878, + 495, + 1934 + ] + }, + { + "test_f1": 0.9917240854480806, + "test_acc": 0.9917241379310345, + "test_auc": 0.9960850830672655, + "val_auc": 1.0, + "k": 16, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 10, + 21724, + 559, + 620, + 3223, + 878, + 495, + 1934, + 610, + 1682, + 1874, + 1872, + 760, + 1023 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5f3643c8c13d143285ecbbf3e52aada8272652a2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8551711279577412, + "test_acc": 0.8551724137931035, + "test_auc": 0.9132545727582662, + "val_auc": 0.9650974025974026, + "k": 1, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495 + ] + }, + { + "test_f1": 0.8850491608730604, + "test_acc": 0.8850574712643678, + "test_auc": 0.9524367187077223, + "val_auc": 0.9715909090909092, + "k": 2, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495, + 1682 + ] + }, + { + "test_f1": 0.9351723863854602, + "test_acc": 0.9351724137931035, + "test_auc": 0.980060981384274, + "val_auc": 0.989448051948052, + "k": 5, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495, + 1682, + 6979, + 27839, + 1027 + ] + }, + { + "test_f1": 0.9567747014106545, + "test_acc": 0.9567816091954023, + "test_auc": 0.9907978313220412, + "val_auc": 0.9959415584415585, + "k": 10, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495, + 1682, + 6979, + 27839, + 1027, + 1708, + 1631, + 1707, + 1841, + 713 + ] + }, + { + "test_f1": 0.959994063360937, + "test_acc": 0.96, + "test_auc": 0.9933471778775909, + "val_auc": 0.9983766233766235, + "k": 16, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495, + 1682, + 6979, + 27839, + 1027, + 1708, + 1631, + 1707, + 1841, + 713, + 2024, + 2616, + 1840, + 532, + 1079, + 1583 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cf907d69bff0dc0889e0b56423fb54488b7ca88d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9024319628151045, + "test_acc": 0.9025287356321839, + "test_auc": 0.9479840291682449, + "val_auc": 0.9468344155844156, + "k": 1, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961 + ] + }, + { + "test_f1": 0.9153999382360288, + "test_acc": 0.9154022988505747, + "test_auc": 0.9668475025704854, + "val_auc": 0.9663149350649352, + "k": 2, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182 + ] + }, + { + "test_f1": 0.9443623824127115, + "test_acc": 0.944367816091954, + "test_auc": 0.9812176998214188, + "val_auc": 0.984577922077922, + "k": 5, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182, + 248, + 8734, + 732 + ] + }, + { + "test_f1": 0.9554010543901167, + "test_acc": 0.9554022988505747, + "test_auc": 0.9843745772227935, + "val_auc": 0.994724025974026, + "k": 10, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182, + 248, + 8734, + 732, + 1264, + 2689, + 859, + 1441, + 1536 + ] + }, + { + "test_f1": 0.9655172413793104, + "test_acc": 0.9655172413793104, + "test_auc": 0.9904908950700795, + "val_auc": 0.9979707792207791, + "k": 16, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182, + 248, + 8734, + 732, + 1264, + 2689, + 859, + 1441, + 1536, + 653, + 1682, + 1482, + 1960, + 306, + 1856 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..142fed34ca28b671ec101ff6f382b62db1038954 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.952620121793032, + "test_acc": 0.9526436781609195, + "test_auc": 0.9829722252286379, + "val_auc": 0.9926948051948052, + "k": 1, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786 + ] + }, + { + "test_f1": 0.9567492443883543, + "test_acc": 0.9567816091954023, + "test_auc": 0.9829722252286379, + "val_auc": 0.9926948051948052, + "k": 2, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558 + ] + }, + { + "test_f1": 0.9567492443883543, + "test_acc": 0.9567816091954023, + "test_auc": 0.9829722252286379, + "val_auc": 0.9926948051948052, + "k": 5, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 271 + ] + }, + { + "test_f1": 0.965968940133108, + "test_acc": 0.9659770114942529, + "test_auc": 0.9912070796579902, + "val_auc": 0.9971590909090909, + "k": 10, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 271, + 67, + 653, + 292, + 1910, + 327 + ] + }, + { + "test_f1": 0.9724100021870672, + "test_acc": 0.9724137931034482, + "test_auc": 0.9953037907895448, + "val_auc": 0.997564935064935, + "k": 16, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 271, + 67, + 653, + 292, + 1910, + 327, + 460, + 276, + 760, + 1051, + 1072, + 1649 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bb7850ce419ff988e96bdf9d699b576820ef4d0b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9627562115697981, + "test_acc": 0.9627586206896551, + "test_auc": 0.9856661615888306, + "val_auc": 0.9801136363636364, + "k": 1, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831 + ] + }, + { + "test_f1": 0.9627586206896551, + "test_acc": 0.9627586206896551, + "test_auc": 0.9884674833594891, + "val_auc": 0.9886363636363635, + "k": 2, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856 + ] + }, + { + "test_f1": 0.9728735288130513, + "test_acc": 0.9728735632183908, + "test_auc": 0.9913068550787381, + "val_auc": 0.9882305194805194, + "k": 5, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 7739, + 1168, + 1167 + ] + }, + { + "test_f1": 0.9770092489512836, + "test_acc": 0.9770114942528736, + "test_auc": 0.9930241760917797, + "val_auc": 0.9914772727272727, + "k": 10, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 7739, + 1168, + 1167, + 1037, + 517, + 822, + 18934, + 1449 + ] + }, + { + "test_f1": 0.9765487497574294, + "test_acc": 0.976551724137931, + "test_auc": 0.9940785824449376, + "val_auc": 0.9959415584415584, + "k": 16, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 7739, + 1168, + 1167, + 1037, + 517, + 822, + 18934, + 1449, + 1252, + 1536, + 581, + 1366, + 779, + 586 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4d6ae167985e4435581bd88bd0bdcf0a75a08452 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.961370687205417, + "test_acc": 0.9613793103448276, + "test_auc": 0.9727350979490232, + "val_auc": 0.9884334415584416, + "k": 1, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840 + ] + }, + { + "test_f1": 0.9558601095190609, + "test_acc": 0.9558620689655173, + "test_auc": 0.9929480761945993, + "val_auc": 0.9991883116883117, + "k": 2, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495 + ] + }, + { + "test_f1": 0.9747125367686433, + "test_acc": 0.9747126436781609, + "test_auc": 0.9959007522052059, + "val_auc": 0.9991883116883117, + "k": 5, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495, + 1765, + 789, + 1763 + ] + }, + { + "test_f1": 0.9806891163364849, + "test_acc": 0.9806896551724138, + "test_auc": 0.9965974890416148, + "val_auc": 0.9995941558441559, + "k": 10, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495, + 1765, + 789, + 1763, + 1682, + 20223, + 901, + 434, + 1797 + ] + }, + { + "test_f1": 0.9806891163364849, + "test_acc": 0.9806896551724138, + "test_auc": 0.9973627157854863, + "val_auc": 1.0, + "k": 16, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495, + 1765, + 789, + 1763, + 1682, + 20223, + 901, + 434, + 1797, + 27505, + 1310, + 1815, + 677, + 2120, + 1684 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7cec81f7eabbe4970d44e311406b79027ae82d9b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.959534669874365, + "test_acc": 0.9595402298850575, + "test_auc": 0.9755465663726393, + "val_auc": 0.9756493506493507, + "k": 1, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879 + ] + }, + { + "test_f1": 0.9747122587999921, + "test_acc": 0.9747126436781609, + "test_auc": 0.9909796255208616, + "val_auc": 0.9971590909090908, + "k": 2, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1051 + ] + }, + { + "test_f1": 0.9885057471264368, + "test_acc": 0.9885057471264368, + "test_auc": 0.9971132772336164, + "val_auc": 1.0, + "k": 5, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1051, + 1293, + 854, + 277 + ] + }, + { + "test_f1": 0.9871263224848184, + "test_acc": 0.9871264367816092, + "test_auc": 0.9972544848206072, + "val_auc": 1.0, + "k": 10, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1051, + 1293, + 854, + 277, + 4455, + 30, + 7367, + 1501, + 30210 + ] + }, + { + "test_f1": 0.9871264367816092, + "test_acc": 0.9871264367816092, + "test_auc": 0.9958331078521564, + "val_auc": 1.0, + "k": 16, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1051, + 1293, + 854, + 277, + 4455, + 30, + 7367, + 1501, + 30210, + 808, + 1454, + 1867, + 1484, + 419, + 613 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..934dda458919984967fe73479498b3c815f46f48 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9714941323571504, + "test_acc": 0.9714942528735632, + "test_auc": 0.9787791208939878, + "val_auc": 0.9772727272727273, + "k": 1, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439 + ] + }, + { + "test_f1": 0.9714941805639601, + "test_acc": 0.9714942528735632, + "test_auc": 0.9828496198387359, + "val_auc": 0.9991883116883117, + "k": 2, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161 + ] + }, + { + "test_f1": 0.9701149425287356, + "test_acc": 0.9701149425287356, + "test_auc": 0.9824471359380919, + "val_auc": 0.9991883116883117, + "k": 5, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161, + 26232, + 1795, + 5535 + ] + }, + { + "test_f1": 0.9696551724137931, + "test_acc": 0.9696551724137931, + "test_auc": 0.9864085583635478, + "val_auc": 1.0, + "k": 10, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161, + 26232, + 1795, + 5535, + 1121, + 822, + 1790, + 493, + 802 + ] + }, + { + "test_f1": 0.9797700892842823, + "test_acc": 0.9797701149425287, + "test_auc": 0.9985177431138049, + "val_auc": 1.0, + "k": 16, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161, + 26232, + 1795, + 5535, + 1121, + 822, + 1790, + 493, + 802, + 1441, + 345, + 2020, + 1631, + 495, + 5797 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fb67716f48d071cc617e8d1be7807c1d09333e48 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9544628325320103, + "test_acc": 0.9544827586206897, + "test_auc": 0.9784396307971209, + "val_auc": 0.9833603896103896, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272 + ] + }, + { + "test_f1": 0.9493658821382959, + "test_acc": 0.9494252873563218, + "test_auc": 0.9784396307971209, + "val_auc": 0.9833603896103896, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866 + ] + }, + { + "test_f1": 0.9632144561076585, + "test_acc": 0.9632183908045977, + "test_auc": 0.9940861924346555, + "val_auc": 0.9955357142857143, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 5703, + 1072 + ] + }, + { + "test_f1": 0.9852872256969354, + "test_acc": 0.9852873563218391, + "test_auc": 0.99792247280697, + "val_auc": 1.0, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 5703, + 1072, + 165, + 1682, + 918, + 6099, + 1787 + ] + }, + { + "test_f1": 0.9843677764383008, + "test_acc": 0.984367816091954, + "test_auc": 0.9977110842036907, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 5703, + 1072, + 165, + 1682, + 918, + 6099, + 1787, + 1726, + 508, + 724, + 1079, + 495, + 1715 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d6b9447f4d342f6d790e0767cc541d7656746d26 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5518506377867791, + "test_acc": 0.579622641509434, + "test_auc": 0.5874719884480513, + "val_auc": 0.6050420168067228, + "k": 1, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475 + ] + }, + { + "test_f1": 0.6060049924358923, + "test_acc": 0.6083018867924528, + "test_auc": 0.6130184478790205, + "val_auc": 0.6152460984393757, + "k": 2, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475, + 1556 + ] + }, + { + "test_f1": 0.608631508910011, + "test_acc": 0.6113207547169811, + "test_auc": 0.6239192831878032, + "val_auc": 0.6828731492597039, + "k": 5, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475, + 1556, + 1132, + 1279, + 1815 + ] + }, + { + "test_f1": 0.5937967449276538, + "test_acc": 0.5949685534591195, + "test_auc": 0.6625628203038058, + "val_auc": 0.6914765906362546, + "k": 10, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475, + 1556, + 1132, + 1279, + 1815, + 776, + 7893, + 238, + 1866, + 7740 + ] + }, + { + "test_f1": 0.6425652838918741, + "test_acc": 0.6427672955974842, + "test_auc": 0.6842872787808887, + "val_auc": 0.7074829931972789, + "k": 16, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475, + 1556, + 1132, + 1279, + 1815, + 776, + 7893, + 238, + 1866, + 7740, + 581, + 1895, + 1148, + 2126, + 6758, + 1028 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9c78be456bb7609cf4e38ad0d4240554fe6a61aa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9484841850106541, + "test_acc": 0.9485057471264368, + "test_auc": 0.9650811394014829, + "val_auc": 0.9740259740259741, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772 + ] + }, + { + "test_f1": 0.9779300547595304, + "test_acc": 0.9779310344827586, + "test_auc": 0.9870384964013204, + "val_auc": 1.0, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963 + ] + }, + { + "test_f1": 0.9595109408823672, + "test_acc": 0.9595402298850575, + "test_auc": 0.9874900224579253, + "val_auc": 1.0, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963, + 217, + 352, + 1942 + ] + }, + { + "test_f1": 0.9608940789721028, + "test_acc": 0.960919540229885, + "test_auc": 0.993550110936739, + "val_auc": 1.0, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963, + 217, + 352, + 1942, + 1682, + 1898, + 111, + 1261, + 581 + ] + }, + { + "test_f1": 0.9608940789721028, + "test_acc": 0.960919540229885, + "test_auc": 0.9935534931543915, + "val_auc": 1.0, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963, + 217, + 352, + 1942, + 1682, + 1898, + 111, + 1261, + 581, + 1726, + 393, + 760, + 1047, + 285, + 3243 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..019b72f9f426f1100c2c7e75624b1fe3b423e2b9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.982988304366428, + "test_acc": 0.9829885057471265, + "test_auc": 0.9967412332918448, + "val_auc": 0.9995941558441559, + "k": 1, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833 + ] + }, + { + "test_f1": 0.9751675319124512, + "test_acc": 0.9751724137931035, + "test_auc": 0.9967412332918448, + "val_auc": 0.9995941558441559, + "k": 2, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675 + ] + }, + { + "test_f1": 0.9880459770114942, + "test_acc": 0.9880459770114942, + "test_auc": 0.9984982953623032, + "val_auc": 0.9995941558441559, + "k": 5, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682 + ] + }, + { + "test_f1": 0.9903448153402388, + "test_acc": 0.9903448275862069, + "test_auc": 0.9982505479192597, + "val_auc": 1.0, + "k": 10, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682, + 829, + 913, + 1588, + 491, + 1281 + ] + }, + { + "test_f1": 0.991724064454526, + "test_acc": 0.9917241379310345, + "test_auc": 0.9993472319930732, + "val_auc": 1.0, + "k": 16, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 596, + 1682, + 829, + 913, + 1588, + 491, + 1281, + 361, + 1181, + 113, + 1515, + 1631, + 1079 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..67cda75e53340c32a7a8de610ef357d5f53cc30a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8587914725255325, + "test_acc": 0.8588505747126437, + "test_auc": 0.9037581511445425, + "val_auc": 0.9364853896103895, + "k": 1, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020 + ] + }, + { + "test_f1": 0.9038809819933558, + "test_acc": 0.9039080459770115, + "test_auc": 0.9365419868499377, + "val_auc": 0.9711850649350648, + "k": 2, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 1482 + ] + }, + { + "test_f1": 0.9402013678163255, + "test_acc": 0.9402298850574713, + "test_auc": 0.982210380702419, + "val_auc": 0.9837662337662337, + "k": 5, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 1482, + 1682, + 760, + 913 + ] + }, + { + "test_f1": 0.9517189736904322, + "test_acc": 0.9517241379310345, + "test_auc": 0.9887329874452081, + "val_auc": 0.9890422077922078, + "k": 10, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 1482, + 1682, + 760, + 913, + 49, + 18571, + 170, + 3842, + 1960 + ] + }, + { + "test_f1": 0.9572279798917432, + "test_acc": 0.9572413793103448, + "test_auc": 0.9930284038638455, + "val_auc": 0.9959415584415584, + "k": 16, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 1482, + 1682, + 760, + 913, + 49, + 18571, + 170, + 3842, + 1960, + 2023, + 364, + 1120, + 2817, + 1264, + 901 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4e830622b81208b6db39bd2fae67f29a9853e04a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9544804492579462, + "test_acc": 0.9544827586206897, + "test_auc": 0.9651022782618108, + "val_auc": 0.9541396103896104, + "k": 1, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123 + ] + }, + { + "test_f1": 0.9572350332478978, + "test_acc": 0.9572413793103448, + "test_auc": 0.9755355741652687, + "val_auc": 0.96875, + "k": 2, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491 + ] + }, + { + "test_f1": 0.969192667136985, + "test_acc": 0.9691954022988506, + "test_auc": 0.9834279790573084, + "val_auc": 0.9849837662337662, + "k": 5, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 866, + 1682, + 1763 + ] + }, + { + "test_f1": 0.967349033333319, + "test_acc": 0.9673563218390805, + "test_auc": 0.9899302924941826, + "val_auc": 0.9793019480519479, + "k": 10, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 866, + 1682, + 1763, + 1281, + 1111, + 167, + 679, + 523 + ] + }, + { + "test_f1": 0.9719492081684445, + "test_acc": 0.9719540229885058, + "test_auc": 0.9900207668163862, + "val_auc": 0.989448051948052, + "k": 16, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 866, + 1682, + 1763, + 1281, + 1111, + 167, + 679, + 523, + 687, + 7316, + 1486, + 204, + 144, + 1310 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..61d21930fdf7a28c3c553d16f1a1595a2cdaaaa5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8515778484674111, + "test_acc": 0.8519540229885058, + "test_auc": 0.8998965041398344, + "val_auc": 0.8944805194805195, + "k": 1, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967 + ] + }, + { + "test_f1": 0.8732593557558297, + "test_acc": 0.8735632183908046, + "test_auc": 0.9251105139617946, + "val_auc": 0.9273538961038962, + "k": 2, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352 + ] + }, + { + "test_f1": 0.9309862530430237, + "test_acc": 0.9310344827586207, + "test_auc": 0.9747475174522431, + "val_auc": 0.976866883116883, + "k": 5, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 495, + 1682, + 178 + ] + }, + { + "test_f1": 0.9457273707686503, + "test_acc": 0.9457471264367816, + "test_auc": 0.9809437401915687, + "val_auc": 0.9837662337662337, + "k": 10, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 495, + 1682, + 178, + 1762, + 12346, + 19873, + 859, + 108 + ] + }, + { + "test_f1": 0.951248291285272, + "test_acc": 0.9512643678160919, + "test_auc": 0.983037332918448, + "val_auc": 0.9963474025974026, + "k": 16, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 495, + 1682, + 178, + 1762, + 12346, + 19873, + 859, + 108, + 182, + 960, + 1168, + 507, + 1037, + 1797 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..667214c92183c679dae50ed551d47816c6c1229c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8954511848844743, + "test_acc": 0.895632183908046, + "test_auc": 0.9488841218410088, + "val_auc": 0.9659090909090909, + "k": 1, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335 + ] + }, + { + "test_f1": 0.9286120425422622, + "test_acc": 0.9287356321839081, + "test_auc": 0.9724653660912386, + "val_auc": 0.9756493506493507, + "k": 2, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443 + ] + }, + { + "test_f1": 0.9760919439152115, + "test_acc": 0.9760919540229885, + "test_auc": 0.9960157476053899, + "val_auc": 0.9987824675324676, + "k": 5, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1281, + 1079 + ] + }, + { + "test_f1": 0.9760919236996066, + "test_acc": 0.9760919540229885, + "test_auc": 0.996621164565182, + "val_auc": 0.9991883116883117, + "k": 10, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1281, + 1079, + 482, + 235, + 878, + 1515, + 21 + ] + }, + { + "test_f1": 0.9770108527815295, + "test_acc": 0.9770114942528736, + "test_auc": 0.9969027341847502, + "val_auc": 1.0, + "k": 16, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1281, + 1079, + 482, + 235, + 878, + 1515, + 21, + 760, + 165, + 571, + 1928, + 335, + 581 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8c09582b55d2e703ca85b0240617042886e18167 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.696056301018089, + "test_acc": 0.699874213836478, + "test_auc": 0.7161652603087069, + "val_auc": 0.7342937174869949, + "k": 1, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313 + ] + }, + { + "test_f1": 0.696056301018089, + "test_acc": 0.699874213836478, + "test_auc": 0.7161652603087069, + "val_auc": 0.7342937174869949, + "k": 2, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 19000 + ] + }, + { + "test_f1": 0.6962269149694807, + "test_acc": 0.70062893081761, + "test_auc": 0.7494621731394913, + "val_auc": 0.7785114045618247, + "k": 5, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 19000, + 4333, + 25757, + 1067 + ] + }, + { + "test_f1": 0.7115958455864763, + "test_acc": 0.7134591194968554, + "test_auc": 0.7752325477778599, + "val_auc": 0.758703481392557, + "k": 10, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 19000, + 4333, + 25757, + 1067, + 1105, + 433, + 6718, + 1928, + 121 + ] + }, + { + "test_f1": 0.7190449158213863, + "test_acc": 0.7202515723270441, + "test_auc": 0.7871720762420522, + "val_auc": 0.7799119647859144, + "k": 16, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 19000, + 4333, + 25757, + 1067, + 1105, + 433, + 6718, + 1928, + 121, + 6636, + 836, + 1019, + 1532, + 26569, + 800 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6a5cfacf2790e8112be13613b64c3b61bae877cf --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6908416280308862, + "test_acc": 0.6925898752751284, + "test_auc": 0.7496533474578096, + "val_auc": 0.7594150641025641, + "k": 1, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394 + ] + }, + { + "test_f1": 0.6955900315294232, + "test_acc": 0.6969919295671313, + "test_auc": 0.7504144758656626, + "val_auc": 0.7628205128205128, + "k": 2, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041 + ] + }, + { + "test_f1": 0.6959835854211198, + "test_acc": 0.6969919295671313, + "test_auc": 0.7474087184190922, + "val_auc": 0.764423076923077, + "k": 5, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584 + ] + }, + { + "test_f1": 0.6971867493257021, + "test_acc": 0.6977256052824652, + "test_auc": 0.7663626459278016, + "val_auc": 0.8110977564102564, + "k": 10, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584, + 24358, + 1551, + 1737, + 31, + 15741 + ] + }, + { + "test_f1": 0.6920226003557831, + "test_acc": 0.6925898752751284, + "test_auc": 0.7645324927547466, + "val_auc": 0.8080929487179488, + "k": 16, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584, + 24358, + 1551, + 1737, + 31, + 15741, + 7649, + 1257, + 1697, + 28053, + 7246, + 943 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..78d332602ab5d9c49725890b14c1e855531aacfa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6517488606669376, + "test_acc": 0.6528301886792452, + "test_auc": 0.7054650499878993, + "val_auc": 0.6502601040416166, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011 + ] + }, + { + "test_f1": 0.6880019370826372, + "test_acc": 0.6880503144654088, + "test_auc": 0.7659339276727298, + "val_auc": 0.729891956782713, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 4520 + ] + }, + { + "test_f1": 0.7353323105536477, + "test_acc": 0.7353459119496856, + "test_auc": 0.8162444723702053, + "val_auc": 0.758703481392557, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 4520, + 1927, + 24630, + 4674 + ] + }, + { + "test_f1": 0.7521596598397087, + "test_acc": 0.7522012578616353, + "test_auc": 0.8303234100121614, + "val_auc": 0.7879151660664265, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 4520, + 1927, + 24630, + 4674, + 5653, + 1520, + 1848, + 1928, + 3089 + ] + }, + { + "test_f1": 0.7612265804411268, + "test_acc": 0.7612578616352201, + "test_auc": 0.8397550628380246, + "val_auc": 0.7999199679871949, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 4520, + 1927, + 24630, + 4674, + 5653, + 1520, + 1848, + 1928, + 3089, + 7614, + 1036, + 6417, + 1470, + 7387, + 320 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9846a49c60419f7470275ba3964e64d5b8f49759 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7221530171020234, + "test_acc": 0.7222641509433962, + "test_auc": 0.7514348800401807, + "val_auc": 0.7360944377751102, + "k": 1, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1643 + ] + }, + { + "test_f1": 0.7499367903949511, + "test_acc": 0.749937106918239, + "test_auc": 0.8184231205046079, + "val_auc": 0.7999199679871949, + "k": 2, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1643, + 1503 + ] + }, + { + "test_f1": 0.768795520230425, + "test_acc": 0.7688050314465409, + "test_auc": 0.8426454550149411, + "val_auc": 0.8171268507402961, + "k": 5, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1643, + 1503, + 7615, + 2472, + 1928 + ] + }, + { + "test_f1": 0.7740086700472445, + "test_acc": 0.7740880503144654, + "test_auc": 0.8517326910633403, + "val_auc": 0.8607442977190876, + "k": 10, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1643, + 1503, + 7615, + 2472, + 1928, + 17225, + 1848, + 5891, + 7032, + 18220 + ] + }, + { + "test_f1": 0.7842505402415085, + "test_acc": 0.7844025157232705, + "test_auc": 0.8572180440468681, + "val_auc": 0.865546218487395, + "k": 16, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1643, + 1503, + 7615, + 2472, + 1928, + 17225, + 1848, + 5891, + 7032, + 18220, + 1666, + 7303, + 6327, + 7708, + 30127, + 446 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..37c4ab2beee280e7c9c61fe740f672273b8c567b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8249811058868729, + "test_acc": 0.8254088050314465, + "test_auc": 0.8536504634247357, + "val_auc": 0.8563425370148059, + "k": 1, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8439927020095286, + "test_acc": 0.8452830188679246, + "test_auc": 0.9187749800261054, + "val_auc": 0.9383753501400559, + "k": 2, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.86454362423461, + "test_acc": 0.8649056603773585, + "test_auc": 0.9307531145605389, + "val_auc": 0.9431772709083635, + "k": 5, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 7284, + 794 + ] + }, + { + "test_f1": 0.8720951326282571, + "test_acc": 0.8722012578616353, + "test_auc": 0.9423039241994492, + "val_auc": 0.9751900760304122, + "k": 10, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 7284, + 794, + 777, + 100, + 1874, + 4675, + 1335 + ] + }, + { + "test_f1": 0.8688200654610815, + "test_acc": 0.8689308176100629, + "test_auc": 0.9447234488967017, + "val_auc": 0.9771908763505402, + "k": 16, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 7284, + 794, + 777, + 100, + 1874, + 4675, + 1335, + 1132, + 6745, + 5660, + 1928, + 1960, + 1417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..73e654f7b347457fc54f82e1f64b6d6c2f9e3277 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7885216856997785, + "test_acc": 0.7886792452830189, + "test_auc": 0.8562060586974285, + "val_auc": 0.8643457382953181, + "k": 1, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433 + ] + }, + { + "test_f1": 0.8641019190453153, + "test_acc": 0.8641509433962264, + "test_auc": 0.9420030500061263, + "val_auc": 0.9603841536614645, + "k": 2, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433, + 18480 + ] + }, + { + "test_f1": 0.9145037819852831, + "test_acc": 0.9147169811320754, + "test_auc": 0.9666873915865601, + "val_auc": 0.9819927971188476, + "k": 5, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433, + 18480, + 1116, + 444, + 7666 + ] + }, + { + "test_f1": 0.9204981533529755, + "test_acc": 0.920503144654088, + "test_auc": 0.9785246456089329, + "val_auc": 0.9939975990396158, + "k": 10, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433, + 18480, + 1116, + 444, + 7666, + 1064, + 3481, + 1335, + 6553, + 15841 + ] + }, + { + "test_f1": 0.9318091310757076, + "test_acc": 0.9318238993710691, + "test_auc": 0.9832482565245524, + "val_auc": 0.9959983993597439, + "k": 16, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433, + 18480, + 1116, + 444, + 7666, + 1064, + 3481, + 1335, + 6553, + 15841, + 1927, + 1928, + 4110, + 988, + 1500, + 993 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9430684b5291a33de6782a6fc98b77dc4c0dc98f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8026117645813822, + "test_acc": 0.8080808080808081, + "test_auc": 0.8053061224489797, + "val_auc": 0.7732371794871795, + "k": 1, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629 + ] + }, + { + "test_f1": 0.7777324263038548, + "test_acc": 0.7777777777777778, + "test_auc": 0.8357142857142857, + "val_auc": 0.8169070512820513, + "k": 2, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625 + ] + }, + { + "test_f1": 0.7676767676767676, + "test_acc": 0.7676767676767676, + "test_auc": 0.8516326530612244, + "val_auc": 0.8511618589743589, + "k": 5, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 794, + 530 + ] + }, + { + "test_f1": 0.8178098497247434, + "test_acc": 0.8181818181818182, + "test_auc": 0.8873469387755102, + "val_auc": 0.8954326923076923, + "k": 10, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 794, + 530, + 7669, + 6400, + 2046, + 777, + 1345 + ] + }, + { + "test_f1": 0.8482990353199857, + "test_acc": 0.8484848484848485, + "test_auc": 0.9040816326530612, + "val_auc": 0.8870192307692308, + "k": 16, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 794, + 530, + 7669, + 6400, + 2046, + 777, + 1345, + 765, + 1335, + 680, + 567, + 1182, + 1758 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2c929ac1558347a8b09b9113e727a3f4e07e80fb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8614072494669509, + "test_acc": 0.8614072494669509, + "test_auc": 0.9262865975631933, + "val_auc": 0.958, + "k": 1, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417 + ] + }, + { + "test_f1": 0.8848320875161462, + "test_acc": 0.8848614072494669, + "test_auc": 0.9540643753409711, + "val_auc": 0.9938, + "k": 2, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417, + 28188 + ] + }, + { + "test_f1": 0.938146065341912, + "test_acc": 0.9381663113006397, + "test_auc": 0.986779414438989, + "val_auc": 0.992, + "k": 5, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417, + 28188, + 32655, + 12485, + 1927 + ] + }, + { + "test_f1": 0.9722743909575886, + "test_acc": 0.9722814498933902, + "test_auc": 0.9895253682487726, + "val_auc": 0.996, + "k": 10, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417, + 28188, + 32655, + 12485, + 1927, + 5406, + 1625, + 4080, + 727, + 1472 + ] + }, + { + "test_f1": 0.9701465392374662, + "test_acc": 0.9701492537313433, + "test_auc": 0.9926532096744863, + "val_auc": 0.996, + "k": 16, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1417, + 28188, + 32655, + 12485, + 1927, + 5406, + 1625, + 4080, + 727, + 1472, + 984, + 2548, + 330, + 3958, + 1348, + 943 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/scr/saebench_scr_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/scr/saebench_scr_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..5b2b64faf9507aee71619d1d46219d04d0accaaf --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/scr/saebench_scr_custom_sae_eval_results.json @@ -0,0 +1,326 @@ +{ + "eval_type_id": "scr", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": true, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "abca7db8-6978-4bdc-8f6f-ce1175773bf1", + "datetime_epoch_millis": 1776879163981, + "eval_result_metrics": { + "scr_metrics": { + "scr_dir1_threshold_2": 0.25709553284875053, + "scr_metric_threshold_2": 0.1848530083458494, + "scr_dir2_threshold_2": 0.17820511858258703, + "scr_dir1_threshold_5": 0.10305584273712019, + "scr_metric_threshold_5": 0.2462936167637781, + "scr_dir2_threshold_5": 0.235936345704021, + "scr_dir1_threshold_10": -0.09653604432334292, + "scr_metric_threshold_10": 0.3016633509439651, + "scr_dir2_threshold_10": 0.28370061329561863, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results", + "scr_dir1_threshold_2": 0.5128207283533913, + "scr_metric_threshold_2": 0.06476677137111979, + "scr_dir2_threshold_2": 0.06476677137111979, + "scr_dir1_threshold_5": 0.4615385791018498, + "scr_metric_threshold_5": 0.21502589713474632, + "scr_dir2_threshold_5": 0.21502589713474632, + "scr_dir1_threshold_10": 0.6025646805840948, + "scr_metric_threshold_10": 0.20725382280374663, + "scr_dir2_threshold_10": 0.20725382280374663, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results", + "scr_dir1_threshold_2": 0.24999988711244006, + "scr_metric_threshold_2": 0.2966360744784594, + "scr_dir2_threshold_2": 0.2966360744784594, + "scr_dir1_threshold_5": -0.08333314518740012, + "scr_metric_threshold_5": 0.27522944141155453, + "scr_dir2_threshold_5": 0.27522944141155453, + "scr_dir1_threshold_10": -0.07575748681586188, + "scr_metric_threshold_10": 0.44342517065701814, + "scr_dir2_threshold_10": 0.44342517065701814, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results", + "scr_dir1_threshold_2": 0.42187508731147705, + "scr_metric_threshold_2": 0.05025135035200431, + "scr_dir2_threshold_2": 0.05025135035200431, + "scr_dir1_threshold_5": -0.21874994179234863, + "scr_metric_threshold_5": 0.04020108028160345, + "scr_dir2_threshold_5": 0.04020108028160345, + "scr_dir1_threshold_10": -2.03124959254644, + "scr_metric_threshold_10": 0.1482412214577287, + "scr_dir2_threshold_10": 0.1482412214577287, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results", + "scr_dir1_threshold_2": 0.21476475165062875, + "scr_metric_threshold_2": 0.04777065348328394, + "scr_dir2_threshold_2": 0.04777065348328394, + "scr_dir1_threshold_5": -0.3422822980192455, + "scr_metric_threshold_5": 0.054140187841943674, + "scr_dir2_threshold_5": 0.054140187841943674, + "scr_dir1_threshold_10": -0.48993301477515244, + "scr_metric_threshold_10": -0.06687887691185689, + "scr_dir2_threshold_10": -0.06687887691185689, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results", + "scr_dir1_threshold_2": 0.04210534901500186, + "scr_metric_threshold_2": 0.48412707235506197, + "scr_dir2_threshold_2": 0.48412707235506197, + "scr_dir1_threshold_5": 0.07894729412163301, + "scr_metric_threshold_5": 0.5595238630238142, + "scr_dir2_threshold_5": 0.5595238630238142, + "scr_dir1_threshold_10": 0.06842111372221286, + "scr_metric_threshold_10": 0.5912697183136902, + "scr_dir2_threshold_10": 0.5912697183136902, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results", + "scr_dir1_threshold_2": 0.2896173902280937, + "scr_metric_threshold_2": 0.12213741326330622, + "scr_dir2_threshold_2": 0.12213741326330622, + "scr_dir1_threshold_5": 0.3934425108216648, + "scr_metric_threshold_5": 0.18702297804342538, + "scr_dir2_threshold_5": 0.18702297804342538, + "scr_dir1_threshold_10": 0.4699452484015855, + "scr_metric_threshold_10": 0.2519083153249215, + "scr_dir2_threshold_10": 0.2519083153249215, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results", + "scr_dir1_threshold_2": 0.19999988910765484, + "scr_metric_threshold_2": 0.287553551452243, + "scr_dir2_threshold_2": 0.287553551452243, + "scr_dir1_threshold_5": 0.3162791071613722, + "scr_metric_threshold_5": 0.4206008506837011, + "scr_dir2_threshold_5": 0.4206008506837011, + "scr_dir1_threshold_10": 0.39069757384182646, + "scr_metric_threshold_10": 0.5450643129054811, + "scr_dir2_threshold_10": 0.5450643129054811, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results", + "scr_dir1_threshold_2": 0.12558118001131657, + "scr_metric_threshold_2": 0.12558118001131657, + "scr_dir2_threshold_2": 0.07239806190521783, + "scr_dir1_threshold_5": 0.21860463568943592, + "scr_metric_threshold_5": 0.21860463568943592, + "scr_dir2_threshold_5": 0.13574646721137937, + "scr_dir1_threshold_10": 0.29302312300099165, + "scr_metric_threshold_10": 0.29302312300099165, + "scr_dir2_threshold_10": 0.1493212218142196, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_scr", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": null +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/tpp/saebench_tpp_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/tpp/saebench_tpp_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..6d5a331bd616f475394c09152bcc0401f559a890 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-146484-tokens-600000000/tpp/saebench_tpp_custom_sae_eval_results.json @@ -0,0 +1,297 @@ +{ + "eval_type_id": "tpp", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": false, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "f6246ad6-d9b2-40b8-8571-fab1cd5642f6", + "datetime_epoch_millis": 1776879341469, + "eval_result_metrics": { + "tpp_metrics": { + "tpp_threshold_2_total_metric": 0.04204999655485153, + "tpp_threshold_2_intended_diff_only": 0.04960000514984131, + "tpp_threshold_2_unintended_diff_only": 0.007550008594989777, + "tpp_threshold_5_total_metric": 0.10030000805854797, + "tpp_threshold_5_intended_diff_only": 0.11940000653266906, + "tpp_threshold_5_unintended_diff_only": 0.019099998474121093, + "tpp_threshold_10_total_metric": 0.13612501472234725, + "tpp_threshold_10_intended_diff_only": 0.18060001730918884, + "tpp_threshold_10_unintended_diff_only": 0.04447500258684158, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results", + "tpp_threshold_2_total_metric": 0.07470000088214875, + "tpp_threshold_2_intended_diff_only": 0.0882000207901001, + "tpp_threshold_2_unintended_diff_only": 0.013500019907951355, + "tpp_threshold_5_total_metric": 0.1766500174999237, + "tpp_threshold_5_intended_diff_only": 0.2104000210762024, + "tpp_threshold_5_unintended_diff_only": 0.033750003576278685, + "tpp_threshold_10_total_metric": 0.2286500245332718, + "tpp_threshold_10_intended_diff_only": 0.309600031375885, + "tpp_threshold_10_unintended_diff_only": 0.08095000684261322, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results", + "tpp_threshold_2_total_metric": 0.009399992227554322, + "tpp_threshold_2_intended_diff_only": 0.01099998950958252, + "tpp_threshold_2_unintended_diff_only": 0.0015999972820281982, + "tpp_threshold_5_total_metric": 0.02394999861717224, + "tpp_threshold_5_intended_diff_only": 0.02839999198913574, + "tpp_threshold_5_unintended_diff_only": 0.004449993371963501, + "tpp_threshold_10_total_metric": 0.04360000491142273, + "tpp_threshold_10_intended_diff_only": 0.05160000324249268, + "tpp_threshold_10_unintended_diff_only": 0.007999998331069947, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_tpp", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1": { + "0": { + "tpp_threshold_2_total_metric": 0.15850000083446503, + "tpp_threshold_2_intended_diff_only": 0.2070000171661377, + "tpp_threshold_2_unintended_diff_only": 0.04850001633167267, + "tpp_threshold_5_total_metric": 0.1900000274181366, + "tpp_threshold_5_intended_diff_only": 0.24800002574920654, + "tpp_threshold_5_unintended_diff_only": 0.057999998331069946, + "tpp_threshold_10_total_metric": 0.22124998271465302, + "tpp_threshold_10_intended_diff_only": 0.3100000023841858, + "tpp_threshold_10_unintended_diff_only": 0.08875001966953278 + }, + "1": { + "tpp_threshold_2_total_metric": -0.000500023365020752, + "tpp_threshold_2_intended_diff_only": 0.0, + "tpp_threshold_2_unintended_diff_only": 0.000500023365020752, + "tpp_threshold_5_total_metric": 0.10349999368190765, + "tpp_threshold_5_intended_diff_only": 0.13899999856948853, + "tpp_threshold_5_unintended_diff_only": 0.03550000488758087, + "tpp_threshold_10_total_metric": 0.15775001049041748, + "tpp_threshold_10_intended_diff_only": 0.19700002670288086, + "tpp_threshold_10_unintended_diff_only": 0.03925001621246338 + }, + "2": { + "tpp_threshold_2_total_metric": 0.1537499874830246, + "tpp_threshold_2_intended_diff_only": 0.16600000858306885, + "tpp_threshold_2_unintended_diff_only": 0.01225002110004425, + "tpp_threshold_5_total_metric": 0.20325003564357758, + "tpp_threshold_5_intended_diff_only": 0.2380000352859497, + "tpp_threshold_5_unintended_diff_only": 0.03474999964237213, + "tpp_threshold_10_total_metric": 0.21800003945827484, + "tpp_threshold_10_intended_diff_only": 0.3030000329017639, + "tpp_threshold_10_unintended_diff_only": 0.08499999344348907 + }, + "6": { + "tpp_threshold_2_total_metric": 0.05275002121925354, + "tpp_threshold_2_intended_diff_only": 0.057000041007995605, + "tpp_threshold_2_unintended_diff_only": 0.004250019788742065, + "tpp_threshold_5_total_metric": 0.24500000476837158, + "tpp_threshold_5_intended_diff_only": 0.2630000114440918, + "tpp_threshold_5_unintended_diff_only": 0.018000006675720215, + "tpp_threshold_10_total_metric": 0.3542500287294388, + "tpp_threshold_10_intended_diff_only": 0.4140000343322754, + "tpp_threshold_10_unintended_diff_only": 0.05975000560283661 + }, + "9": { + "tpp_threshold_2_total_metric": 0.009000018239021301, + "tpp_threshold_2_intended_diff_only": 0.01100003719329834, + "tpp_threshold_2_unintended_diff_only": 0.0020000189542770386, + "tpp_threshold_5_total_metric": 0.14150002598762512, + "tpp_threshold_5_intended_diff_only": 0.1640000343322754, + "tpp_threshold_5_unintended_diff_only": 0.02250000834465027, + "tpp_threshold_10_total_metric": 0.19200006127357483, + "tpp_threshold_10_intended_diff_only": 0.3240000605583191, + "tpp_threshold_10_unintended_diff_only": 0.13199999928474426 + } + }, + "canrager/amazon_reviews_mcauley_1and5": { + "1": { + "tpp_threshold_2_total_metric": 0.0104999840259552, + "tpp_threshold_2_intended_diff_only": 0.0059999823570251465, + "tpp_threshold_2_unintended_diff_only": -0.004500001668930054, + "tpp_threshold_5_total_metric": 0.011499986052513123, + "tpp_threshold_5_intended_diff_only": 0.010999977588653564, + "tpp_threshold_5_unintended_diff_only": -0.0005000084638595581, + "tpp_threshold_10_total_metric": 0.016749978065490723, + "tpp_threshold_10_intended_diff_only": 0.02499997615814209, + "tpp_threshold_10_unintended_diff_only": 0.008249998092651367 + }, + "2": { + "tpp_threshold_2_total_metric": 0.005500033497810364, + "tpp_threshold_2_intended_diff_only": 0.012000024318695068, + "tpp_threshold_2_unintended_diff_only": 0.006499990820884705, + "tpp_threshold_5_total_metric": 0.0335000604391098, + "tpp_threshold_5_intended_diff_only": 0.04300004243850708, + "tpp_threshold_5_unintended_diff_only": 0.009499981999397278, + "tpp_threshold_10_total_metric": 0.0675000548362732, + "tpp_threshold_10_intended_diff_only": 0.08000004291534424, + "tpp_threshold_10_unintended_diff_only": 0.012499988079071045 + }, + "3": { + "tpp_threshold_2_total_metric": 0.01624998450279236, + "tpp_threshold_2_intended_diff_only": 0.014999985694885254, + "tpp_threshold_2_unintended_diff_only": -0.0012499988079071045, + "tpp_threshold_5_total_metric": 0.019750013947486877, + "tpp_threshold_5_intended_diff_only": 0.022000014781951904, + "tpp_threshold_5_unintended_diff_only": 0.002250000834465027, + "tpp_threshold_10_total_metric": 0.015749990940093994, + "tpp_threshold_10_intended_diff_only": 0.02399998903274536, + "tpp_threshold_10_unintended_diff_only": 0.008249998092651367 + }, + "5": { + "tpp_threshold_2_total_metric": -0.0034999996423721313, + "tpp_threshold_2_intended_diff_only": 0.004999995231628418, + "tpp_threshold_2_unintended_diff_only": 0.00849999487400055, + "tpp_threshold_5_total_metric": 0.01024995744228363, + "tpp_threshold_5_intended_diff_only": 0.012999951839447021, + "tpp_threshold_5_unintended_diff_only": 0.002749994397163391, + "tpp_threshold_10_total_metric": 0.01874999701976776, + "tpp_threshold_10_intended_diff_only": 0.023000001907348633, + "tpp_threshold_10_unintended_diff_only": 0.004250004887580872 + }, + "6": { + "tpp_threshold_2_total_metric": 0.018249958753585815, + "tpp_threshold_2_intended_diff_only": 0.01699995994567871, + "tpp_threshold_2_unintended_diff_only": -0.0012499988079071045, + "tpp_threshold_5_total_metric": 0.04474997520446777, + "tpp_threshold_5_intended_diff_only": 0.05299997329711914, + "tpp_threshold_5_unintended_diff_only": 0.008249998092651367, + "tpp_threshold_10_total_metric": 0.09925000369548798, + "tpp_threshold_10_intended_diff_only": 0.10600000619888306, + "tpp_threshold_10_unintended_diff_only": 0.006750002503395081 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..95c97d5d195bbedc8b332d9c207fab5c83b228ad --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9756504661239563, + "test_acc": 0.9756563245823389, + "test_auc": 0.9797421932529875, + "val_auc": 0.9856284856284856, + "k": 1, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431 + ] + }, + { + "test_f1": 0.9756504661239563, + "test_acc": 0.9756563245823389, + "test_auc": 0.992789285271623, + "val_auc": 0.9927069927069927, + "k": 2, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928 + ] + }, + { + "test_f1": 0.9789957948374844, + "test_acc": 0.9789976133651551, + "test_auc": 0.997110975014035, + "val_auc": 0.9982839982839983, + "k": 5, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 126 + ] + }, + { + "test_f1": 0.9847242144272614, + "test_acc": 0.98472553699284, + "test_auc": 0.9987760376794476, + "val_auc": 0.9957099957099957, + "k": 10, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 126, + 1472, + 1411, + 542, + 786, + 461 + ] + }, + { + "test_f1": 0.9923627476154805, + "test_acc": 0.99236276849642, + "test_auc": 0.9990959265659062, + "val_auc": 0.993993993993994, + "k": 16, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 126, + 1472, + 1411, + 542, + 786, + 461, + 1848, + 5951, + 5020, + 1634, + 1872, + 468 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..09a03de6af85c78a76693352536e53b50010658a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9299734806089139, + "test_acc": 0.9300064808813999, + "test_auc": 0.9603519418291298, + "val_auc": 0.9594873914840844, + "k": 1, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441 + ] + }, + { + "test_f1": 0.9507451772332857, + "test_acc": 0.9507453013609851, + "test_auc": 0.9623327486677016, + "val_auc": 0.9570070276973957, + "k": 2, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441, + 6838 + ] + }, + { + "test_f1": 0.9487891635854852, + "test_acc": 0.948801036941024, + "test_auc": 0.9688749554780482, + "val_auc": 0.9632079371641175, + "k": 5, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441, + 6838, + 4110, + 1500, + 943 + ] + }, + { + "test_f1": 0.9591555217723507, + "test_acc": 0.9591704471808166, + "test_auc": 0.9823239450817525, + "val_auc": 0.9847044233154196, + "k": 10, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441, + 6838, + 4110, + 1500, + 943, + 1411, + 1697, + 30737, + 1180, + 601 + ] + }, + { + "test_f1": 0.9630454720797459, + "test_acc": 0.9630589760207389, + "test_auc": 0.9866786959940322, + "val_auc": 0.9913187267465895, + "k": 16, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1441, + 6838, + 4110, + 1500, + 943, + 1411, + 1697, + 30737, + 1180, + 601, + 1417, + 654, + 908, + 698, + 1703, + 1619 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7a9d22a60074aceb0b90fd0f5184fb8d5c7b1d0f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5609393780573025, + "test_acc": 0.579874213836478, + "test_auc": 0.5869312503101143, + "val_auc": 0.5958383353341338, + "k": 1, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921 + ] + }, + { + "test_f1": 0.5647700175765595, + "test_acc": 0.5816352201257862, + "test_auc": 0.6035505939512262, + "val_auc": 0.5958383353341337, + "k": 2, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777 + ] + }, + { + "test_f1": 0.5759677215942643, + "test_acc": 0.589937106918239, + "test_auc": 0.6273806654724522, + "val_auc": 0.5918367346938775, + "k": 5, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 1380 + ] + }, + { + "test_f1": 0.5938204160721695, + "test_acc": 0.5959748427672956, + "test_auc": 0.6260167952860596, + "val_auc": 0.6502601040416166, + "k": 10, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 1380, + 2046, + 505, + 969, + 1335, + 559 + ] + }, + { + "test_f1": 0.5928052532663504, + "test_acc": 0.5957232704402515, + "test_auc": 0.6327280492213473, + "val_auc": 0.6574629851940776, + "k": 16, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 1380, + 2046, + 505, + 969, + 1335, + 559, + 1178, + 6411, + 1533, + 1310, + 765, + 26301 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e984119d4b6fc1bcdf9c1bfa5ce22a47b0b71228 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6739381974366184, + "test_acc": 0.6835220125786163, + "test_auc": 0.6914172503566947, + "val_auc": 0.73109243697479, + "k": 1, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567 + ] + }, + { + "test_f1": 0.7110750290546596, + "test_acc": 0.7124528301886792, + "test_auc": 0.7378421257287056, + "val_auc": 0.709483793517407, + "k": 2, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470 + ] + }, + { + "test_f1": 0.7053322327147186, + "test_acc": 0.7066666666666667, + "test_auc": 0.7559622961726068, + "val_auc": 0.797719087635054, + "k": 5, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 1164, + 433, + 1345 + ] + }, + { + "test_f1": 0.7210667308294312, + "test_acc": 0.7225157232704402, + "test_auc": 0.7769714664433506, + "val_auc": 0.8135254101640655, + "k": 10, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 1164, + 433, + 1345, + 1629, + 559, + 4196, + 1731, + 1335 + ] + }, + { + "test_f1": 0.7285826650337119, + "test_acc": 0.7295597484276729, + "test_auc": 0.7810278885188332, + "val_auc": 0.8165266106442579, + "k": 16, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4470, + 1164, + 433, + 1345, + 1629, + 559, + 4196, + 1731, + 1335, + 6411, + 628, + 4505, + 993, + 794, + 1551 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a7139e6056f550f9168cc8380f05c44f39785df6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.773280817115881, + "test_acc": 0.7733333333333333, + "test_auc": 0.815963344232481, + "val_auc": 0.8609443777511006, + "k": 1, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411 + ] + }, + { + "test_f1": 0.7500201934431802, + "test_acc": 0.7511949685534591, + "test_auc": 0.8221198099518094, + "val_auc": 0.8673469387755103, + "k": 2, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786 + ] + }, + { + "test_f1": 0.7722777255480272, + "test_acc": 0.7723270440251573, + "test_auc": 0.8457158400832778, + "val_auc": 0.868547418967587, + "k": 5, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1758 + ] + }, + { + "test_f1": 0.7994760533217744, + "test_acc": 0.799748427672956, + "test_auc": 0.8886394613276033, + "val_auc": 0.9253701480592237, + "k": 10, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1758, + 1417, + 1335, + 181, + 5054, + 1360 + ] + }, + { + "test_f1": 0.8140659547310335, + "test_acc": 0.8143396226415094, + "test_auc": 0.9009240394556569, + "val_auc": 0.913765506202481, + "k": 16, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1758, + 1417, + 1335, + 181, + 5054, + 1360, + 1183, + 19471, + 1108, + 191, + 126, + 993 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b75258bda634bf47a7d1e115abf1b395e847e414 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8426002379192219, + "test_acc": 0.8427672955974843, + "test_auc": 0.8938525465829704, + "val_auc": 0.9403761504601841, + "k": 1, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8749475644064316, + "test_acc": 0.8749685534591195, + "test_auc": 0.9357587396548389, + "val_auc": 0.9611844737895159, + "k": 2, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8739522126134908, + "test_acc": 0.8739622641509434, + "test_auc": 0.9429936437953337, + "val_auc": 0.9623849539815927, + "k": 5, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 505, + 821 + ] + }, + { + "test_f1": 0.8782357570634072, + "test_acc": 0.8782389937106918, + "test_auc": 0.944437131090519, + "val_auc": 0.9655862344937975, + "k": 10, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 505, + 821, + 4675, + 7284, + 100, + 7867, + 5256 + ] + }, + { + "test_f1": 0.8789861131430634, + "test_acc": 0.8789937106918239, + "test_auc": 0.9448624307495704, + "val_auc": 0.9611844737895159, + "k": 16, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 505, + 821, + 4675, + 7284, + 100, + 7867, + 5256, + 1697, + 6745, + 22141, + 473, + 1438, + 1108 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..faf1cc9d63c3d7772ce2f00c2d2f63f0543d6029 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5251097003112032, + "test_acc": 0.5384615384615384, + "test_auc": 0.5433273302588616, + "val_auc": 0.5703794369645043, + "k": 1, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264 + ] + }, + { + "test_f1": 0.5462748326591974, + "test_acc": 0.5476923076923077, + "test_auc": 0.5806144174773622, + "val_auc": 0.5971032231742146, + "k": 2, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945 + ] + }, + { + "test_f1": 0.5568783807498845, + "test_acc": 0.5641025641025641, + "test_auc": 0.5883630120846939, + "val_auc": 0.6164830681354548, + "k": 5, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 4531 + ] + }, + { + "test_f1": 0.5531759993591284, + "test_acc": 0.56, + "test_auc": 0.6110933450028613, + "val_auc": 0.6332109343125256, + "k": 10, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 4531, + 386, + 1523, + 470, + 1851, + 7073 + ] + }, + { + "test_f1": 0.5683676848889536, + "test_acc": 0.5702564102564103, + "test_auc": 0.6220188002827616, + "val_auc": 0.6523867809057528, + "k": 16, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 4531, + 386, + 1523, + 470, + 1851, + 7073, + 913, + 1924, + 1588, + 3301, + 551, + 5888 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..86cd91c45993664d2a3275dd79357394aa4a2d94 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.47555090752907453, + "test_acc": 0.5046153846153846, + "test_auc": 0.5088236779210288, + "val_auc": 0.5226438188494492, + "k": 1, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341 + ] + }, + { + "test_f1": 0.4760843848570316, + "test_acc": 0.5046153846153846, + "test_auc": 0.5054111825495674, + "val_auc": 0.5324357405140759, + "k": 2, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 239 + ] + }, + { + "test_f1": 0.5107452381143942, + "test_acc": 0.5128205128205128, + "test_auc": 0.5278364526879187, + "val_auc": 0.5381476948184415, + "k": 5, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 239, + 827, + 456, + 1409 + ] + }, + { + "test_f1": 0.5085512568104517, + "test_acc": 0.5107692307692308, + "test_auc": 0.5100607600902145, + "val_auc": 0.5197878416972664, + "k": 10, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 239, + 827, + 456, + 1409, + 238, + 538, + 1264, + 326, + 400 + ] + }, + { + "test_f1": 0.5048690975909911, + "test_acc": 0.5056410256410256, + "test_auc": 0.5172476184064362, + "val_auc": 0.5589555283557732, + "k": 16, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6341, + 239, + 827, + 456, + 1409, + 238, + 538, + 1264, + 326, + 400, + 3301, + 8141, + 26171, + 6990, + 347, + 16809 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..960f24e078fd399dd350e753b40803218ea5aa4b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5363214849854526, + "test_acc": 0.5364102564102564, + "test_auc": 0.5451829535126401, + "val_auc": 0.5883312933496532, + "k": 1, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815 + ] + }, + { + "test_f1": 0.5480495278029346, + "test_acc": 0.5487179487179488, + "test_auc": 0.5548692227421146, + "val_auc": 0.6091391268869849, + "k": 2, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156 + ] + }, + { + "test_f1": 0.5316036448122602, + "test_acc": 0.5394871794871795, + "test_auc": 0.5399232504123608, + "val_auc": 0.6087311301509588, + "k": 5, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 12594, + 1770, + 16809 + ] + }, + { + "test_f1": 0.52657510722897, + "test_acc": 0.5333333333333333, + "test_auc": 0.5425804524186219, + "val_auc": 0.609547123623011, + "k": 10, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 12594, + 1770, + 16809, + 698, + 6990, + 360, + 191, + 4347 + ] + }, + { + "test_f1": 0.5237654611802373, + "test_acc": 0.5312820512820513, + "test_auc": 0.5424984010502576, + "val_auc": 0.6230110159118726, + "k": 16, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 12594, + 1770, + 16809, + 698, + 6990, + 360, + 191, + 4347, + 1611, + 14220, + 1409, + 1684, + 538, + 6129 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fc97a0cf9b0a19ce554295f4bf8aee9f3b12b170 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8270130700037241, + "test_acc": 0.8282828282828283, + "test_auc": 0.8612244897959184, + "val_auc": 0.8705929487179488, + "k": 1, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997 + ] + }, + { + "test_f1": 0.8164983164983166, + "test_acc": 0.8181818181818182, + "test_auc": 0.8783673469387756, + "val_auc": 0.8762019230769231, + "k": 2, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369 + ] + }, + { + "test_f1": 0.807255349190833, + "test_acc": 0.8080808080808081, + "test_auc": 0.8724489795918368, + "val_auc": 0.921474358974359, + "k": 5, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 2335, + 1156, + 1499 + ] + }, + { + "test_f1": 0.807255349190833, + "test_acc": 0.8080808080808081, + "test_auc": 0.8775510204081632, + "val_auc": 0.938301282051282, + "k": 10, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 2335, + 1156, + 1499, + 3526, + 314, + 4531, + 1279, + 2832 + ] + }, + { + "test_f1": 0.7977322242028124, + "test_acc": 0.797979797979798, + "test_auc": 0.8897959183673468, + "val_auc": 0.953926282051282, + "k": 16, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 2335, + 1156, + 1499, + 3526, + 314, + 4531, + 1279, + 2832, + 1734, + 12946, + 1684, + 6990, + 1473, + 224 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6d4925b9efe1c9d33f02d63bfcd8e0a520186dda --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.89867992527567, + "test_acc": 0.898989898989899, + "test_auc": 0.9020408163265304, + "val_auc": 0.9342948717948718, + "k": 1, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8684184718931782, + "test_acc": 0.8686868686868687, + "test_auc": 0.9099999999999999, + "val_auc": 0.9517227564102565, + "k": 2, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889 + ] + }, + { + "test_f1": 0.8784159103308039, + "test_acc": 0.8787878787878788, + "test_auc": 0.926530612244898, + "val_auc": 0.9801682692307693, + "k": 5, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 2335, + 2832 + ] + }, + { + "test_f1": 0.888410991636798, + "test_acc": 0.8888888888888888, + "test_auc": 0.9467346938775512, + "val_auc": 0.9837740384615383, + "k": 10, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 2335, + 2832, + 4201, + 1924, + 1116, + 3526, + 2000 + ] + }, + { + "test_f1": 0.8380531997553273, + "test_acc": 0.8383838383838383, + "test_auc": 0.9589795918367348, + "val_auc": 0.9891826923076923, + "k": 16, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 2335, + 2832, + 4201, + 1924, + 1116, + 3526, + 2000, + 581, + 2900, + 12946, + 7997, + 7481, + 5388 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8f445745c48d1b84b8c4cd9c939c4e0ce960da39 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8473644735326978, + "test_acc": 0.8484848484848485, + "test_auc": 0.8906122448979592, + "val_auc": 0.827724358974359, + "k": 1, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388 + ] + }, + { + "test_f1": 0.8780918288711562, + "test_acc": 0.8787878787878788, + "test_auc": 0.9199999999999999, + "val_auc": 0.8756009615384615, + "k": 2, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335 + ] + }, + { + "test_f1": 0.8280722400293172, + "test_acc": 0.8282828282828283, + "test_auc": 0.94, + "val_auc": 0.9419070512820513, + "k": 5, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 5888, + 1596 + ] + }, + { + "test_f1": 0.8484539175567042, + "test_acc": 0.8484848484848485, + "test_auc": 0.9571428571428572, + "val_auc": 0.9483173076923077, + "k": 10, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 5888, + 1596, + 2832, + 7384, + 4201, + 1769, + 5081 + ] + }, + { + "test_f1": 0.8786393345216874, + "test_acc": 0.8787878787878788, + "test_auc": 0.9553061224489795, + "val_auc": 0.9507211538461537, + "k": 16, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 5888, + 1596, + 2832, + 7384, + 4201, + 1769, + 5081, + 3526, + 2000, + 2900, + 1499, + 8141, + 386 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8188853d18b8558eb932d9ddb93ea2a2f4164d5d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6451343202990732, + "test_acc": 0.6662332032943217, + "test_auc": 0.6793324925858396, + "val_auc": 0.77, + "k": 1, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.6568144589814678, + "test_acc": 0.6814044213263979, + "test_auc": 0.7786589426122195, + "val_auc": 0.8655999999999999, + "k": 2, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388 + ] + }, + { + "test_f1": 0.7080905264273415, + "test_acc": 0.7087126137841352, + "test_auc": 0.8125498849358391, + "val_auc": 0.9128000000000001, + "k": 5, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 789, + 5352, + 337 + ] + }, + { + "test_f1": 0.7461878328135294, + "test_acc": 0.7477243172951885, + "test_auc": 0.8527126131664666, + "val_auc": 0.9192, + "k": 10, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 789, + 5352, + 337, + 2000, + 12946, + 2659, + 4889, + 7384 + ] + }, + { + "test_f1": 0.7876725682871819, + "test_acc": 0.788036410923277, + "test_auc": 0.8826589816934498, + "val_auc": 0.9368000000000001, + "k": 16, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 789, + 5352, + 337, + 2000, + 12946, + 2659, + 4889, + 7384, + 239, + 2335, + 7481, + 1847, + 7118, + 4201 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..362f0ad24b0386895823d7216798a987afac0afe --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6297971644100129, + "test_acc": 0.6501950585175552, + "test_auc": 0.6680477873259569, + "val_auc": 0.6923999999999999, + "k": 1, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.722156063507838, + "test_acc": 0.7256176853055917, + "test_auc": 0.7595880537697605, + "val_auc": 0.7448, + "k": 2, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384 + ] + }, + { + "test_f1": 0.7708867776228415, + "test_acc": 0.7715648027741656, + "test_auc": 0.8677513712250914, + "val_auc": 0.9114, + "k": 5, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 3526, + 4201, + 5388 + ] + }, + { + "test_f1": 0.8031646608130136, + "test_acc": 0.8036410923276983, + "test_auc": 0.9022352960628667, + "val_auc": 0.9072, + "k": 10, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 3526, + 4201, + 5388, + 7481, + 1924, + 2832, + 2900, + 1091 + ] + }, + { + "test_f1": 0.822570461818294, + "test_acc": 0.8227134807108799, + "test_auc": 0.9174980196338088, + "val_auc": 0.9296, + "k": 16, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 3526, + 4201, + 5388, + 7481, + 1924, + 2832, + 2900, + 1091, + 3842, + 6244, + 2659, + 789, + 3825, + 5352 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1ac6a0ff7cf1441a5541d31de9a42912654b376d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.818299379170509, + "test_acc": 0.8192457737321196, + "test_auc": 0.8508983422042714, + "val_auc": 0.8692, + "k": 1, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8242875609255654, + "test_acc": 0.8266146510619853, + "test_auc": 0.8508983422042714, + "val_auc": 0.8692, + "k": 2, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889 + ] + }, + { + "test_f1": 0.826570810592993, + "test_acc": 0.8287819679237104, + "test_auc": 0.9021481148567297, + "val_auc": 0.9268000000000001, + "k": 5, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 12946, + 14154 + ] + }, + { + "test_f1": 0.826570810592993, + "test_acc": 0.8287819679237104, + "test_auc": 0.9021488664188516, + "val_auc": 0.9268000000000001, + "k": 10, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 12946, + 14154, + 5352, + 4531, + 1924, + 5388, + 1158 + ] + }, + { + "test_f1": 0.8233731497940172, + "test_acc": 0.8257477243172952, + "test_auc": 0.9173492103336786, + "val_auc": 0.9516000000000001, + "k": 16, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 12946, + 14154, + 5352, + 4531, + 1924, + 5388, + 1158, + 1091, + 3526, + 4201, + 1851, + 415, + 2832 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bec1aa679090e3f416e8d337419b207766d9f50a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9333093409351353, + "test_acc": 0.9333333333333333, + "test_auc": 0.9748396842495036, + "val_auc": 0.9818441452468379, + "k": 1, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958 + ] + }, + { + "test_f1": 0.9384194331983805, + "test_acc": 0.9384615384615385, + "test_auc": 0.9748396842495036, + "val_auc": 0.9818441452468379, + "k": 2, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818 + ] + }, + { + "test_f1": 0.9661538461538461, + "test_acc": 0.9661538461538461, + "test_auc": 0.99115107550409, + "val_auc": 0.9957160342717258, + "k": 5, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 5888, + 508, + 2780 + ] + }, + { + "test_f1": 0.9640760363626538, + "test_acc": 0.9641025641025641, + "test_auc": 0.9916244487831152, + "val_auc": 0.996736026111791, + "k": 10, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 5888, + 508, + 2780, + 1499, + 4586, + 1279, + 5463, + 12605 + ] + }, + { + "test_f1": 0.9764092141414832, + "test_acc": 0.9764102564102564, + "test_auc": 0.9940333928030431, + "val_auc": 0.9971440228478172, + "k": 16, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 5888, + 508, + 2780, + 1499, + 4586, + 1279, + 5463, + 12605, + 1596, + 3475, + 730, + 779, + 386, + 1158 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f0471c5b695aa5aac681ea0bd34a4aeae42f8ecc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9278796844181461, + "test_acc": 0.9282051282051282, + "test_auc": 0.9296967044804255, + "val_auc": 0.9651162790697674, + "k": 1, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065 + ] + }, + { + "test_f1": 0.9289000582612018, + "test_acc": 0.9292307692307692, + "test_auc": 0.9796028713771164, + "val_auc": 0.99265605875153, + "k": 2, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508 + ] + }, + { + "test_f1": 0.9351669323216011, + "test_acc": 0.9353846153846154, + "test_auc": 0.9880541623186455, + "val_auc": 0.9959200326397388, + "k": 5, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 12813, + 3291, + 3475 + ] + }, + { + "test_f1": 0.9517938576002196, + "test_acc": 0.9517948717948718, + "test_auc": 0.9936694213485037, + "val_auc": 0.9989800081599347, + "k": 10, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 12813, + 3291, + 3475, + 1279, + 992, + 1499, + 581, + 778 + ] + }, + { + "test_f1": 0.9753840974930813, + "test_acc": 0.9753846153846154, + "test_auc": 0.9971892146632106, + "val_auc": 1.0, + "k": 16, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 12813, + 3291, + 3475, + 1279, + 992, + 1499, + 581, + 778, + 2708, + 3958, + 5406, + 343, + 1482, + 305 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e604b3071b87b668932a4029be2869660ae7ea27 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9620428971737486, + "test_acc": 0.9620512820512821, + "test_auc": 0.9808483690712626, + "val_auc": 0.9983680130558955, + "k": 1, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037 + ] + }, + { + "test_f1": 0.965121527631649, + "test_acc": 0.9651282051282051, + "test_auc": 0.9824978119635102, + "val_auc": 0.9983680130558956, + "k": 2, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508 + ] + }, + { + "test_f1": 0.9723031472405105, + "test_acc": 0.9723076923076923, + "test_auc": 0.9918916417006092, + "val_auc": 1.0, + "k": 5, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 5472, + 1253 + ] + }, + { + "test_f1": 0.977433761007462, + "test_acc": 0.9774358974358974, + "test_auc": 0.9967347763153465, + "val_auc": 1.0, + "k": 10, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 5472, + 1253, + 1079, + 582, + 1279, + 26034, + 3958 + ] + }, + { + "test_f1": 0.9764075264749102, + "test_acc": 0.9764102564102564, + "test_auc": 0.997420641599623, + "val_auc": 1.0, + "k": 16, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 5472, + 1253, + 1079, + 582, + 1279, + 26034, + 3958, + 3475, + 493, + 778, + 343, + 581, + 12813 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..aca7f247910664c225653dc0dd438f78603d3232 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5984177593952192, + "test_acc": 0.623753792804508, + "test_auc": 0.6295941113604626, + "val_auc": 0.6468, + "k": 1, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351 + ] + }, + { + "test_f1": 0.6518359256535529, + "test_acc": 0.6571304724750758, + "test_auc": 0.6823176973339085, + "val_auc": 0.6623999999999999, + "k": 2, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1894 + ] + }, + { + "test_f1": 0.6719327430813585, + "test_acc": 0.6749024707412223, + "test_auc": 0.7287649880276155, + "val_auc": 0.7664000000000001, + "k": 5, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1894, + 1441, + 559, + 3078 + ] + }, + { + "test_f1": 0.7004855917743081, + "test_acc": 0.7017771998266147, + "test_auc": 0.7651578806549413, + "val_auc": 0.7851999999999999, + "k": 10, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1894, + 1441, + 559, + 3078, + 654, + 3274, + 343, + 1602, + 1028 + ] + }, + { + "test_f1": 0.7072564985777962, + "test_acc": 0.7078456870394452, + "test_auc": 0.7829349553046007, + "val_auc": 0.8200000000000001, + "k": 16, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1894, + 1441, + 559, + 3078, + 654, + 3274, + 343, + 1602, + 1028, + 52, + 619, + 4510, + 1960, + 2046, + 30138 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8da8003a0fd5a1bd72a176889efd6808376ece23 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.622456864785246, + "test_acc": 0.6415257910706545, + "test_auc": 0.6497010285879201, + "val_auc": 0.6732, + "k": 1, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274 + ] + }, + { + "test_f1": 0.6234414136804511, + "test_acc": 0.6410923276983095, + "test_auc": 0.6573064614801866, + "val_auc": 0.7160000000000002, + "k": 2, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121 + ] + }, + { + "test_f1": 0.7366418081907561, + "test_acc": 0.7368877329865626, + "test_auc": 0.7857127289070334, + "val_auc": 0.8109999999999999, + "k": 5, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 698, + 52, + 3078 + ] + }, + { + "test_f1": 0.746821427928401, + "test_acc": 0.7468573905504985, + "test_auc": 0.8168942897813104, + "val_auc": 0.8403999999999999, + "k": 10, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 698, + 52, + 3078, + 1602, + 4510, + 1524, + 567, + 1028 + ] + }, + { + "test_f1": 0.7654857404094577, + "test_acc": 0.765496315561335, + "test_auc": 0.8530158684826411, + "val_auc": 0.8532, + "k": 16, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 698, + 52, + 3078, + 1602, + 4510, + 1524, + 567, + 1028, + 1895, + 2046, + 1351, + 1894, + 1960, + 1158 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a6921f511f7cbd0a5d258618cbb6d7bcb2690c55 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5497700591991302, + "test_acc": 0.5799739921976593, + "test_auc": 0.5855210054097442, + "val_auc": 0.5528, + "k": 1, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.5510860443670675, + "test_acc": 0.58777633289987, + "test_auc": 0.6519057360724266, + "val_auc": 0.6931999999999999, + "k": 2, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508 + ] + }, + { + "test_f1": 0.6393894906442027, + "test_acc": 0.6454269614217598, + "test_auc": 0.713463183226336, + "val_auc": 0.7487999999999999, + "k": 5, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 121, + 581 + ] + }, + { + "test_f1": 0.6613623589712865, + "test_acc": 0.6627654963155614, + "test_auc": 0.7366052840829664, + "val_auc": 0.7952, + "k": 10, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 121, + 581, + 925, + 52, + 3274, + 4654, + 51 + ] + }, + { + "test_f1": 0.6752011071984935, + "test_acc": 0.6757693974859125, + "test_auc": 0.7535898364751135, + "val_auc": 0.8004, + "k": 16, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 121, + 581, + 925, + 52, + 3274, + 4654, + 51, + 7893, + 4510, + 1441, + 567, + 957, + 1524 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a9a04a7a03debc8d4fb41928c584a208b3796a46 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.59920246215814, + "test_acc": 0.6183368869936035, + "test_auc": 0.631151118385161, + "val_auc": 0.6103999999999999, + "k": 1, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504 + ] + }, + { + "test_f1": 0.5437643941755704, + "test_acc": 0.5543710021321961, + "test_auc": 0.6489907255864702, + "val_auc": 0.6872, + "k": 2, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 696 + ] + }, + { + "test_f1": 0.6700575487997347, + "test_acc": 0.6716417910447762, + "test_auc": 0.7720949263502457, + "val_auc": 0.7656, + "k": 5, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 696, + 4672, + 1048, + 927 + ] + }, + { + "test_f1": 0.6957673501919631, + "test_acc": 0.697228144989339, + "test_auc": 0.8039098017821422, + "val_auc": 0.8012, + "k": 10, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 696, + 4672, + 1048, + 927, + 1902, + 1818, + 7129, + 29085, + 1927 + ] + }, + { + "test_f1": 0.7006759614772183, + "test_acc": 0.7014925373134329, + "test_auc": 0.783106019276232, + "val_auc": 0.798, + "k": 16, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 696, + 4672, + 1048, + 927, + 1902, + 1818, + 7129, + 29085, + 1927, + 6905, + 741, + 437, + 499, + 246, + 988 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4b49a39003ddc0c15db4c7ef3b7b65804716fe74 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9292929292929293, + "test_acc": 0.9292929292929293, + "test_auc": 0.9620408163265306, + "val_auc": 0.9508856682769727, + "k": 1, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299 + ] + }, + { + "test_f1": 0.9494949494949495, + "test_acc": 0.9494949494949495, + "test_auc": 0.9702040816326529, + "val_auc": 0.9537037037037037, + "k": 2, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114 + ] + }, + { + "test_f1": 0.9595877138734282, + "test_acc": 0.9595959595959596, + "test_auc": 0.9971428571428571, + "val_auc": 0.970209339774557, + "k": 5, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 432, + 23633, + 2535 + ] + }, + { + "test_f1": 0.9393815708101421, + "test_acc": 0.9393939393939394, + "test_auc": 0.9951020408163265, + "val_auc": 0.9887278582930756, + "k": 10, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 432, + 23633, + 2535, + 6877, + 1232, + 19940, + 10739, + 373 + ] + }, + { + "test_f1": 0.979793856936714, + "test_acc": 0.9797979797979798, + "test_auc": 0.9922448979591836, + "val_auc": 0.9975845410628019, + "k": 16, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 432, + 23633, + 2535, + 6877, + 1232, + 19940, + 10739, + 373, + 55, + 1744, + 3959, + 1420, + 15752, + 1960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3ea5bd41c6b0045ecb7d991b1285cd204f038258 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7574272133095662, + "test_acc": 0.7575757575757576, + "test_auc": 0.8387755102040816, + "val_auc": 0.8079710144927537, + "k": 1, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791 + ] + }, + { + "test_f1": 0.8478331704138156, + "test_acc": 0.8484848484848485, + "test_auc": 0.9171428571428571, + "val_auc": 0.9086151368760065, + "k": 2, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 114 + ] + }, + { + "test_f1": 0.92920621648266, + "test_acc": 0.9292929292929293, + "test_auc": 0.9853061224489796, + "val_auc": 0.9593397745571659, + "k": 5, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 114, + 278, + 1657, + 385 + ] + }, + { + "test_f1": 0.9393568033273916, + "test_acc": 0.9393939393939394, + "test_auc": 0.9873469387755102, + "val_auc": 0.9851046698872785, + "k": 10, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 114, + 278, + 1657, + 385, + 1396, + 521, + 4580, + 5387, + 429 + ] + }, + { + "test_f1": 0.9595712022182611, + "test_acc": 0.9595959595959596, + "test_auc": 0.9926530612244898, + "val_auc": 0.9911433172302737, + "k": 16, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 114, + 278, + 1657, + 385, + 1396, + 521, + 4580, + 5387, + 429, + 1299, + 6088, + 23557, + 19000, + 1360, + 25757 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7034430e9b8aca3f3ba1c41a4fdd5e70faed87a3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9393568033273916, + "test_acc": 0.9393939393939394, + "test_auc": 0.9559183673469388, + "val_auc": 0.9154589371980677, + "k": 1, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9595132999388319, + "test_acc": 0.9595959595959596, + "test_auc": 0.9951020408163266, + "val_auc": 0.9786634460547504, + "k": 2, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244 + ] + }, + { + "test_f1": 0.9696784111400831, + "test_acc": 0.9696969696969697, + "test_auc": 0.9975510204081632, + "val_auc": 0.9879227053140097, + "k": 5, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244, + 1610, + 1501, + 1550 + ] + }, + { + "test_f1": 0.9696784111400831, + "test_acc": 0.9696969696969697, + "test_auc": 0.9975510204081632, + "val_auc": 0.9911433172302737, + "k": 10, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244, + 1610, + 1501, + 1550, + 177, + 25895, + 25152, + 114, + 15752 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 0.9983673469387755, + "val_auc": 0.9959742351046699, + "k": 16, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244, + 1610, + 1501, + 1550, + 177, + 25895, + 25152, + 114, + 15752, + 3644, + 7577, + 23791, + 1360, + 1960, + 5948 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1db1522ab16774631bb25cb524b831ffea188825 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8754522127315547, + "test_acc": 0.8755555555555555, + "test_auc": 0.9322060682680151, + "val_auc": 0.9119318181818181, + "k": 1, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752 + ] + }, + { + "test_f1": 0.9199968394437421, + "test_acc": 0.92, + "test_auc": 0.9490360303413401, + "val_auc": 0.9228896103896105, + "k": 2, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309 + ] + }, + { + "test_f1": 0.937775319567355, + "test_acc": 0.9377777777777778, + "test_auc": 0.9599399494310998, + "val_auc": 0.9811282467532467, + "k": 5, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580 + ] + }, + { + "test_f1": 0.9466645596291614, + "test_acc": 0.9466666666666667, + "test_auc": 0.9627054361567636, + "val_auc": 0.9874188311688311, + "k": 10, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 5563, + 3301, + 12400, + 943, + 1541 + ] + }, + { + "test_f1": 0.951109179545599, + "test_acc": 0.9511111111111111, + "test_auc": 0.9689475347661188, + "val_auc": 0.9902597402597403, + "k": 16, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 5563, + 3301, + 12400, + 943, + 1541, + 1073, + 1295, + 969, + 1232, + 1915, + 1327 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f011fe0aeff98204e35b4c998c9767714905f9f9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7997151955064179, + "test_acc": 0.8, + "test_auc": 0.8264064475347661, + "val_auc": 0.8429383116883117, + "k": 1, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406 + ] + }, + { + "test_f1": 0.8165388362060397, + "test_acc": 0.8177777777777778, + "test_auc": 0.8701801517067003, + "val_auc": 0.854301948051948, + "k": 2, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854 + ] + }, + { + "test_f1": 0.7943127074945256, + "test_acc": 0.7955555555555556, + "test_auc": 0.8968473451327433, + "val_auc": 0.8743912337662338, + "k": 5, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1232, + 1804 + ] + }, + { + "test_f1": 0.8394228634850166, + "test_acc": 0.84, + "test_auc": 0.9185366624525916, + "val_auc": 0.885551948051948, + "k": 10, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1232, + 1804, + 1182, + 15752, + 6662, + 1570, + 311 + ] + }, + { + "test_f1": 0.8259448597698413, + "test_acc": 0.8266666666666667, + "test_auc": 0.9224083438685209, + "val_auc": 0.8916396103896104, + "k": 16, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1232, + 1804, + 1182, + 15752, + 6662, + 1570, + 311, + 800, + 3481, + 322, + 5475, + 920, + 1858 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a656c3c261a174c1bd4f8f659cfa16ee7cae74fa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9199810336652442, + "test_acc": 0.92, + "test_auc": 0.9518805309734513, + "val_auc": 0.9419642857142856, + "k": 1, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9243817167104839, + "test_acc": 0.9244444444444444, + "test_auc": 0.9631005056890012, + "val_auc": 0.9586038961038961, + "k": 2, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16 + ] + }, + { + "test_f1": 0.9555292050944225, + "test_acc": 0.9555555555555556, + "test_auc": 0.9798119469026549, + "val_auc": 0.9655032467532467, + "k": 5, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 21467 + ] + }, + { + "test_f1": 0.9466540224434961, + "test_acc": 0.9466666666666667, + "test_auc": 0.9897281921618204, + "val_auc": 0.9862012987012987, + "k": 10, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 21467, + 15752, + 1960, + 20486, + 79, + 1835 + ] + }, + { + "test_f1": 0.9510917908760189, + "test_acc": 0.9511111111111111, + "test_auc": 0.9896491782553729, + "val_auc": 0.9874188311688312, + "k": 16, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 21467, + 15752, + 1960, + 20486, + 79, + 1835, + 1417, + 7033, + 4580, + 969, + 177, + 1596 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5b030dac5a562a7edfaebf7cfc8f556323463d25 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6550287732211579, + "test_acc": 0.6601647160814911, + "test_auc": 0.6811572854177408, + "val_auc": 0.7288000000000001, + "k": 1, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649 + ] + }, + { + "test_f1": 0.6967552450035771, + "test_acc": 0.7026441265713047, + "test_auc": 0.758119501383626, + "val_auc": 0.802, + "k": 2, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394 + ] + }, + { + "test_f1": 0.7130905779685286, + "test_acc": 0.7152145643693107, + "test_auc": 0.778696520718313, + "val_auc": 0.8508000000000001, + "k": 5, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 943, + 28053, + 1928 + ] + }, + { + "test_f1": 0.7335097348969201, + "test_acc": 0.7347204161248374, + "test_auc": 0.7987369246979847, + "val_auc": 0.8644, + "k": 10, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 943, + 28053, + 1928, + 7614, + 13, + 7246, + 1737, + 1560 + ] + }, + { + "test_f1": 0.7462858331956599, + "test_acc": 0.7468573905504985, + "test_auc": 0.8204127278548463, + "val_auc": 0.8716, + "k": 16, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 943, + 28053, + 1928, + 7614, + 13, + 7246, + 1737, + 1560, + 253, + 811, + 1551, + 17225, + 584, + 1848 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..57150634cca3d125a492687c1e34a16ed5fc989f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6150693371041017, + "test_acc": 0.6150845253576073, + "test_auc": 0.6450879402838801, + "val_auc": 0.6482, + "k": 1, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943 + ] + }, + { + "test_f1": 0.6375366757371758, + "test_acc": 0.6380580840918942, + "test_auc": 0.6773325857795427, + "val_auc": 0.7064, + "k": 2, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253 + ] + }, + { + "test_f1": 0.6520403115757332, + "test_acc": 0.6527958387516255, + "test_auc": 0.7111994781152626, + "val_auc": 0.76, + "k": 5, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 7649, + 1011, + 28053 + ] + }, + { + "test_f1": 0.6446697181809008, + "test_acc": 0.6454269614217598, + "test_auc": 0.7132065247617172, + "val_auc": 0.764, + "k": 10, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 7649, + 1011, + 28053, + 1551, + 5394, + 32305, + 13, + 7614 + ] + }, + { + "test_f1": 0.6716348875593185, + "test_acc": 0.6718682271348071, + "test_auc": 0.7336756949319161, + "val_auc": 0.7595999999999999, + "k": 16, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 7649, + 1011, + 28053, + 1551, + 5394, + 32305, + 13, + 7614, + 1928, + 1411, + 30127, + 1875, + 4674, + 319 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2dce6b9c9f5bc2dcba8c4b552e00c48b50610536 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6125557653454348, + "test_acc": 0.6259211096662332, + "test_auc": 0.643935043988931, + "val_auc": 0.6902, + "k": 1, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253 + ] + }, + { + "test_f1": 0.6207555101976201, + "test_acc": 0.635023840485479, + "test_auc": 0.7060685635092541, + "val_auc": 0.7388, + "k": 2, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394 + ] + }, + { + "test_f1": 0.6868937295483557, + "test_acc": 0.6900736887732987, + "test_auc": 0.7609724311982455, + "val_auc": 0.7522, + "k": 5, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 1541, + 794, + 1099 + ] + }, + { + "test_f1": 0.7280191777051398, + "test_acc": 0.729952319029042, + "test_auc": 0.8113736902151121, + "val_auc": 0.7699999999999999, + "k": 10, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 1541, + 794, + 1099, + 844, + 7649, + 100, + 28053, + 1960 + ] + }, + { + "test_f1": 0.7365576048287248, + "test_acc": 0.7381881231035977, + "test_auc": 0.8215862921081469, + "val_auc": 0.7796000000000001, + "k": 16, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 1541, + 794, + 1099, + 844, + 7649, + 100, + 28053, + 1960, + 7246, + 7143, + 1737, + 24568, + 4481, + 4675 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a8478150e226f2c92d8ee616407bacf3853c0907 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8092508746726291, + "test_acc": 0.8128813559322033, + "test_auc": 0.8224233599188094, + "val_auc": 0.7964975845410629, + "k": 1, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.8092508746726291, + "test_acc": 0.8128813559322033, + "test_auc": 0.8439197214224518, + "val_auc": 0.8756038647342995, + "k": 2, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5020 + ] + }, + { + "test_f1": 0.8085961610293066, + "test_acc": 0.8122033898305084, + "test_auc": 0.8940763661367959, + "val_auc": 0.8812399355877617, + "k": 5, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5020, + 1847, + 266, + 6838 + ] + }, + { + "test_f1": 0.8389070164332545, + "test_acc": 0.84, + "test_auc": 0.9254448378947833, + "val_auc": 0.9367954911433172, + "k": 10, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5020, + 1847, + 266, + 6838, + 4413, + 776, + 508, + 1697, + 1103 + ] + }, + { + "test_f1": 0.8640874404682592, + "test_acc": 0.864406779661017, + "test_auc": 0.9382430052251677, + "val_auc": 0.9468599033816425, + "k": 16, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5020, + 1847, + 266, + 6838, + 4413, + 776, + 508, + 1697, + 1103, + 7335, + 32278, + 1180, + 27673, + 1433, + 1628 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b46a96634cb754b6dd4b76c88c049fc6de4b7278 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7844110851799176, + "test_acc": 0.7911864406779661, + "test_auc": 0.796876666188643, + "val_auc": 0.710950080515298, + "k": 1, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528 + ] + }, + { + "test_f1": 0.7843036275307428, + "test_acc": 0.7911864406779661, + "test_auc": 0.8454448011237237, + "val_auc": 0.7528180354267311, + "k": 2, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 1628 + ] + }, + { + "test_f1": 0.8119192241256458, + "test_acc": 0.8128813559322033, + "test_auc": 0.8889918478560633, + "val_auc": 0.8289049919484702, + "k": 5, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 1628, + 4791, + 1815, + 2023 + ] + }, + { + "test_f1": 0.8224115753930111, + "test_acc": 0.8230508474576271, + "test_auc": 0.9124701694778142, + "val_auc": 0.8393719806763286, + "k": 10, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 1628, + 4791, + 1815, + 2023, + 416, + 1103, + 1866, + 203, + 4558 + ] + }, + { + "test_f1": 0.8211993641221327, + "test_acc": 0.8223728813559322, + "test_auc": 0.9162575886274466, + "val_auc": 0.8353462157809984, + "k": 16, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 1628, + 4791, + 1815, + 2023, + 416, + 1103, + 1866, + 203, + 4558, + 1180, + 16223, + 20997, + 31149, + 4185, + 461 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..94c99374d749649421fb9214ea9924a3bad646cb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6560922559198916, + "test_acc": 0.6610169491525424, + "test_auc": 0.6917095968788725, + "val_auc": 0.7159822866344606, + "k": 1, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020 + ] + }, + { + "test_f1": 0.7083087848898152, + "test_acc": 0.7111864406779661, + "test_auc": 0.7817003305718266, + "val_auc": 0.8357487922705313, + "k": 2, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266 + ] + }, + { + "test_f1": 0.7281529838804975, + "test_acc": 0.7294915254237289, + "test_auc": 0.8083639452405378, + "val_auc": 0.8993558776167472, + "k": 5, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 431, + 1866, + 373 + ] + }, + { + "test_f1": 0.7888698485321134, + "test_acc": 0.7911864406779661, + "test_auc": 0.8709004864811198, + "val_auc": 0.9404186795491143, + "k": 10, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 431, + 1866, + 373, + 1103, + 1628, + 32278, + 4413, + 18827 + ] + }, + { + "test_f1": 0.8059384928592034, + "test_acc": 0.8074576271186441, + "test_auc": 0.8797108323864785, + "val_auc": 0.9219001610305959, + "k": 16, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 431, + 1866, + 373, + 1103, + 1628, + 32278, + 4413, + 18827, + 895, + 27673, + 416, + 16223, + 5804, + 1758 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..72080f5320293e03305056346cb07d6a62b0550c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5766058671220736, + "test_acc": 0.5801749271137027, + "test_auc": 0.5888072895416837, + "val_auc": 0.5254101640656262, + "k": 1, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381 + ] + }, + { + "test_f1": 0.5448401657747003, + "test_acc": 0.5539358600583091, + "test_auc": 0.5888922888616892, + "val_auc": 0.5554221688675469, + "k": 2, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219 + ] + }, + { + "test_f1": 0.6239067055393586, + "test_acc": 0.6239067055393586, + "test_auc": 0.6576907384740922, + "val_auc": 0.6318527410964385, + "k": 5, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219, + 673, + 1122, + 5897 + ] + }, + { + "test_f1": 0.6587296752389654, + "test_acc": 0.6588921282798834, + "test_auc": 0.7172412620699035, + "val_auc": 0.6784713885554221, + "k": 10, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219, + 673, + 1122, + 5897, + 1799, + 993, + 1411, + 799, + 275 + ] + }, + { + "test_f1": 0.705524340845061, + "test_acc": 0.7055393586005831, + "test_auc": 0.7479940160478716, + "val_auc": 0.7523009203681472, + "k": 16, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219, + 673, + 1122, + 5897, + 1799, + 993, + 1411, + 799, + 275, + 1697, + 1508, + 304, + 8006, + 801, + 31278 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..938b7a238598e9afbf297f4a45abb75fcbf9e748 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7307295042357645, + "test_acc": 0.7434402332361516, + "test_auc": 0.7504929960560316, + "val_auc": 0.7192877150860345, + "k": 1, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7794746793986154, + "test_acc": 0.7842565597667639, + "test_auc": 0.800999592003264, + "val_auc": 0.8147258903561425, + "k": 2, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275 + ] + }, + { + "test_f1": 0.8181566923235742, + "test_acc": 0.8192419825072886, + "test_auc": 0.8843669250645995, + "val_auc": 0.8755502200880352, + "k": 5, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 8006, + 191, + 534 + ] + }, + { + "test_f1": 0.8297283044833185, + "test_acc": 0.8309037900874635, + "test_auc": 0.906296749626003, + "val_auc": 0.9111644657863145, + "k": 10, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 8006, + 191, + 534, + 433, + 1928, + 1799, + 1122, + 142 + ] + }, + { + "test_f1": 0.8856655490431259, + "test_acc": 0.8862973760932945, + "test_auc": 0.9454304365565076, + "val_auc": 0.9319727891156462, + "k": 16, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 8006, + 191, + 534, + 433, + 1928, + 1799, + 1122, + 142, + 799, + 1411, + 801, + 786, + 1239, + 506 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a69fd83542a0a7b82607d78856c0414e75082763 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6863732188712761, + "test_acc": 0.6909620991253644, + "test_auc": 0.6946484428124575, + "val_auc": 0.6914765906362546, + "k": 1, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.6907255261541587, + "test_acc": 0.6909620991253644, + "test_auc": 0.7453760369917042, + "val_auc": 0.7527010804321728, + "k": 2, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191 + ] + }, + { + "test_f1": 0.7696401410930706, + "test_acc": 0.7696793002915452, + "test_auc": 0.837515299877601, + "val_auc": 0.8319327731092436, + "k": 5, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 5897, + 1239, + 433 + ] + }, + { + "test_f1": 0.7871683935226884, + "test_acc": 0.7871720116618076, + "test_auc": 0.8615191078471374, + "val_auc": 0.8567426970788315, + "k": 10, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 5897, + 1239, + 433, + 275, + 534, + 8006, + 1464, + 1195 + ] + }, + { + "test_f1": 0.8015637189267025, + "test_acc": 0.8017492711370262, + "test_auc": 0.8675540595675234, + "val_auc": 0.8663465386154462, + "k": 16, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 5897, + 1239, + 433, + 275, + 534, + 8006, + 1464, + 1195, + 1596, + 3233, + 4003, + 142, + 1411, + 167 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f2741f8cf1eea4837fa8b6a5e92a4386f89b2c74 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320 + ] + }, + { + "test_f1": 0.5237940076649754, + "test_acc": 0.5252525252525253, + "test_auc": 0.5391836734693877, + "val_auc": 0.52, + "k": 2, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 5, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 1073, + 13612, + 859 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 10, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 1073, + 13612, + 859, + 1116, + 987, + 1030, + 1703, + 473 + ] + }, + { + "test_f1": 0.5546462993271503, + "test_acc": 0.5555555555555556, + "test_auc": 0.5551020408163265, + "val_auc": 0.5006060606060606, + "k": 16, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 2779, + 1073, + 13612, + 859, + 1116, + 987, + 1030, + 1703, + 473, + 7666, + 1600, + 1377, + 1011, + 10244, + 1848 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2d7d49aa012128192df3ec5a2a45e28434a03e49 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5333523918429579, + "test_acc": 0.5353535353535354, + "test_auc": 0.553469387755102, + "val_auc": 0.5955555555555555, + "k": 1, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520 + ] + }, + { + "test_f1": 0.5340254297823707, + "test_acc": 0.5454545454545454, + "test_auc": 0.5487755102040817, + "val_auc": 0.6355555555555555, + "k": 2, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534 + ] + }, + { + "test_f1": 0.5780112420403936, + "test_acc": 0.5858585858585859, + "test_auc": 0.5826530612244898, + "val_auc": 0.6662626262626263, + "k": 5, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534, + 1788, + 2046, + 859 + ] + }, + { + "test_f1": 0.5453617526701124, + "test_acc": 0.5454545454545454, + "test_auc": 0.5861224489795919, + "val_auc": 0.6973737373737373, + "k": 10, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534, + 1788, + 2046, + 859, + 7666, + 320, + 1116, + 347, + 1366 + ] + }, + { + "test_f1": 0.49391624923539823, + "test_acc": 0.494949494949495, + "test_auc": 0.5514285714285714, + "val_auc": 0.6872727272727271, + "k": 16, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534, + 1788, + 2046, + 859, + 7666, + 320, + 1116, + 347, + 1366, + 1030, + 1818, + 5129, + 1791, + 680, + 2489 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7eb67049b640f24a1b8fb793b43bf1a0e1f0c07d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5757575757575758, + "test_acc": 0.5757575757575758, + "test_auc": 0.5973469387755102, + "val_auc": 0.5846464646464646, + "k": 1, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666 + ] + }, + { + "test_f1": 0.5959595959595959, + "test_acc": 0.5959595959595959, + "test_auc": 0.5918367346938775, + "val_auc": 0.6016161616161616, + "k": 2, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245 + ] + }, + { + "test_f1": 0.515052566481138, + "test_acc": 0.5151515151515151, + "test_auc": 0.5824489795918367, + "val_auc": 0.6371717171717172, + "k": 5, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 7026 + ] + }, + { + "test_f1": 0.5530033725275726, + "test_acc": 0.5555555555555556, + "test_auc": 0.5973469387755102, + "val_auc": 0.6662626262626263, + "k": 10, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 7026, + 27174, + 1116, + 4985, + 373, + 111 + ] + }, + { + "test_f1": 0.5550108932461874, + "test_acc": 0.5555555555555556, + "test_auc": 0.5918367346938775, + "val_auc": 0.6379797979797981, + "k": 16, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 2779, + 7026, + 27174, + 1116, + 4985, + 373, + 111, + 5883, + 2595, + 680, + 1804, + 2214, + 16012 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ec00e2f004a30f27edcfd62efc5e1b5e83e0a8f7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5085176399022184, + "test_acc": 0.5525423728813559, + "test_auc": 0.5545792839203832, + "val_auc": 0.623389694041868, + "k": 1, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217 + ] + }, + { + "test_f1": 0.5464122160525243, + "test_acc": 0.5538983050847458, + "test_auc": 0.564919305909477, + "val_auc": 0.6795491143317232, + "k": 2, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1960 + ] + }, + { + "test_f1": 0.5832838957062655, + "test_acc": 0.5877966101694915, + "test_auc": 0.6125222005272971, + "val_auc": 0.6078904991948471, + "k": 5, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1960, + 1319, + 1541, + 539 + ] + }, + { + "test_f1": 0.5659915793780979, + "test_acc": 0.5661016949152542, + "test_auc": 0.6044647420694017, + "val_auc": 0.6084943639291466, + "k": 10, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1960, + 1319, + 1541, + 539, + 186, + 311, + 1812, + 20, + 24957 + ] + }, + { + "test_f1": 0.5748541883925075, + "test_acc": 0.5755932203389831, + "test_auc": 0.5969882663548481, + "val_auc": 0.6336553945249598, + "k": 16, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 217, + 1960, + 1319, + 1541, + 539, + 186, + 311, + 1812, + 20, + 24957, + 443, + 836, + 373, + 1255, + 26270, + 6088 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..43cb954c58048bbbe7e222faccf5fd2edfded66d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.623709788344485, + "test_acc": 0.6237288135593221, + "test_auc": 0.6468893522042412, + "val_auc": 0.6610305958132046, + "k": 1, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319 + ] + }, + { + "test_f1": 0.6508464947226172, + "test_acc": 0.6508474576271186, + "test_auc": 0.7053516600294903, + "val_auc": 0.6821658615136876, + "k": 2, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541 + ] + }, + { + "test_f1": 0.6637437135391987, + "test_acc": 0.6644067796610169, + "test_auc": 0.7232122830047839, + "val_auc": 0.7041062801932367, + "k": 5, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1960, + 1874 + ] + }, + { + "test_f1": 0.6743080960532066, + "test_acc": 0.6752542372881356, + "test_auc": 0.732177986637397, + "val_auc": 0.7411433172302737, + "k": 10, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1960, + 1874, + 1791, + 186, + 1255, + 1335, + 908 + ] + }, + { + "test_f1": 0.6751033663985004, + "test_acc": 0.6759322033898305, + "test_auc": 0.7355112831996706, + "val_auc": 0.7302737520128825, + "k": 16, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1960, + 1874, + 1791, + 186, + 1255, + 1335, + 908, + 311, + 765, + 217, + 253, + 821, + 17442 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e42b42b6e3a33aa688f5f2e20cbc1848d691c458 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5917652621269602, + "test_acc": 0.6, + "test_auc": 0.610918063047659, + "val_auc": 0.640901771336554, + "k": 1, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.5882580916415593, + "test_acc": 0.5979661016949153, + "test_auc": 0.610918063047659, + "val_auc": 0.640901771336554, + "k": 2, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255 + ] + }, + { + "test_f1": 0.597495641619723, + "test_acc": 0.6033898305084746, + "test_auc": 0.6419767386276306, + "val_auc": 0.6660628019323671, + "k": 5, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 17442, + 1791 + ] + }, + { + "test_f1": 0.6000101204972192, + "test_acc": 0.6047457627118644, + "test_auc": 0.6337841465253187, + "val_auc": 0.6950483091787439, + "k": 10, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 17442, + 1791, + 521, + 186, + 4511, + 12485, + 1942 + ] + }, + { + "test_f1": 0.6219414383558981, + "test_acc": 0.6250847457627119, + "test_auc": 0.6614405798060695, + "val_auc": 0.6892109500805152, + "k": 16, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 20, + 17442, + 1791, + 521, + 186, + 4511, + 12485, + 1942, + 5144, + 311, + 1319, + 6212, + 567, + 583 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5cf0217d181cf3eec0322d706e9db498bf8ca6ec --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5631248024639055, + "test_acc": 0.5632754342431762, + "test_auc": 0.5808580858085809, + "val_auc": 0.569827931172469, + "k": 1, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067 + ] + }, + { + "test_f1": 0.5990151076150722, + "test_acc": 0.6004962779156328, + "test_auc": 0.6458302546672577, + "val_auc": 0.6174469787915167, + "k": 2, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268 + ] + }, + { + "test_f1": 0.5922839988024678, + "test_acc": 0.5930521091811415, + "test_auc": 0.6556327274518498, + "val_auc": 0.6190476190476191, + "k": 5, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 24630, + 5731, + 124 + ] + }, + { + "test_f1": 0.5880132306548067, + "test_acc": 0.5880893300248139, + "test_auc": 0.6412491995468204, + "val_auc": 0.6446578631452582, + "k": 10, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 24630, + 5731, + 124, + 1441, + 4802, + 8006, + 1619, + 1551 + ] + }, + { + "test_f1": 0.6077230758958396, + "test_acc": 0.607940446650124, + "test_auc": 0.6616792276242549, + "val_auc": 0.6610644257703082, + "k": 16, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 24630, + 5731, + 124, + 1441, + 4802, + 8006, + 1619, + 1551, + 1101, + 2480, + 2, + 1373, + 898, + 8013 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4597b5291c94e697cefee48f4ab48e31c0addd2b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5739869169758436, + "test_acc": 0.6054590570719603, + "test_auc": 0.6127038076942023, + "val_auc": 0.6276510604241696, + "k": 1, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575 + ] + }, + { + "test_f1": 0.6298581259125967, + "test_acc": 0.630272952853598, + "test_auc": 0.6545490369932516, + "val_auc": 0.6626650660264105, + "k": 2, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711 + ] + }, + { + "test_f1": 0.6299721857736773, + "test_acc": 0.630272952853598, + "test_auc": 0.6492783606718882, + "val_auc": 0.7380952380952381, + "k": 5, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 1625, + 583, + 213 + ] + }, + { + "test_f1": 0.6200385531769975, + "test_acc": 0.6203473945409429, + "test_auc": 0.6500911285158367, + "val_auc": 0.7174869947979191, + "k": 10, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 1625, + 583, + 213, + 434, + 1619, + 4802, + 10, + 1799 + ] + }, + { + "test_f1": 0.620277250008516, + "test_acc": 0.6203473945409429, + "test_auc": 0.6507314910595536, + "val_auc": 0.7306922769107643, + "k": 16, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 1625, + 583, + 213, + 434, + 1619, + 4802, + 10, + 1799, + 1551, + 680, + 1653, + 433, + 765, + 1381 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ab070cd65c503fcee7c2c685a763c9f648fbcbd0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.547433974162972, + "test_acc": 0.5483870967741935, + "test_auc": 0.5764740653169794, + "val_auc": 0.5520208083233293, + "k": 1, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619 + ] + }, + { + "test_f1": 0.5849946055450892, + "test_acc": 0.5856079404466501, + "test_auc": 0.6189842864883502, + "val_auc": 0.6518607442977191, + "k": 2, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042 + ] + }, + { + "test_f1": 0.6564978745399129, + "test_acc": 0.6575682382133995, + "test_auc": 0.7103344662824491, + "val_auc": 0.7394957983193277, + "k": 5, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 2480, + 1067 + ] + }, + { + "test_f1": 0.6746735324587365, + "test_acc": 0.674937965260546, + "test_auc": 0.735579528102064, + "val_auc": 0.7559023609443777, + "k": 10, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 2480, + 1067, + 1662, + 10, + 772, + 1625, + 1872 + ] + }, + { + "test_f1": 0.6911665696712426, + "test_acc": 0.6923076923076923, + "test_auc": 0.7455544061868874, + "val_auc": 0.7635054021608643, + "k": 16, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 2480, + 1067, + 1662, + 10, + 772, + 1625, + 1872, + 1653, + 2536, + 517, + 125, + 1094, + 898 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fd02afc45c057883edc3495b273dfee1dcfd5ca2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8346139578393897, + "test_acc": 0.8366800535475234, + "test_auc": 0.8564823443391494, + "val_auc": 0.8412121212121212, + "k": 1, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731 + ] + }, + { + "test_f1": 0.8287882835798244, + "test_acc": 0.8299866131191432, + "test_auc": 0.8979154420725151, + "val_auc": 0.9042424242424243, + "k": 2, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226 + ] + }, + { + "test_f1": 0.8602845521247463, + "test_acc": 0.8621151271753681, + "test_auc": 0.95505440782211, + "val_auc": 0.9551515151515152, + "k": 5, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 1158, + 737 + ] + }, + { + "test_f1": 0.8537527469486939, + "test_acc": 0.8540829986613119, + "test_auc": 0.9534988745681066, + "val_auc": 0.9612121212121212, + "k": 10, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 1158, + 737, + 6567, + 805, + 5872, + 1499, + 1274 + ] + }, + { + "test_f1": 0.8565366870554146, + "test_acc": 0.856760374832664, + "test_auc": 0.9545812963255007, + "val_auc": 0.9632323232323232, + "k": 16, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 23226, + 4648, + 1158, + 737, + 6567, + 805, + 5872, + 1499, + 1274, + 1629, + 13701, + 23697, + 1708, + 292, + 2046 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b6377bb8768edee71db247c7c3b00ab5cb310ac9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8733047151706635, + "test_acc": 0.8741633199464525, + "test_auc": 0.8908581955814252, + "val_auc": 0.8888888888888888, + "k": 1, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226 + ] + }, + { + "test_f1": 0.8773207386280631, + "test_acc": 0.8781793842034806, + "test_auc": 0.9243308339665381, + "val_auc": 0.9268686868686868, + "k": 2, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731 + ] + }, + { + "test_f1": 0.8828641471321304, + "test_acc": 0.8835341365461847, + "test_auc": 0.9683839658212785, + "val_auc": 0.9727272727272727, + "k": 5, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4648, + 4505, + 32 + ] + }, + { + "test_f1": 0.8842793556550892, + "test_acc": 0.8848728246318608, + "test_auc": 0.9696025863428481, + "val_auc": 0.9701010101010101, + "k": 10, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4648, + 4505, + 32, + 1065, + 6567, + 1682, + 1425, + 805 + ] + }, + { + "test_f1": 0.8842793556550892, + "test_acc": 0.8848728246318608, + "test_auc": 0.9696025863428481, + "val_auc": 0.9701010101010101, + "k": 16, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4731, + 4648, + 4505, + 32, + 1065, + 6567, + 1682, + 1425, + 805, + 22848, + 2046, + 176, + 22775, + 3255, + 23 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3a9d10f807fde94f16937c2690827c09e3923737 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8594893648948498, + "test_acc": 0.8607764390896921, + "test_auc": 0.8764784734269042, + "val_auc": 0.8616161616161615, + "k": 1, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648 + ] + }, + { + "test_f1": 0.8647498114343598, + "test_acc": 0.8661311914323963, + "test_auc": 0.9136786569368182, + "val_auc": 0.8820202020202019, + "k": 2, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505 + ] + }, + { + "test_f1": 0.8952104715156923, + "test_acc": 0.8955823293172691, + "test_auc": 0.9753050135481928, + "val_auc": 0.9658585858585859, + "k": 5, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 4731, + 23226, + 32 + ] + }, + { + "test_f1": 0.8873369798954491, + "test_acc": 0.8875502008032129, + "test_auc": 0.9733767257817091, + "val_auc": 0.9656565656565655, + "k": 10, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 4731, + 23226, + 32, + 1682, + 1158, + 1305, + 2832, + 1065 + ] + }, + { + "test_f1": 0.8887153740710226, + "test_acc": 0.8888888888888888, + "test_auc": 0.9755344009404884, + "val_auc": 0.9638383838383838, + "k": 16, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 4731, + 23226, + 32, + 1682, + 1158, + 1305, + 2832, + 1065, + 1101, + 2302, + 152, + 22848, + 923, + 1425 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e3e6bb3533fb084c514b75e39c8d5fadfbeacff0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7390011712640726, + "test_acc": 0.7538461538461538, + "test_auc": 0.753858518194365, + "val_auc": 0.7072623419012648, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518 + ] + }, + { + "test_f1": 0.767917002981525, + "test_acc": 0.7764102564102564, + "test_auc": 0.7820379035244217, + "val_auc": 0.7466340269277845, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541 + ] + }, + { + "test_f1": 0.7805622745331364, + "test_acc": 0.7846153846153846, + "test_auc": 0.8763338607062309, + "val_auc": 0.8419012647898817, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 1791, + 186, + 1335 + ] + }, + { + "test_f1": 0.7826521576328207, + "test_acc": 0.7866666666666666, + "test_auc": 0.8799188743393812, + "val_auc": 0.8523051815585475, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 1791, + 186, + 1335, + 821, + 2581, + 7813, + 1030, + 253 + ] + }, + { + "test_f1": 0.8063702411822147, + "test_acc": 0.8092307692307692, + "test_auc": 0.897513633150436, + "val_auc": 0.890656874745002, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 1791, + 186, + 1335, + 821, + 2581, + 7813, + 1030, + 253, + 505, + 1361, + 4895, + 1438, + 302, + 1417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6ccd34ca95ea83854baae93f1e2b3f4b32946236 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8562335864344115, + "test_acc": 0.8584873949579832, + "test_auc": 0.8619539142098906, + "val_auc": 0.8679245283018868, + "k": 1, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852 + ] + }, + { + "test_f1": 0.8972761380959635, + "test_acc": 0.8978151260504201, + "test_auc": 0.9073859651025735, + "val_auc": 0.9339622641509434, + "k": 2, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782 + ] + }, + { + "test_f1": 0.9068879304901383, + "test_acc": 0.9072268907563025, + "test_auc": 0.9621224446999443, + "val_auc": 0.9815335206744279, + "k": 5, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 1275 + ] + }, + { + "test_f1": 0.9153185378514811, + "test_acc": 0.9156302521008404, + "test_auc": 0.9704316893362547, + "val_auc": 0.9755118426334806, + "k": 10, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 1275, + 173, + 8815, + 11077, + 6919, + 259 + ] + }, + { + "test_f1": 0.9168291357576848, + "test_acc": 0.9169747899159664, + "test_auc": 0.9789836287249352, + "val_auc": 0.9853472501003613, + "k": 16, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 1275, + 173, + 8815, + 11077, + 6919, + 259, + 23246, + 1688, + 1283, + 2883, + 1152, + 98 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bd5fbcffd756a21c6286a67820fa2bb3bc75742d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7558319277234338, + "test_acc": 0.7620168067226891, + "test_auc": 0.7793579300171379, + "val_auc": 0.7962665596146126, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021 + ] + }, + { + "test_f1": 0.7691958245890731, + "test_acc": 0.7697478991596639, + "test_auc": 0.8803727285217403, + "val_auc": 0.8904054596547571, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914 + ] + }, + { + "test_f1": 0.7806980148013093, + "test_acc": 0.7818487394957984, + "test_auc": 0.8938603650996811, + "val_auc": 0.923323966278603, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 1180, + 1275 + ] + }, + { + "test_f1": 0.8140814741082297, + "test_acc": 0.8141176470588235, + "test_auc": 0.9151874489301545, + "val_auc": 0.9510236852669611, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 1180, + 1275, + 6919, + 1091, + 4790, + 98, + 1067 + ] + }, + { + "test_f1": 0.8278810738041636, + "test_acc": 0.8278991596638655, + "test_auc": 0.9224881771048007, + "val_auc": 0.951023685266961, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1493, + 1180, + 1275, + 6919, + 1091, + 4790, + 98, + 1067, + 1500, + 1794, + 27957, + 1152, + 347, + 964 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2fe1ef6232354381d66980347651192d8f8cfa01 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7548770083315461, + "test_acc": 0.7633613445378151, + "test_auc": 0.7671450058210586, + "val_auc": 0.7342432757928543, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235 + ] + }, + { + "test_f1": 0.7369760108667898, + "test_acc": 0.7438655462184874, + "test_auc": 0.7817473660614211, + "val_auc": 0.7442794058610999, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 5189 + ] + }, + { + "test_f1": 0.8077750881318762, + "test_acc": 0.8107563025210084, + "test_auc": 0.8240930808946353, + "val_auc": 0.7888398233641108, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 5189, + 692, + 764, + 1521 + ] + }, + { + "test_f1": 0.8212453654010239, + "test_acc": 0.8228571428571428, + "test_auc": 0.9103807821911765, + "val_auc": 0.9130871136089924, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 5189, + 692, + 764, + 1521, + 576, + 42, + 17338, + 3494, + 30916 + ] + }, + { + "test_f1": 0.8311470324279631, + "test_acc": 0.8322689075630252, + "test_auc": 0.930256442935549, + "val_auc": 0.9245283018867925, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 5189, + 692, + 764, + 1521, + 576, + 42, + 17338, + 3494, + 30916, + 964, + 8189, + 2046, + 32655, + 1980, + 24178 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3dff431bfbe0e6f2bf6f29b487ebfb8c02bc8d2f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9458822795663652, + "test_acc": 0.9458823529411765, + "test_auc": 0.9865740539876059, + "val_auc": 0.9851465274989964, + "k": 1, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9865542572447943, + "test_acc": 0.9865546218487395, + "test_auc": 0.9993632087409883, + "val_auc": 0.9995985547972703, + "k": 2, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847 + ] + }, + { + "test_f1": 0.9966386440685691, + "test_acc": 0.9966386554621849, + "test_auc": 0.999867128012669, + "val_auc": 1.0, + "k": 5, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 461, + 279 + ] + }, + { + "test_f1": 0.996302490929125, + "test_acc": 0.9963025210084033, + "test_auc": 0.999893340853707, + "val_auc": 1.0, + "k": 10, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 461, + 279, + 786, + 1411, + 1697, + 431, + 4148 + ] + }, + { + "test_f1": 0.9966386341940444, + "test_acc": 0.9966386554621849, + "test_auc": 0.9999597768473726, + "val_auc": 1.0, + "k": 16, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 461, + 279, + 786, + 1411, + 1697, + 431, + 4148, + 27673, + 1758, + 126, + 1360, + 1634, + 5020 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2d0d2fdbec1ee092539e3bbec5ed2bbc700bca0f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9442011132666112, + "test_acc": 0.944201680672269, + "test_auc": 0.9877986004150667, + "val_auc": 0.9879566439181052, + "k": 1, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9626821607536452, + "test_acc": 0.9626890756302521, + "test_auc": 0.995617258172983, + "val_auc": 0.9967884383781613, + "k": 2, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9673814693298005, + "test_acc": 0.9673949579831933, + "test_auc": 0.9968933263914499, + "val_auc": 0.9971898835808912, + "k": 5, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 431 + ] + }, + { + "test_f1": 0.9915962398620117, + "test_acc": 0.9915966386554622, + "test_auc": 0.9994427511551727, + "val_auc": 1.0, + "k": 10, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 431, + 1758, + 1847, + 279, + 461, + 126 + ] + }, + { + "test_f1": 0.999327731092437, + "test_acc": 0.999327731092437, + "test_auc": 0.9999977402723242, + "val_auc": 1.0, + "k": 16, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 431, + 1758, + 1847, + 279, + 461, + 126, + 1628, + 1551, + 1360, + 1472, + 1914, + 1634 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ac2aa5d21064810739d0952596745dc6ef75d45a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9297474546547736, + "test_acc": 0.9297478991596638, + "test_auc": 0.9793854083056741, + "val_auc": 0.9775190686471297, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9670564041209704, + "test_acc": 0.9670588235294117, + "test_auc": 0.995945596604262, + "val_auc": 0.9963869931754316, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9952940474616242, + "test_acc": 0.9952941176470588, + "test_auc": 0.9996931289816402, + "val_auc": 1.0, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 461 + ] + }, + { + "test_f1": 0.9959663263956778, + "test_acc": 0.9959663865546219, + "test_auc": 0.9998237412412955, + "val_auc": 1.0, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 461, + 279, + 1411, + 4148, + 1758, + 431 + ] + }, + { + "test_f1": 0.9979831928215676, + "test_acc": 0.9979831932773109, + "test_auc": 0.999980566341989, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1847, + 461, + 279, + 1411, + 4148, + 1758, + 431, + 1360, + 27673, + 1697, + 1472, + 1634, + 1914 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e9534d41b221cee4d08d05c6d865bb56ffd8f4d7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7642938817245678, + "test_acc": 0.7652830188679245, + "test_auc": 0.8275321278450776, + "val_auc": 0.8507402961184475, + "k": 1, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.8107168382924519, + "test_acc": 0.8110691823899371, + "test_auc": 0.8810095094978525, + "val_auc": 0.9335734293717487, + "k": 2, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585 + ] + }, + { + "test_f1": 0.8551623771245024, + "test_acc": 0.8553459119496856, + "test_auc": 0.9230255716483095, + "val_auc": 0.9603841536614646, + "k": 5, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 20310, + 776, + 508 + ] + }, + { + "test_f1": 0.8718299395725779, + "test_acc": 0.8719496855345912, + "test_auc": 0.9376648669065221, + "val_auc": 0.9687875150060025, + "k": 10, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 20310, + 776, + 508, + 1866, + 32700, + 27673, + 1327, + 16910 + ] + }, + { + "test_f1": 0.8718299395725779, + "test_acc": 0.8719496855345912, + "test_auc": 0.9376643605974041, + "val_auc": 0.9687875150060025, + "k": 16, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 20310, + 776, + 508, + 1866, + 32700, + 27673, + 1327, + 16910, + 400, + 5888, + 431, + 1499, + 7976, + 3832 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bbef515effe293494341e623d1e62fda1b46263c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6524902483693862, + "test_acc": 0.652931854199683, + "test_auc": 0.6635372714486638, + "val_auc": 0.7278911564625851, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1327 + ] + }, + { + "test_f1": 0.6877043682388629, + "test_acc": 0.6909667194928685, + "test_auc": 0.7597347799879445, + "val_auc": 0.7995198079231693, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1327, + 508 + ] + }, + { + "test_f1": 0.7701736972441955, + "test_acc": 0.7702060221870047, + "test_auc": 0.8355937311633513, + "val_auc": 0.9119647859143658, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1327, + 508, + 776, + 1042, + 21219 + ] + }, + { + "test_f1": 0.7954392907815062, + "test_acc": 0.7955625990491284, + "test_auc": 0.8851617440225035, + "val_auc": 0.9371748699479793, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1327, + 508, + 776, + 1042, + 21219, + 913, + 431, + 28521, + 1288, + 5186 + ] + }, + { + "test_f1": 0.8128674022822016, + "test_acc": 0.8129952456418383, + "test_auc": 0.8931685754470564, + "val_auc": 0.9311724689875951, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1327, + 508, + 776, + 1042, + 21219, + 913, + 431, + 28521, + 1288, + 5186, + 435, + 19666, + 1681, + 1866, + 20310, + 2023 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6c2045762b3b9f91e855cdc62edfe11cef8b2226 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6978555741549517, + "test_acc": 0.6986794717887155, + "test_auc": 0.7052867321527394, + "val_auc": 0.7202093397745573, + "k": 1, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.7262905162064826, + "test_acc": 0.7262905162064826, + "test_auc": 0.7835327891532927, + "val_auc": 0.7729468599033816, + "k": 2, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501 + ] + }, + { + "test_f1": 0.7262881494328289, + "test_acc": 0.7262905162064826, + "test_auc": 0.8038040721269138, + "val_auc": 0.7854267310789049, + "k": 5, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501, + 347, + 1447, + 21219 + ] + }, + { + "test_f1": 0.737094837935174, + "test_acc": 0.737094837935174, + "test_auc": 0.8347744881018262, + "val_auc": 0.8494363929146538, + "k": 10, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501, + 347, + 1447, + 21219, + 1528, + 5186, + 1417, + 2837, + 776 + ] + }, + { + "test_f1": 0.7681930742350217, + "test_acc": 0.7683073229291717, + "test_auc": 0.8457099704851504, + "val_auc": 0.8152173913043479, + "k": 16, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501, + 347, + 1447, + 21219, + 1528, + 5186, + 1417, + 2837, + 776, + 1866, + 559, + 5020, + 32700, + 1804, + 777 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d8f1c6128f76428e296b2ede31c18fffcad7fc60 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.737814416276898, + "test_acc": 0.7382920110192838, + "test_auc": 0.7593042316799223, + "val_auc": 0.7330932372949179, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7327580310924489, + "test_acc": 0.7327823691460055, + "test_auc": 0.7726306842328943, + "val_auc": 0.78031212484994, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186 + ] + }, + { + "test_f1": 0.7548209366391184, + "test_acc": 0.7548209366391184, + "test_auc": 0.8022281585817498, + "val_auc": 0.7979191676670668, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 6501, + 3507, + 1681 + ] + }, + { + "test_f1": 0.7353564473202626, + "test_acc": 0.7355371900826446, + "test_auc": 0.8025013660372776, + "val_auc": 0.8079231692677071, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 6501, + 3507, + 1681, + 12612, + 1377, + 2247, + 508, + 3301 + ] + }, + { + "test_f1": 0.7487928112104222, + "test_acc": 0.7493112947658402, + "test_auc": 0.835043409629045, + "val_auc": 0.7955182072829132, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 6501, + 3507, + 1681, + 12612, + 1377, + 2247, + 508, + 3301, + 601, + 1585, + 4336, + 1447, + 1417, + 1688 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9605f93676854c8b8da8083d5257c89971317b61 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6349640721496733, + "test_acc": 0.6382389937106918, + "test_auc": 0.6596239490288485, + "val_auc": 0.6514605842336935, + "k": 1, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27673 + ] + }, + { + "test_f1": 0.6619592520156323, + "test_acc": 0.6621383647798742, + "test_auc": 0.7099403162811798, + "val_auc": 0.7198879551820727, + "k": 2, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27673, + 508 + ] + }, + { + "test_f1": 0.6815001212453041, + "test_acc": 0.6815094339622642, + "test_auc": 0.7466769666818222, + "val_auc": 0.7791116446578631, + "k": 5, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27673, + 508, + 4605, + 1585, + 776 + ] + }, + { + "test_f1": 0.6845109307393541, + "test_acc": 0.6845283018867925, + "test_auc": 0.7503156837350222, + "val_auc": 0.7709083633453381, + "k": 10, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27673, + 508, + 4605, + 1585, + 776, + 1804, + 1866, + 126, + 2837, + 30052 + ] + }, + { + "test_f1": 0.6842593444948006, + "test_acc": 0.6842767295597484, + "test_auc": 0.7484466436262265, + "val_auc": 0.7699079631852741, + "k": 16, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 27673, + 508, + 4605, + 1585, + 776, + 1804, + 1866, + 126, + 2837, + 30052, + 3832, + 20310, + 1411, + 559, + 322, + 1042 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c9182cf016bc7e9fdbbafacc792f84608c89cf0e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7993015310872965, + "test_acc": 0.8, + "test_auc": 0.8439623650306468, + "val_auc": 0.8283313325330133, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293 + ] + }, + { + "test_f1": 0.8052800612165061, + "test_acc": 0.8052830188679245, + "test_auc": 0.8884459246672789, + "val_auc": 0.8865546218487395, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062 + ] + }, + { + "test_f1": 0.8296067034198026, + "test_acc": 0.829685534591195, + "test_auc": 0.9112808456172363, + "val_auc": 0.9275710284113645, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1441 + ] + }, + { + "test_f1": 0.8343595282174633, + "test_acc": 0.8344654088050314, + "test_auc": 0.918990794287618, + "val_auc": 0.945578231292517, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1441, + 1166, + 409, + 1305, + 1116, + 1313 + ] + }, + { + "test_f1": 0.8379273950551852, + "test_acc": 0.8379874213836478, + "test_auc": 0.9217896710914708, + "val_auc": 0.9555822328931572, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1441, + 1166, + 409, + 1305, + 1116, + 1313, + 1928, + 524, + 14876, + 765, + 14, + 1737 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..563e148d0906260ebb07277cedd320fc34b4fc8d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6272499430394167, + "test_acc": 0.6363636363636364, + "test_auc": 0.6579591836734693, + "val_auc": 0.6158169409875552, + "k": 1, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377 + ] + }, + { + "test_f1": 0.6363900926499145, + "test_acc": 0.6464646464646465, + "test_auc": 0.6991836734693877, + "val_auc": 0.6557607386591731, + "k": 2, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027 + ] + }, + { + "test_f1": 0.6666666666666666, + "test_acc": 0.6666666666666666, + "test_auc": 0.7281632653061225, + "val_auc": 0.6842633480529907, + "k": 5, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339 + ] + }, + { + "test_f1": 0.5957120221826104, + "test_acc": 0.5959595959595959, + "test_auc": 0.72, + "val_auc": 0.7547169811320754, + "k": 10, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339, + 681, + 100, + 1551, + 524, + 6086 + ] + }, + { + "test_f1": 0.6361408199643493, + "test_acc": 0.6363636363636364, + "test_auc": 0.7310204081632653, + "val_auc": 0.788438378161381, + "k": 16, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339, + 681, + 100, + 1551, + 524, + 6086, + 1790, + 957, + 505, + 272, + 56, + 1202 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..aa02f494465ef34fe52a0e71a0d6c9eca5d49a6b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6688259072771345, + "test_acc": 0.6722268557130943, + "test_auc": 0.6931636060100166, + "val_auc": 0.7910628019323671, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.675658192440463, + "test_acc": 0.6805671392827356, + "test_auc": 0.7203311074012243, + "val_auc": 0.8627214170692431, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114 + ] + }, + { + "test_f1": 0.7374589495056673, + "test_acc": 0.7406171809841534, + "test_auc": 0.800681691708403, + "val_auc": 0.8719806763285024, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 4481 + ] + }, + { + "test_f1": 0.7706008121287446, + "test_acc": 0.7723102585487907, + "test_auc": 0.8433889816360602, + "val_auc": 0.8892914653784219, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 4481, + 238, + 3276, + 6086, + 30704, + 801 + ] + }, + { + "test_f1": 0.7958306870979438, + "test_acc": 0.7964970809007507, + "test_auc": 0.8793711741791875, + "val_auc": 0.9166666666666666, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 4481, + 238, + 3276, + 6086, + 30704, + 801, + 6719, + 6400, + 4534, + 363, + 1335, + 1670 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fde7eac72898c83226ea73cc01b9d9acf0b34184 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.5735670219593688, + "test_acc": 0.5934591194968554, + "test_auc": 0.6344149446249718, + "val_auc": 0.5450180072028812, + "k": 2, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727 + ] + }, + { + "test_f1": 0.608894639024476, + "test_acc": 0.6103144654088051, + "test_auc": 0.6591380188529263, + "val_auc": 0.5602240896358542, + "k": 5, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1928 + ] + }, + { + "test_f1": 0.6229960898641206, + "test_acc": 0.6233962264150943, + "test_auc": 0.6867916102553924, + "val_auc": 0.5820328131252501, + "k": 10, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1928, + 6400, + 486, + 293, + 2138, + 4913 + ] + }, + { + "test_f1": 0.6275812686231992, + "test_acc": 0.6284276729559748, + "test_auc": 0.6932349000900218, + "val_auc": 0.5776310524209683, + "k": 16, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 3727, + 1960, + 588, + 1928, + 6400, + 486, + 293, + 2138, + 4913, + 1739, + 765, + 2034, + 7143, + 30704, + 1438 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0ffc325ea115af71556e520309303bff7ea00892 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.521232068425025, + "test_acc": 0.5362058647516457, + "test_auc": 0.546713749534424, + "val_auc": 0.5501806503412284, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498 + ] + }, + { + "test_f1": 0.5522243385225211, + "test_acc": 0.5523638539796529, + "test_auc": 0.5841890381915594, + "val_auc": 0.5194700923323967, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727 + ] + }, + { + "test_f1": 0.5494550187686323, + "test_acc": 0.5499700777977259, + "test_auc": 0.6015471449445606, + "val_auc": 0.5620232838217584, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 844, + 293, + 860 + ] + }, + { + "test_f1": 0.5644847745301597, + "test_acc": 0.5649311789347696, + "test_auc": 0.6031652006990803, + "val_auc": 0.576876756322762, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 844, + 293, + 860, + 1336, + 2797, + 150, + 100, + 27404 + ] + }, + { + "test_f1": 0.561930172986926, + "test_acc": 0.5619389587073609, + "test_auc": 0.6004727387330602, + "val_auc": 0.5535929345644319, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 844, + 293, + 860, + 1336, + 2797, + 150, + 100, + 27404, + 19092, + 5060, + 5704, + 378, + 7400, + 22608 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..27118ef477cce5951c87264ab8aa32c340a328bd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5480897703390899, + "test_acc": 0.5630188679245283, + "test_auc": 0.5810157877309148, + "val_auc": 0.6510604241696678, + "k": 1, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.5892994932888765, + "test_acc": 0.6062893081761006, + "test_auc": 0.6114654712370855, + "val_auc": 0.7006802721088435, + "k": 2, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 6659 + ] + }, + { + "test_f1": 0.6042480143977995, + "test_acc": 0.6203773584905661, + "test_auc": 0.6193981959193511, + "val_auc": 0.7098839535814326, + "k": 5, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 6659, + 1345, + 13203, + 6400 + ] + }, + { + "test_f1": 0.6129529093827806, + "test_acc": 0.6238993710691824, + "test_auc": 0.6966618533546524, + "val_auc": 0.8023209283713485, + "k": 10, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 6659, + 1345, + 13203, + 6400, + 1377, + 777, + 7143, + 1596, + 1327 + ] + }, + { + "test_f1": 0.6131935600496689, + "test_acc": 0.6236477987421384, + "test_auc": 0.705478214024965, + "val_auc": 0.8211284513805522, + "k": 16, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 6659, + 1345, + 13203, + 6400, + 1377, + 777, + 7143, + 1596, + 1327, + 27965, + 13614, + 1550, + 448, + 26139, + 121 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4f5691e3b9135ba6ca3fb2a2d32c93b549c93eb3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9952200531821748, + "test_acc": 0.9952201257861635, + "test_auc": 0.9994704006626574, + "val_auc": 1.0, + "k": 1, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.99345891002436, + "test_acc": 0.9934591194968554, + "test_auc": 0.9994817926178106, + "val_auc": 1.0, + "k": 2, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1500 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.9999521537883568, + "val_auc": 1.0, + "k": 5, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1500, + 1625, + 1703, + 1525 + ] + }, + { + "test_f1": 0.9977358447577434, + "test_acc": 0.9977358490566037, + "test_auc": 0.9999756971623399, + "val_auc": 1.0, + "k": 10, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1500, + 1625, + 1703, + 1525, + 347, + 1551, + 6718, + 1232, + 1255 + ] + }, + { + "test_f1": 0.9992452824456872, + "test_acc": 0.999245283018868, + "test_auc": 0.9999913927449954, + "val_auc": 1.0, + "k": 16, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1500, + 1625, + 1703, + 1525, + 347, + 1551, + 6718, + 1232, + 1255, + 1335, + 15467, + 1042, + 11655, + 549, + 1758 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..110f4599bab48405d37e0e9964350517b5dc9c81 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6495593048495152, + "test_acc": 0.649559748427673, + "test_auc": 0.6921310196356802, + "val_auc": 0.7639055622248899, + "k": 1, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6500240833725932, + "test_acc": 0.650314465408805, + "test_auc": 0.7027595872163024, + "val_auc": 0.7523009203681472, + "k": 2, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928 + ] + }, + { + "test_f1": 0.6792103222946334, + "test_acc": 0.6792452830188679, + "test_auc": 0.7514766505424091, + "val_auc": 0.8379351740696278, + "k": 5, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 6400 + ] + }, + { + "test_f1": 0.6827473776993825, + "test_acc": 0.6827672955974843, + "test_auc": 0.7510997034041187, + "val_auc": 0.8251300520208084, + "k": 10, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 6400, + 56, + 5827, + 152, + 1255, + 30704 + ] + }, + { + "test_f1": 0.6986149788547047, + "test_acc": 0.6986163522012578, + "test_auc": 0.7758749274712188, + "val_auc": 0.8443377350940376, + "k": 16, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 6400, + 56, + 5827, + 152, + 1255, + 30704, + 406, + 1327, + 1736, + 1396, + 1691, + 2046 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..291f3beb5d10478f8d21cd06fad7840f15abae11 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6744673306306844, + "test_acc": 0.6759748427672956, + "test_auc": 0.7210327895910946, + "val_auc": 0.7322929171668668, + "k": 1, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.669187173477787, + "test_acc": 0.6714465408805032, + "test_auc": 0.7210327895910946, + "val_auc": 0.7322929171668668, + "k": 2, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1360 + ] + }, + { + "test_f1": 0.669187173477787, + "test_acc": 0.6714465408805032, + "test_auc": 0.7210327895910946, + "val_auc": 0.7322929171668668, + "k": 5, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1360, + 7143, + 30704, + 6400 + ] + }, + { + "test_f1": 0.7127536364242256, + "test_acc": 0.7129559748427673, + "test_auc": 0.785358856713507, + "val_auc": 0.7783113245298119, + "k": 10, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1360, + 7143, + 30704, + 6400, + 1928, + 1011, + 19940, + 598, + 13614 + ] + }, + { + "test_f1": 0.723589795686132, + "test_acc": 0.7237735849056603, + "test_auc": 0.7983021430039725, + "val_auc": 0.8143257302921169, + "k": 16, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1360, + 7143, + 30704, + 6400, + 1928, + 1011, + 19940, + 598, + 13614, + 253, + 1960, + 836, + 5827, + 1019, + 7068 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0ad3f7df5f671d1706285d0d3bafd2f369ba5e72 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7756216428168373, + "test_acc": 0.7758212877792379, + "test_auc": 0.8050448044567408, + "val_auc": 0.7447999999999999, + "k": 1, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062 + ] + }, + { + "test_f1": 0.7941426495740193, + "test_acc": 0.7952693823915901, + "test_auc": 0.8849110900769331, + "val_auc": 0.8246000000000001, + "k": 2, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448 + ] + }, + { + "test_f1": 0.8271104266636357, + "test_acc": 0.8275952693823916, + "test_auc": 0.9053578029709026, + "val_auc": 0.8324, + "k": 5, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 1662 + ] + }, + { + "test_f1": 0.8444138245817677, + "test_acc": 0.8444152431011827, + "test_auc": 0.9189370041105057, + "val_auc": 0.8508, + "k": 10, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 1662, + 30127, + 296, + 6778, + 1030, + 140 + ] + }, + { + "test_f1": 0.8520358738998899, + "test_acc": 0.8520367936925098, + "test_auc": 0.927744559616699, + "val_auc": 0.8924, + "k": 16, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 1662, + 30127, + 296, + 6778, + 1030, + 140, + 628, + 2073, + 1313, + 2046, + 1942, + 2831 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..29c9eed3f1bd37d8cc4eec4a10aa0d515c9be842 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5184027317877234, + "test_acc": 0.5240251572327044, + "test_auc": 0.5329723686861987, + "val_auc": 0.6016406562625051, + "k": 1, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011 + ] + }, + { + "test_f1": 0.5202054853352679, + "test_acc": 0.5368553459119497, + "test_auc": 0.5808431363217047, + "val_auc": 0.5864345738295318, + "k": 2, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1377 + ] + }, + { + "test_f1": 0.5205944992913223, + "test_acc": 0.521006289308176, + "test_auc": 0.5383000063794949, + "val_auc": 0.645858343337335, + "k": 5, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1377, + 1411, + 800, + 1928 + ] + }, + { + "test_f1": 0.5213283979342271, + "test_acc": 0.5217610062893082, + "test_auc": 0.5383000063794949, + "val_auc": 0.645858343337335, + "k": 10, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1377, + 1411, + 800, + 1928, + 1703, + 786, + 18461, + 6718, + 334 + ] + }, + { + "test_f1": 0.5477570340978981, + "test_acc": 0.5479245283018868, + "test_auc": 0.5943129840948054, + "val_auc": 0.6478591436574629, + "k": 16, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1377, + 1411, + 800, + 1928, + 1703, + 786, + 18461, + 6718, + 334, + 1596, + 15467, + 1417, + 1791, + 1643, + 448 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c26335db88d033672546419cf789a8079f73929a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9934589779181365, + "test_acc": 0.9934591194968554, + "test_auc": 0.9968834142246534, + "val_auc": 1.0, + "k": 1, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.993710570020631, + "test_acc": 0.9937106918238994, + "test_auc": 0.9973114985838534, + "val_auc": 1.0, + "k": 2, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 292 + ] + }, + { + "test_f1": 0.9919496060527676, + "test_acc": 0.9919496855345912, + "test_auc": 0.9990613028953793, + "val_auc": 1.0, + "k": 5, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 292, + 26988, + 487, + 6935 + ] + }, + { + "test_f1": 0.9927043184801099, + "test_acc": 0.9927044025157232, + "test_auc": 0.9993658478298072, + "val_auc": 1.0, + "k": 10, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 292, + 26988, + 487, + 6935, + 6475, + 1744, + 152, + 347, + 334 + ] + }, + { + "test_f1": 0.9944653352458714, + "test_acc": 0.9944654088050314, + "test_auc": 0.999360025274951, + "val_auc": 1.0, + "k": 16, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 292, + 26988, + 487, + 6935, + 6475, + 1744, + 152, + 347, + 334, + 821, + 559, + 3417, + 1960, + 771, + 7368 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bd9054eeffe28ce6eec2c2d530d5041a041b253a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8573833090422625, + "test_acc": 0.8576100628930817, + "test_auc": 0.9001219192355947, + "val_auc": 0.9231692677070829, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487 + ] + }, + { + "test_f1": 0.8943044633799548, + "test_acc": 0.8943396226415095, + "test_auc": 0.9377243582278775, + "val_auc": 0.9719887955182072, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744 + ] + }, + { + "test_f1": 0.9101431267652809, + "test_acc": 0.910188679245283, + "test_auc": 0.9666273939560869, + "val_auc": 0.982593037214886, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 6475, + 26988, + 24947 + ] + }, + { + "test_f1": 0.9330318671382632, + "test_acc": 0.9330817610062893, + "test_auc": 0.9800096502517875, + "val_auc": 0.9823929571828732, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 6475, + 26988, + 24947, + 771, + 7368, + 1524, + 1028, + 559 + ] + }, + { + "test_f1": 0.9471459274485569, + "test_acc": 0.9471698113207547, + "test_auc": 0.9868822902184117, + "val_auc": 0.981592637054822, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 6475, + 26988, + 24947, + 771, + 7368, + 1524, + 1028, + 559, + 152, + 878, + 3417, + 6797, + 1288, + 6935 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..faf96604fb4ad813e7d6d93f12834fad40704492 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8556524980858146, + "test_acc": 0.8572151898734177, + "test_auc": 0.8639038266697842, + "val_auc": 0.816126450580232, + "k": 1, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104 + ] + }, + { + "test_f1": 0.8555588418790218, + "test_acc": 0.8572151898734177, + "test_auc": 0.9161488008072554, + "val_auc": 0.9103641456582634, + "k": 2, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555 + ] + }, + { + "test_f1": 0.8605553023922425, + "test_acc": 0.8617721518987341, + "test_auc": 0.9334865395895631, + "val_auc": 0.9211684673869548, + "k": 5, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 292, + 337, + 1708 + ] + }, + { + "test_f1": 0.8616547979656971, + "test_acc": 0.8627848101265823, + "test_auc": 0.9365793780687398, + "val_auc": 0.9259703881552621, + "k": 10, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 292, + 337, + 1708, + 1815, + 10639, + 1745, + 677, + 1380 + ] + }, + { + "test_f1": 0.8759157745573835, + "test_acc": 0.8759493670886076, + "test_auc": 0.9460711927117302, + "val_auc": 0.9423769507803121, + "k": 16, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 292, + 337, + 1708, + 1815, + 10639, + 1745, + 677, + 1380, + 3825, + 1851, + 918, + 3502, + 3203, + 472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b01da39ff6eef168e1089c40a797c37b502d6406 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7428634207719175, + "test_acc": 0.7497435897435898, + "test_auc": 0.764647221193658, + "val_auc": 0.7645858833129335, + "k": 1, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468 + ] + }, + { + "test_f1": 0.7709182079801772, + "test_acc": 0.7743589743589744, + "test_auc": 0.8434754434981654, + "val_auc": 0.8176254589963282, + "k": 2, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804 + ] + }, + { + "test_f1": 0.8587487526527945, + "test_acc": 0.8594871794871795, + "test_auc": 0.9014752415255665, + "val_auc": 0.8943288453692371, + "k": 5, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804, + 1351, + 1132, + 1474 + ] + }, + { + "test_f1": 0.8578586974327772, + "test_acc": 0.8584615384615385, + "test_auc": 0.9141027367287171, + "val_auc": 0.8967768257853937, + "k": 10, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804, + 1351, + 1132, + 1474, + 343, + 6475, + 805, + 930, + 3417 + ] + }, + { + "test_f1": 0.8681935991345182, + "test_acc": 0.8687179487179487, + "test_auc": 0.9198926179015048, + "val_auc": 0.8951448388412893, + "k": 16, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1804, + 1351, + 1132, + 1474, + 343, + 6475, + 805, + 930, + 3417, + 27324, + 21810, + 1305, + 7595, + 1473, + 26988 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0f498af424c43f16332c114557a4d8ec9601cb77 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9100051632975643, + "test_acc": 0.910188679245283, + "test_auc": 0.9270536404131888, + "val_auc": 0.9811924769907963, + "k": 1, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.9074206781214923, + "test_acc": 0.9076729559748428, + "test_auc": 0.9515057126857774, + "val_auc": 0.987595038015206, + "k": 2, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 11381 + ] + }, + { + "test_f1": 0.9276672192936306, + "test_acc": 0.9277987421383648, + "test_auc": 0.9773522868463929, + "val_auc": 0.9947979191676671, + "k": 5, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 11381, + 1279, + 1447, + 1556 + ] + }, + { + "test_f1": 0.9343048136828592, + "test_acc": 0.9343396226415094, + "test_auc": 0.9807241789944497, + "val_auc": 0.9943977591036415, + "k": 10, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 11381, + 1279, + 1447, + 1556, + 1012, + 238, + 7740, + 26750, + 4877 + ] + }, + { + "test_f1": 0.9343032744469593, + "test_acc": 0.9343396226415094, + "test_auc": 0.9808739199160742, + "val_auc": 0.9943977591036415, + "k": 16, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 11381, + 1279, + 1447, + 1556, + 1012, + 238, + 7740, + 26750, + 4877, + 88, + 6828, + 8129, + 27068, + 3301, + 262 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8bd28e571dea152368a998356137657c86eccf2e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8404762170699583, + "test_acc": 0.8430769230769231, + "test_auc": 0.8449418487225233, + "val_auc": 0.8659730722154223, + "k": 1, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628 + ] + }, + { + "test_f1": 0.8990857243373654, + "test_acc": 0.8994871794871795, + "test_auc": 0.9148033291816744, + "val_auc": 0.9055487556099552, + "k": 2, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416 + ] + }, + { + "test_f1": 0.8893028015988149, + "test_acc": 0.8902564102564102, + "test_auc": 0.9158868280203318, + "val_auc": 0.9092207262341901, + "k": 5, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 238, + 737 + ] + }, + { + "test_f1": 0.8969709777402085, + "test_acc": 0.8974358974358975, + "test_auc": 0.930460413370586, + "val_auc": 0.9151366789065689, + "k": 10, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 238, + 737, + 9381, + 805, + 1132, + 930, + 4211 + ] + }, + { + "test_f1": 0.8990857243373654, + "test_acc": 0.8994871794871795, + "test_auc": 0.9392399097855724, + "val_auc": 0.9281925744594043, + "k": 16, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 238, + 737, + 9381, + 805, + 1132, + 930, + 4211, + 1524, + 334, + 2062, + 1274, + 52, + 1288 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..da6cfe2c2d617bcb876e8db7b115b91d6797299e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8707276195755217, + "test_acc": 0.8707692307692307, + "test_auc": 0.9138502709799037, + "val_auc": 0.9400244798041615, + "k": 1, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524 + ] + }, + { + "test_f1": 0.8686281241565452, + "test_acc": 0.8687179487179487, + "test_auc": 0.9238373952267143, + "val_auc": 0.9410444716442268, + "k": 2, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947 + ] + }, + { + "test_f1": 0.8727192936425106, + "test_acc": 0.8728205128205129, + "test_auc": 0.9265934796512606, + "val_auc": 0.9420644634842922, + "k": 5, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 888 + ] + }, + { + "test_f1": 0.8727192936425106, + "test_acc": 0.8728205128205129, + "test_auc": 0.9281692867001043, + "val_auc": 0.9408404732762139, + "k": 10, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 888, + 969, + 1474, + 3203, + 1682, + 805 + ] + }, + { + "test_f1": 0.8696613267801492, + "test_acc": 0.8697435897435898, + "test_auc": 0.93109999326758, + "val_auc": 0.9530803753569971, + "k": 16, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 24947, + 3816, + 930, + 888, + 969, + 1474, + 3203, + 1682, + 805, + 1214, + 6631, + 1260, + 776, + 3417, + 5173 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f5386ee12a4df2ad5e1e02f9376847c7495cbc86 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8094940138931124, + "test_acc": 0.8123076923076923, + "test_auc": 0.8260321641363988, + "val_auc": 0.8004895960832313, + "k": 1, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.7977313302567773, + "test_acc": 0.8020512820512821, + "test_auc": 0.8630289157437642, + "val_auc": 0.8155854753161975, + "k": 2, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815 + ] + }, + { + "test_f1": 0.8643454076215444, + "test_acc": 0.8646153846153846, + "test_auc": 0.9264209613895715, + "val_auc": 0.8700530395756834, + "k": 5, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 1524, + 487, + 334 + ] + }, + { + "test_f1": 0.8880956552200334, + "test_acc": 0.8882051282051282, + "test_auc": 0.9415520752684553, + "val_auc": 0.9031007751937985, + "k": 10, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 1524, + 487, + 334, + 6758, + 1866, + 6935, + 930, + 2046 + ] + }, + { + "test_f1": 0.885049608636977, + "test_acc": 0.8851282051282051, + "test_auc": 0.94653616992628, + "val_auc": 0.920032639738882, + "k": 16, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 1524, + 487, + 334, + 6758, + 1866, + 6935, + 930, + 2046, + 913, + 10639, + 1221, + 1158, + 508, + 24947 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..795ff0e54b06e9556780d249f155b91bbf4860e5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8538146334448481, + "test_acc": 0.8543589743589743, + "test_auc": 0.8811349176961659, + "val_auc": 0.8853529171766626, + "k": 1, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52 + ] + }, + { + "test_f1": 0.8548257153474546, + "test_acc": 0.8553846153846154, + "test_auc": 0.8811349176961659, + "val_auc": 0.8853529171766626, + "k": 2, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230 + ] + }, + { + "test_f1": 0.8537772740398561, + "test_acc": 0.8543589743589743, + "test_auc": 0.8924285521930858, + "val_auc": 0.8931048551611587, + "k": 5, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 18545, + 2884, + 1274 + ] + }, + { + "test_f1": 0.8589795740101306, + "test_acc": 0.8594871794871795, + "test_auc": 0.9002213283064597, + "val_auc": 0.9122807017543859, + "k": 10, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 18545, + 2884, + 1274, + 2131, + 17546, + 888, + 2046, + 117 + ] + }, + { + "test_f1": 0.8611950529896035, + "test_acc": 0.8615384615384616, + "test_auc": 0.91284040798465, + "val_auc": 0.9473684210526315, + "k": 16, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 18545, + 2884, + 1274, + 2131, + 17546, + 888, + 2046, + 117, + 1193, + 238, + 821, + 1682, + 2484, + 508 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b2b8f42781d35ae726c8a271af735c5588acb27c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8757764035412042, + "test_acc": 0.8759358288770054, + "test_auc": 0.9225598931166381, + "val_auc": 0.9216161616161617, + "k": 1, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839 + ] + }, + { + "test_f1": 0.8811216108192637, + "test_acc": 0.881283422459893, + "test_auc": 0.9378237156609748, + "val_auc": 0.934949494949495, + "k": 2, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682 + ] + }, + { + "test_f1": 0.8866138959908392, + "test_acc": 0.8866310160427807, + "test_auc": 0.9467459140906678, + "val_auc": 0.9426262626262627, + "k": 5, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 343 + ] + }, + { + "test_f1": 0.8940987499004838, + "test_acc": 0.8941176470588236, + "test_auc": 0.9522776771170776, + "val_auc": 0.9486868686868687, + "k": 10, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 343, + 3417, + 930, + 340, + 888, + 874 + ] + }, + { + "test_f1": 0.9047958458701318, + "test_acc": 0.9048128342245989, + "test_auc": 0.9573473160196928, + "val_auc": 0.9402020202020203, + "k": 16, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 343, + 3417, + 930, + 340, + 888, + 874, + 1895, + 238, + 1441, + 456, + 1744, + 2884 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dab104eea98559daf06d2a0fd4fd966e566d530c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9503404186303577, + "test_acc": 0.9503448275862069, + "test_auc": 0.987156874019157, + "val_auc": 0.9951298701298701, + "k": 1, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554 + ] + }, + { + "test_f1": 0.9632157314793794, + "test_acc": 0.9632183908045977, + "test_auc": 0.9916966556631852, + "val_auc": 0.9995941558441559, + "k": 2, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072 + ] + }, + { + "test_f1": 0.9719539518450812, + "test_acc": 0.9719540229885058, + "test_auc": 0.9951084677201147, + "val_auc": 1.0, + "k": 5, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 1101, + 760, + 192 + ] + }, + { + "test_f1": 0.9751723508137717, + "test_acc": 0.9751724137931035, + "test_auc": 0.9944751474646896, + "val_auc": 1.0, + "k": 10, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 1101, + 760, + 192, + 4323, + 1065, + 1622, + 1682, + 763 + ] + }, + { + "test_f1": 0.9779307732329513, + "test_acc": 0.9779310344827586, + "test_auc": 0.9952657408409544, + "val_auc": 1.0, + "k": 16, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 1101, + 760, + 192, + 4323, + 1065, + 1622, + 1682, + 763, + 1876, + 770, + 1684, + 1872, + 652, + 432 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..35bf9aa31690fd77eca9dc2481dbbc2c79dfc979 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9617669670241205, + "test_acc": 0.9617773530817009, + "test_auc": 0.9716863806450553, + "val_auc": 0.9841269841269842, + "k": 1, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286 + ] + }, + { + "test_f1": 0.961297714687038, + "test_acc": 0.9612995699952221, + "test_auc": 0.9879725556584322, + "val_auc": 0.9905619905619905, + "k": 2, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905 + ] + }, + { + "test_f1": 0.9737209944203152, + "test_acc": 0.9737219302436694, + "test_auc": 0.9966644204236452, + "val_auc": 0.9991419991419992, + "k": 5, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581 + ] + }, + { + "test_f1": 0.9813643070665401, + "test_acc": 0.9813664596273292, + "test_auc": 0.9977638011545324, + "val_auc": 0.9995709995709996, + "k": 10, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581, + 2874, + 1079, + 3877, + 1682, + 1684 + ] + }, + { + "test_f1": 0.9885322111539042, + "test_acc": 0.9885332059245103, + "test_auc": 0.9987572614827761, + "val_auc": 1.0, + "k": 16, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581, + 2874, + 1079, + 3877, + 1682, + 1684, + 760, + 1156, + 239, + 948, + 1596, + 88 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..981b0b40adfb13e03634fe33ea8382701b906fb2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9788488756993098, + "test_acc": 0.9788505747126437, + "test_auc": 0.9928195519238053, + "val_auc": 0.9837662337662337, + "k": 1, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1789 + ] + }, + { + "test_f1": 0.9898848650314882, + "test_acc": 0.9898850574712643, + "test_auc": 0.9990732723632232, + "val_auc": 1.0, + "k": 2, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1789, + 300 + ] + }, + { + "test_f1": 0.992183908045977, + "test_acc": 0.992183908045977, + "test_auc": 0.9996803804318415, + "val_auc": 1.0, + "k": 5, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1789, + 300, + 14559, + 2236, + 334 + ] + }, + { + "test_f1": 0.9908045977011494, + "test_acc": 0.9908045977011494, + "test_auc": 0.9996871448671465, + "val_auc": 1.0, + "k": 10, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1789, + 300, + 14559, + 2236, + 334, + 982, + 3825, + 1158, + 1934, + 1268 + ] + }, + { + "test_f1": 0.991724116937924, + "test_acc": 0.9917241379310345, + "test_auc": 0.9996727704421235, + "val_auc": 1.0, + "k": 16, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1789, + 300, + 14559, + 2236, + 334, + 982, + 3825, + 1158, + 1934, + 1268, + 1872, + 1592, + 896, + 913, + 724, + 1410 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..20dad14120359b8f05ed132156bcaa9b84695585 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9457464841976717, + "test_acc": 0.9457471264367816, + "test_auc": 0.9772123085664809, + "val_auc": 0.961038961038961, + "k": 1, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095 + ] + }, + { + "test_f1": 0.9576999691180144, + "test_acc": 0.9577011494252874, + "test_auc": 0.9864508360842037, + "val_auc": 0.9732142857142858, + "k": 2, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023 + ] + }, + { + "test_f1": 0.9618373863654072, + "test_acc": 0.9618390804597701, + "test_auc": 0.9879423940689432, + "val_auc": 0.9703733766233766, + "k": 5, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 494, + 1789 + ] + }, + { + "test_f1": 0.977928459044762, + "test_acc": 0.9779310344827586, + "test_auc": 0.9956453947724444, + "val_auc": 0.9866071428571428, + "k": 10, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 494, + 1789, + 1301, + 1682, + 1079, + 211, + 1701 + ] + }, + { + "test_f1": 0.9802281296550799, + "test_acc": 0.9802298850574712, + "test_auc": 0.9961028397099411, + "val_auc": 0.9849837662337663, + "k": 16, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 494, + 1789, + 1301, + 1682, + 1079, + 211, + 1701, + 1193, + 235, + 2011, + 902, + 586, + 292 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..62b3d6670ad389dd4601b9f3e5c6911550b4f72c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9719510228106321, + "test_acc": 0.9719540229885058, + "test_auc": 0.9949993912008225, + "val_auc": 0.9874188311688311, + "k": 1, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32 + ] + }, + { + "test_f1": 0.9848272333992235, + "test_acc": 0.9848275862068966, + "test_auc": 0.9979562949834948, + "val_auc": 0.9955357142857143, + "k": 2, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683 + ] + }, + { + "test_f1": 0.994482756288126, + "test_acc": 0.9944827586206897, + "test_auc": 0.9997632447643271, + "val_auc": 0.9955357142857143, + "k": 5, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 816, + 1047 + ] + }, + { + "test_f1": 0.9926436688306481, + "test_acc": 0.9926436781609196, + "test_auc": 0.998741815033281, + "val_auc": 0.9955357142857143, + "k": 10, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 816, + 1047, + 1558, + 87, + 232, + 495, + 1458 + ] + }, + { + "test_f1": 0.995862067216093, + "test_acc": 0.9958620689655172, + "test_auc": 0.999906143460144, + "val_auc": 0.9955357142857143, + "k": 16, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 816, + 1047, + 1558, + 87, + 232, + 495, + 1458, + 1627, + 1023, + 868, + 1051, + 1682, + 862 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7844f0bf82aa1f0dd2c566982146ec42c22bec80 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8211909248021941, + "test_acc": 0.8213836477987422, + "test_auc": 0.8694890784060174, + "val_auc": 0.8891556622649059, + "k": 1, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.889231622247651, + "test_acc": 0.889308176100629, + "test_auc": 0.9486159533952584, + "val_auc": 0.990796318527411, + "k": 2, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028 + ] + }, + { + "test_f1": 0.9121916882202565, + "test_acc": 0.9122012578616352, + "test_auc": 0.9635321997409723, + "val_auc": 0.9959983993597439, + "k": 5, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 3417, + 1839, + 1305 + ] + }, + { + "test_f1": 0.9164586361253299, + "test_acc": 0.9164779874213836, + "test_auc": 0.9674301470625464, + "val_auc": 0.9967987194877952, + "k": 10, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 3417, + 1839, + 1305, + 778, + 5539, + 1895, + 1602, + 4663 + ] + }, + { + "test_f1": 0.9209903279279742, + "test_acc": 0.9210062893081761, + "test_auc": 0.968834901710211, + "val_auc": 0.9979991996798719, + "k": 16, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 3417, + 1839, + 1305, + 778, + 5539, + 1895, + 1602, + 4663, + 988, + 1478, + 30649, + 1279, + 1815, + 343 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cdfb5ffdc25db55d477e3916f1bc27e5ea4c7e51 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999805522484982, + "val_auc": 1.0, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999797066940852, + "val_auc": 1.0, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686 + ] + }, + { + "test_f1": 0.9995402298850574, + "test_acc": 0.9995402298850574, + "test_auc": 0.9999966177823475, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 24163, + 4787 + ] + }, + { + "test_f1": 0.9995402298850574, + "test_acc": 0.9995402298850574, + "test_auc": 0.9999966177823476, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 24163, + 4787, + 23428, + 547, + 1840, + 4276, + 3818 + ] + }, + { + "test_f1": 0.9995402298850574, + "test_acc": 0.9995402298850574, + "test_auc": 0.9999966177823475, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 24163, + 4787, + 23428, + 547, + 1840, + 4276, + 3818, + 6777, + 1362, + 104, + 1234, + 1931, + 452 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5caf15d86cea7942b625429f28bdf2f8d4ec666e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9857458609141273, + "test_acc": 0.9857471264367816, + "test_auc": 0.9951718843010986, + "val_auc": 1.0, + "k": 1, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211 + ] + }, + { + "test_f1": 0.9885055721818145, + "test_acc": 0.9885057471264368, + "test_auc": 0.9961451174305969, + "val_auc": 1.0, + "k": 2, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586 + ] + }, + { + "test_f1": 0.9940229354395258, + "test_acc": 0.9940229885057471, + "test_auc": 0.9976688064830348, + "val_auc": 1.0, + "k": 5, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 10, + 559, + 21724 + ] + }, + { + "test_f1": 0.9949424517599984, + "test_acc": 0.9949425287356322, + "test_auc": 0.9975681855078739, + "val_auc": 1.0, + "k": 10, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 10, + 559, + 21724, + 620, + 3223, + 878, + 495, + 1874 + ] + }, + { + "test_f1": 0.994942483833445, + "test_acc": 0.9949425287356322, + "test_auc": 0.9979360016775799, + "val_auc": 1.0, + "k": 16, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 10, + 559, + 21724, + 620, + 3223, + 878, + 495, + 1874, + 1603, + 1872, + 1682, + 760, + 1023, + 610 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1f852b5eb2a4bd61edfef3ea1448afad362d854f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8565508144333954, + "test_acc": 0.856551724137931, + "test_auc": 0.9241081092050435, + "val_auc": 0.9624594155844156, + "k": 1, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495 + ] + }, + { + "test_f1": 0.9057444963914839, + "test_acc": 0.9057471264367816, + "test_auc": 0.9613074300557392, + "val_auc": 0.9878246753246753, + "k": 2, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495, + 6979 + ] + }, + { + "test_f1": 0.9291940552204188, + "test_acc": 0.9291954022988506, + "test_auc": 0.9781001407002543, + "val_auc": 0.9939123376623377, + "k": 5, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495, + 6979, + 1682, + 1708, + 1027 + ] + }, + { + "test_f1": 0.9489643305246611, + "test_acc": 0.9489655172413793, + "test_auc": 0.9918877509605498, + "val_auc": 0.9951298701298702, + "k": 10, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495, + 6979, + 1682, + 1708, + 1027, + 117, + 27304, + 1631, + 27839, + 1840 + ] + }, + { + "test_f1": 0.9627546368469049, + "test_acc": 0.9627586206896551, + "test_auc": 0.99500953785378, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 495, + 6979, + 1682, + 1708, + 1027, + 117, + 27304, + 1631, + 27839, + 1840, + 1841, + 2616, + 425, + 1079, + 2024, + 1707 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c0ae16a09679daecf19aea6dc1841f5606a8a20e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9028651813787946, + "test_acc": 0.9029885057471264, + "test_auc": 0.9470137554791926, + "val_auc": 0.9525162337662338, + "k": 1, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961 + ] + }, + { + "test_f1": 0.9149387527421725, + "test_acc": 0.9149425287356322, + "test_auc": 0.9663782198712052, + "val_auc": 0.9683441558441559, + "k": 2, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182 + ] + }, + { + "test_f1": 0.9333311348254015, + "test_acc": 0.9333333333333333, + "test_auc": 0.9729101277125386, + "val_auc": 0.9776785714285714, + "k": 5, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182, + 248, + 732, + 1264 + ] + }, + { + "test_f1": 0.9489629279558792, + "test_acc": 0.9489655172413793, + "test_auc": 0.9830415606905134, + "val_auc": 0.9882305194805194, + "k": 10, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182, + 248, + 732, + 1264, + 8734, + 1536, + 1441, + 859, + 653 + ] + }, + { + "test_f1": 0.9613780367197086, + "test_acc": 0.9613793103448276, + "test_auc": 0.9896064451539586, + "val_auc": 0.9967532467532467, + "k": 16, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182, + 248, + 732, + 1264, + 8734, + 1536, + 1441, + 859, + 653, + 1682, + 4861, + 1482, + 1960, + 1856, + 656 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b7cbc48525853a7272f54740cf80cde31f79c765 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9553721864881621, + "test_acc": 0.9554022988505747, + "test_auc": 0.9800415336327722, + "val_auc": 0.9833603896103895, + "k": 1, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786 + ] + }, + { + "test_f1": 0.9636667786177174, + "test_acc": 0.9636781609195403, + "test_auc": 0.9888885694572218, + "val_auc": 0.984577922077922, + "k": 2, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558 + ] + }, + { + "test_f1": 0.9650491817583271, + "test_acc": 0.9650574712643678, + "test_auc": 0.98894775826614, + "val_auc": 0.9862012987012987, + "k": 5, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 67 + ] + }, + { + "test_f1": 0.9659668965159478, + "test_acc": 0.9659770114942529, + "test_auc": 0.9904672195465122, + "val_auc": 0.987012987012987, + "k": 10, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 67, + 271, + 292, + 1910, + 653, + 276 + ] + }, + { + "test_f1": 0.9701012782861059, + "test_acc": 0.9701149425287356, + "test_auc": 0.9926884909897722, + "val_auc": 0.9890422077922078, + "k": 16, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 1558, + 3355, + 3930, + 67, + 271, + 292, + 1910, + 653, + 276, + 760, + 327, + 460, + 1051, + 1649, + 990 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..150608666e98621e853999614d908bfb29f1823a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9687355000046165, + "test_acc": 0.968735632183908, + "test_auc": 0.9864567549650956, + "val_auc": 0.9931006493506493, + "k": 1, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831 + ] + }, + { + "test_f1": 0.9678159558871053, + "test_acc": 0.967816091954023, + "test_auc": 0.9883076735754099, + "val_auc": 0.9935064935064934, + "k": 2, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856 + ] + }, + { + "test_f1": 0.9770100071528539, + "test_acc": 0.9770114942528736, + "test_auc": 0.9918995887223333, + "val_auc": 0.9935064935064934, + "k": 5, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 1167, + 7739, + 1168 + ] + }, + { + "test_f1": 0.981147751531588, + "test_acc": 0.9811494252873563, + "test_auc": 0.9944244141999027, + "val_auc": 0.9967532467532467, + "k": 10, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 1167, + 7739, + 1168, + 1037, + 18934, + 517, + 1449, + 1252 + ] + }, + { + "test_f1": 0.9788481065845637, + "test_acc": 0.9788505747126437, + "test_auc": 0.9947161304724281, + "val_auc": 0.9979707792207793, + "k": 16, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 1167, + 7739, + 1168, + 1037, + 18934, + 517, + 1449, + 1252, + 822, + 1536, + 1366, + 581, + 1101, + 779 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..23d01abcc9fdad7420211bb32069ddcb433f1340 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9096960781890385, + "test_acc": 0.9098850574712644, + "test_auc": 0.9303618296444613, + "val_auc": 0.9711850649350648, + "k": 1, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765 + ] + }, + { + "test_f1": 0.9415545636964161, + "test_acc": 0.9416091954022988, + "test_auc": 0.9724856593971535, + "val_auc": 0.994724025974026, + "k": 2, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765, + 5495 + ] + }, + { + "test_f1": 0.9733331642215929, + "test_acc": 0.9733333333333334, + "test_auc": 0.9969450119054061, + "val_auc": 1.0, + "k": 5, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765, + 5495, + 23840, + 789, + 1763 + ] + }, + { + "test_f1": 0.9806896061884891, + "test_acc": 0.9806896551724138, + "test_auc": 0.9981186414308134, + "val_auc": 1.0, + "k": 10, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765, + 5495, + 23840, + 789, + 1763, + 1682, + 20223, + 1051, + 901, + 1797 + ] + }, + { + "test_f1": 0.9834482338758478, + "test_acc": 0.983448275862069, + "test_auc": 0.9983714622003356, + "val_auc": 1.0, + "k": 16, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1765, + 5495, + 23840, + 789, + 1763, + 1682, + 20223, + 1051, + 901, + 1797, + 434, + 29148, + 2120, + 27505, + 1815, + 1310 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e215f95b8d7d13e2f4ba5373e36f7fe14c85d5b5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9590662690974125, + "test_acc": 0.9590804597701149, + "test_auc": 0.9711966116943558, + "val_auc": 0.9695616883116883, + "k": 1, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879 + ] + }, + { + "test_f1": 0.9774709214763567, + "test_acc": 0.9774712643678161, + "test_auc": 0.9897455388549163, + "val_auc": 0.9849837662337662, + "k": 2, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293 + ] + }, + { + "test_f1": 0.9921838882191208, + "test_acc": 0.992183908045977, + "test_auc": 0.9971851493587314, + "val_auc": 1.0, + "k": 5, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 1051, + 277, + 1283 + ] + }, + { + "test_f1": 0.9908045588248873, + "test_acc": 0.9908045977011494, + "test_auc": 0.9947651726283893, + "val_auc": 1.0, + "k": 10, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 1051, + 277, + 1283, + 571, + 4455, + 7367, + 1501, + 30210 + ] + }, + { + "test_f1": 0.9880459466881123, + "test_acc": 0.9880459770114942, + "test_auc": 0.9930952026624815, + "val_auc": 1.0, + "k": 16, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 1051, + 277, + 1283, + 571, + 4455, + 7367, + 1501, + 30210, + 30, + 91, + 1867, + 1484, + 1454, + 3268 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..43007eaeb89508247a6c348c8fc43dbc5c2df482 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9783907132386588, + "test_acc": 0.9783908045977011, + "test_auc": 0.9845753963959087, + "val_auc": 1.0, + "k": 1, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439 + ] + }, + { + "test_f1": 0.9719535961236276, + "test_acc": 0.9719540229885058, + "test_auc": 0.9871234746198387, + "val_auc": 1.0, + "k": 2, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161 + ] + }, + { + "test_f1": 0.9737928707733055, + "test_acc": 0.9737931034482759, + "test_auc": 0.986711266843444, + "val_auc": 1.0, + "k": 5, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161, + 26232, + 1121, + 1795 + ] + }, + { + "test_f1": 0.9756316172775409, + "test_acc": 0.975632183908046, + "test_auc": 0.9922352738243412, + "val_auc": 1.0, + "k": 10, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161, + 26232, + 1121, + 1795, + 5535, + 1790, + 802, + 1441, + 5797 + ] + }, + { + "test_f1": 0.9829879447485509, + "test_acc": 0.9829885057471265, + "test_auc": 0.9974176768223388, + "val_auc": 1.0, + "k": 16, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161, + 26232, + 1121, + 1795, + 5535, + 1790, + 802, + 1441, + 5797, + 493, + 822, + 5672, + 2020, + 345, + 495 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4851630a694bc80674cfa129a2eeb8c4590aba74 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9581532235325337, + "test_acc": 0.9581609195402299, + "test_auc": 0.9793760823096488, + "val_auc": 0.9967532467532467, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272 + ] + }, + { + "test_f1": 0.9586114505772325, + "test_acc": 0.9586206896551724, + "test_auc": 0.9793760823096488, + "val_auc": 0.9967532467532467, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866 + ] + }, + { + "test_f1": 0.9645948571253962, + "test_acc": 0.9645977011494253, + "test_auc": 0.9942451566643217, + "val_auc": 0.9979707792207793, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 5703, + 3198, + 1072 + ] + }, + { + "test_f1": 0.9825274724527052, + "test_acc": 0.9825287356321839, + "test_auc": 0.996965305211321, + "val_auc": 1.0, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 5703, + 3198, + 1072, + 165, + 1682, + 918, + 1726, + 508 + ] + }, + { + "test_f1": 0.9848274899879238, + "test_acc": 0.9848275862068966, + "test_auc": 0.9963844093295091, + "val_auc": 0.9987824675324676, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 5703, + 3198, + 1072, + 165, + 1682, + 918, + 1726, + 508, + 495, + 1715, + 1079, + 1787, + 1281, + 1337 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b4079c83715344107a74424e64aa7e96f873f169 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5661869999135863, + "test_acc": 0.5667924528301886, + "test_auc": 0.5721279108976962, + "val_auc": 0.584233693477391, + "k": 1, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556 + ] + }, + { + "test_f1": 0.5418203930652862, + "test_acc": 0.5441509433962264, + "test_auc": 0.5820224568346162, + "val_auc": 0.6246498599439776, + "k": 2, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556, + 1279 + ] + }, + { + "test_f1": 0.6095439264217442, + "test_acc": 0.6115723270440252, + "test_auc": 0.6193690831450708, + "val_auc": 0.6594637855142057, + "k": 5, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556, + 1279, + 6475, + 1815, + 1132 + ] + }, + { + "test_f1": 0.6077838484659012, + "test_acc": 0.6077987421383648, + "test_auc": 0.6527809281456226, + "val_auc": 0.6730692276910764, + "k": 10, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556, + 1279, + 6475, + 1815, + 1132, + 6758, + 238, + 776, + 925, + 1895 + ] + }, + { + "test_f1": 0.6160808991491125, + "test_acc": 0.6161006289308176, + "test_auc": 0.6700068301100006, + "val_auc": 0.657062825130052, + "k": 16, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1556, + 1279, + 6475, + 1815, + 1132, + 6758, + 238, + 776, + 925, + 1895, + 7740, + 581, + 1028, + 7893, + 445, + 26750 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..27f2db9837711f1a4cefa89269512201eeedfa37 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.97332937552961, + "test_acc": 0.9733333333333334, + "test_auc": 0.9856492505005683, + "val_auc": 0.9987824675324676, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963 + ] + }, + { + "test_f1": 0.9691871952586756, + "test_acc": 0.9691954022988506, + "test_auc": 0.9874397119703447, + "val_auc": 0.9991883116883117, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 3772 + ] + }, + { + "test_f1": 0.9793103360804659, + "test_acc": 0.9793103448275862, + "test_auc": 0.9923688714216139, + "val_auc": 1.0, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 3772, + 217, + 1942, + 352 + ] + }, + { + "test_f1": 0.9829879447485509, + "test_acc": 0.9829885057471265, + "test_auc": 0.9929032618107041, + "val_auc": 0.9995941558441559, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 3772, + 217, + 1942, + 352, + 1682, + 1261, + 1898, + 111, + 581 + ] + }, + { + "test_f1": 0.9820688897086743, + "test_acc": 0.9820689655172414, + "test_auc": 0.9939432937388386, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 3772, + 217, + 1942, + 352, + 1682, + 1261, + 1898, + 111, + 581, + 1726, + 1047, + 760, + 1441, + 2046, + 285 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0d138d13c05675ac11a38e9bd809875049d07b90 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9816087677547362, + "test_acc": 0.9816091954022989, + "test_auc": 0.9966469539747822, + "val_auc": 0.9995941558441559, + "k": 1, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833 + ] + }, + { + "test_f1": 0.9811490666481517, + "test_acc": 0.9811494252873563, + "test_auc": 0.9977296864007792, + "val_auc": 0.9995941558441559, + "k": 2, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675 + ] + }, + { + "test_f1": 0.9871264204536344, + "test_acc": 0.9871264367816092, + "test_auc": 0.9989650413983441, + "val_auc": 0.9995941558441559, + "k": 5, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 1682, + 829 + ] + }, + { + "test_f1": 0.9880459264723536, + "test_acc": 0.9880459770114942, + "test_auc": 0.9992406921370205, + "val_auc": 1.0, + "k": 10, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 1682, + 829, + 596, + 1588, + 913, + 491, + 361 + ] + }, + { + "test_f1": 0.991724116937924, + "test_acc": 0.9917241379310345, + "test_auc": 0.9991485267059906, + "val_auc": 1.0, + "k": 16, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 1682, + 829, + 596, + 1588, + 913, + 491, + 361, + 1278, + 1181, + 1281, + 1079, + 1515, + 1631 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ac82685bfe374a3e4450c9f872ed5037f4fbc531 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8560598230046317, + "test_acc": 0.8560919540229885, + "test_auc": 0.9106823962335625, + "val_auc": 0.942775974025974, + "k": 1, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020 + ] + }, + { + "test_f1": 0.9108021087923347, + "test_acc": 0.9108045977011494, + "test_auc": 0.9619661169435576, + "val_auc": 0.9845779220779222, + "k": 2, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760 + ] + }, + { + "test_f1": 0.9351541553160916, + "test_acc": 0.9351724137931035, + "test_auc": 0.9816337125926726, + "val_auc": 0.9882305194805194, + "k": 5, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 49 + ] + }, + { + "test_f1": 0.9549261214858282, + "test_acc": 0.9549425287356322, + "test_auc": 0.9909872355105795, + "val_auc": 0.9931006493506493, + "k": 10, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 49, + 913, + 18571, + 170, + 1960, + 364 + ] + }, + { + "test_f1": 0.9641294233835463, + "test_acc": 0.9641379310344828, + "test_auc": 0.9938375994371991, + "val_auc": 0.9955357142857143, + "k": 16, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1482, + 1682, + 49, + 913, + 18571, + 170, + 1960, + 364, + 1264, + 2023, + 6935, + 3842, + 2817, + 1120 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c91d929dc2766a5b08d8e92c3ba33757b95078ed --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9434364145421934, + "test_acc": 0.9434482758620689, + "test_auc": 0.9579167568591375, + "val_auc": 0.9594155844155845, + "k": 1, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123 + ] + }, + { + "test_f1": 0.9392849323339852, + "test_acc": 0.9393103448275862, + "test_auc": 0.9730483758590832, + "val_auc": 0.9703733766233766, + "k": 2, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491 + ] + }, + { + "test_f1": 0.95494104283966, + "test_acc": 0.9549425287356322, + "test_auc": 0.978241348287245, + "val_auc": 0.9797077922077922, + "k": 5, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 1763, + 1281, + 1682 + ] + }, + { + "test_f1": 0.9618334167696293, + "test_acc": 0.9618390804597701, + "test_auc": 0.9927781197575626, + "val_auc": 0.9866071428571429, + "k": 10, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 1763, + 1281, + 1682, + 167, + 866, + 679, + 1111, + 1486 + ] + }, + { + "test_f1": 0.9705685409822119, + "test_acc": 0.9705747126436781, + "test_auc": 0.9933015179392825, + "val_auc": 0.9935064935064934, + "k": 16, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 1763, + 1281, + 1682, + 167, + 866, + 679, + 1111, + 1486, + 7316, + 204, + 523, + 144, + 1942, + 969 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..922911500ac87dd2911ed3ac22e910d2113da54a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8567294790820809, + "test_acc": 0.8570114942528736, + "test_auc": 0.9021562483088912, + "val_auc": 0.9030032467532467, + "k": 1, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967 + ] + }, + { + "test_f1": 0.8782407474460152, + "test_acc": 0.8786206896551724, + "test_auc": 0.9276201194599275, + "val_auc": 0.932224025974026, + "k": 2, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352 + ] + }, + { + "test_f1": 0.9369874383033796, + "test_acc": 0.9370114942528736, + "test_auc": 0.9765933627360788, + "val_auc": 0.9732142857142857, + "k": 5, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 495, + 1682, + 19873 + ] + }, + { + "test_f1": 0.9438834910675611, + "test_acc": 0.9439080459770115, + "test_auc": 0.9822492762054224, + "val_auc": 0.9874188311688311, + "k": 10, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 495, + 1682, + 19873, + 178, + 960, + 859, + 108, + 1762 + ] + }, + { + "test_f1": 0.9526272552475672, + "test_acc": 0.9526436781609195, + "test_auc": 0.9882484847664917, + "val_auc": 0.9971590909090908, + "k": 16, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7967, + 352, + 495, + 1682, + 19873, + 178, + 960, + 859, + 108, + 1762, + 1168, + 1242, + 12346, + 1797, + 1037, + 507 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..604213045a22a3d551ff1b892f87b20c53d9c119 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8981986380025944, + "test_acc": 0.8983908045977012, + "test_auc": 0.9502661805292494, + "val_auc": 0.9622564935064936, + "k": 1, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335 + ] + }, + { + "test_f1": 0.9244754752008021, + "test_acc": 0.9245977011494253, + "test_auc": 0.9674808228259105, + "val_auc": 0.9691558441558441, + "k": 2, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443 + ] + }, + { + "test_f1": 0.9760919236996066, + "test_acc": 0.9760919540229885, + "test_auc": 0.9959261188375994, + "val_auc": 0.9983766233766234, + "k": 5, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1079, + 1281 + ] + }, + { + "test_f1": 0.9779308385454031, + "test_acc": 0.9779310344827586, + "test_auc": 0.9957756101520644, + "val_auc": 0.9979707792207793, + "k": 10, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1079, + 1281, + 1515, + 21, + 482, + 878, + 165 + ] + }, + { + "test_f1": 0.9816086822252236, + "test_acc": 0.9816091954022989, + "test_auc": 0.9971640104984036, + "val_auc": 1.0, + "k": 16, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1443, + 1682, + 1079, + 1281, + 1515, + 21, + 482, + 878, + 165, + 235, + 1928, + 571, + 335, + 559, + 232 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fcb5ad750a064fdd91312cf3fa38c3730edef8ea --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6967420246028291, + "test_acc": 0.6986163522012578, + "test_auc": 0.721712762736459, + "val_auc": 0.7422969187675069, + "k": 1, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313 + ] + }, + { + "test_f1": 0.696973610754988, + "test_acc": 0.6991194968553459, + "test_auc": 0.7267012998980293, + "val_auc": 0.7559023609443778, + "k": 2, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757 + ] + }, + { + "test_f1": 0.6964767873921975, + "test_acc": 0.6983647798742139, + "test_auc": 0.721712762736459, + "val_auc": 0.7422969187675069, + "k": 5, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757, + 19000, + 24630, + 1105 + ] + }, + { + "test_f1": 0.7021413402325055, + "test_acc": 0.7041509433962264, + "test_auc": 0.7738774114237514, + "val_auc": 0.7458983593437375, + "k": 10, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757, + 19000, + 24630, + 1105, + 4333, + 6718, + 1067, + 1928, + 121 + ] + }, + { + "test_f1": 0.7141966461722155, + "test_acc": 0.7154716981132075, + "test_auc": 0.7842395338310689, + "val_auc": 0.7623049219687876, + "k": 16, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 25757, + 19000, + 24630, + 1105, + 4333, + 6718, + 1067, + 1928, + 121, + 1532, + 433, + 26569, + 836, + 19189, + 800 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7a4460c5efbd082fb0a9dda5d306faedf22c0635 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6906954659754081, + "test_acc": 0.6918561995597946, + "test_auc": 0.7566391497754293, + "val_auc": 0.7728365384615385, + "k": 1, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394 + ] + }, + { + "test_f1": 0.6952214627065075, + "test_acc": 0.6962582538517975, + "test_auc": 0.7571365208142243, + "val_auc": 0.7748397435897436, + "k": 2, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041 + ] + }, + { + "test_f1": 0.6928201443284661, + "test_acc": 0.6940572267057961, + "test_auc": 0.7570805396583428, + "val_auc": 0.782051282051282, + "k": 5, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584 + ] + }, + { + "test_f1": 0.6955976180116479, + "test_acc": 0.6962582538517975, + "test_auc": 0.7635722006192378, + "val_auc": 0.8040865384615385, + "k": 10, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584, + 1551, + 465, + 28053, + 13, + 1257 + ] + }, + { + "test_f1": 0.6981375652919007, + "test_acc": 0.698459280997799, + "test_auc": 0.7711511878770654, + "val_auc": 0.8175080128205128, + "k": 16, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 1928, + 584, + 1551, + 465, + 28053, + 13, + 1257, + 7649, + 31, + 28397, + 9221, + 1697, + 943 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ab904ea5b5a5489f191523f6a7091fb1889822e8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6483554265263931, + "test_acc": 0.6490566037735849, + "test_auc": 0.6811543645364893, + "val_auc": 0.65906362545018, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927 + ] + }, + { + "test_f1": 0.6726987681914542, + "test_acc": 0.6727044025157233, + "test_auc": 0.7428473710911669, + "val_auc": 0.6948779511804722, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927, + 1011 + ] + }, + { + "test_f1": 0.7396175988541265, + "test_acc": 0.7396226415094339, + "test_auc": 0.8236543063109406, + "val_auc": 0.7887154861944778, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927, + 1011, + 4520, + 24630, + 1848 + ] + }, + { + "test_f1": 0.7466098964836733, + "test_acc": 0.7466666666666667, + "test_auc": 0.827811483900889, + "val_auc": 0.7883153261304522, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927, + 1011, + 4520, + 24630, + 1848, + 4674, + 1520, + 56, + 1928, + 6417 + ] + }, + { + "test_f1": 0.7549576665472392, + "test_acc": 0.7549685534591195, + "test_auc": 0.8411024779780849, + "val_auc": 0.7871148459383754, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1927, + 1011, + 4520, + 24630, + 1848, + 4674, + 1520, + 56, + 1928, + 6417, + 1036, + 5653, + 7387, + 4481, + 1500, + 7614 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d99e1e2496119b8e4f7adf1a70809e6471cc942f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7300275245662056, + "test_acc": 0.730062893081761, + "test_auc": 0.7890606092518877, + "val_auc": 0.7621048419367746, + "k": 1, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503 + ] + }, + { + "test_f1": 0.7443854322504775, + "test_acc": 0.7444025157232704, + "test_auc": 0.8062162608261547, + "val_auc": 0.7943177270908364, + "k": 2, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472 + ] + }, + { + "test_f1": 0.7682499649071134, + "test_acc": 0.7683018867924528, + "test_auc": 0.8381775302038704, + "val_auc": 0.8479391756702681, + "k": 5, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472, + 1643, + 1928, + 17225 + ] + }, + { + "test_f1": 0.7770552759248259, + "test_acc": 0.7771069182389937, + "test_auc": 0.8570458989467759, + "val_auc": 0.8839535814325731, + "k": 10, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472, + 1643, + 1928, + 17225, + 1848, + 7303, + 7032, + 5891, + 16477 + ] + }, + { + "test_f1": 0.7801917795537893, + "test_acc": 0.780377358490566, + "test_auc": 0.8618203939287461, + "val_auc": 0.8847539015606243, + "k": 16, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1503, + 2472, + 1643, + 1928, + 17225, + 1848, + 7303, + 7032, + 5891, + 16477, + 18220, + 7615, + 7708, + 2, + 446, + 30127 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fbb75b95b12a2b0ed82810207bcf496d568c536e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8214792341379004, + "test_acc": 0.8216352201257862, + "test_auc": 0.8571772861628755, + "val_auc": 0.8589435774309724, + "k": 1, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8499076697028964, + "test_acc": 0.8508176100628931, + "test_auc": 0.9184206902208418, + "val_auc": 0.9385754301720688, + "k": 2, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8611268512834996, + "test_acc": 0.8616352201257862, + "test_auc": 0.9285623150073059, + "val_auc": 0.940376150460184, + "k": 5, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 7284, + 100 + ] + }, + { + "test_f1": 0.8743480154159458, + "test_acc": 0.8744654088050314, + "test_auc": 0.9378538467847852, + "val_auc": 0.9451780712284914, + "k": 10, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 7284, + 100, + 1132, + 4675, + 794, + 777, + 505 + ] + }, + { + "test_f1": 0.8716160477065445, + "test_acc": 0.8716981132075472, + "test_auc": 0.9420131761884847, + "val_auc": 0.9591836734693877, + "k": 16, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 7284, + 100, + 1132, + 4675, + 794, + 777, + 505, + 1335, + 1874, + 1928, + 6745, + 1960, + 5256 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c6005d2a99809bc7073cbfadbb31c937ce3e1a52 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7935680232797132, + "test_acc": 0.7937106918238994, + "test_auc": 0.8550409148398189, + "val_auc": 0.8591436574629852, + "k": 1, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433 + ] + }, + { + "test_f1": 0.8695534244449179, + "test_acc": 0.869685534591195, + "test_auc": 0.9509829485215268, + "val_auc": 0.954781912765106, + "k": 2, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433, + 18480 + ] + }, + { + "test_f1": 0.9230824791757481, + "test_acc": 0.9232704402515723, + "test_auc": 0.968649086263935, + "val_auc": 0.9857943177270908, + "k": 5, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433, + 18480, + 1116, + 444, + 7666 + ] + }, + { + "test_f1": 0.9197344275185475, + "test_acc": 0.919748427672956, + "test_auc": 0.9783451590266308, + "val_auc": 0.9967987194877951, + "k": 10, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433, + 18480, + 1116, + 444, + 7666, + 3481, + 1927, + 6553, + 1335, + 15841 + ] + }, + { + "test_f1": 0.9272672329671755, + "test_acc": 0.9272955974842767, + "test_auc": 0.9815941446363132, + "val_auc": 0.9983993597438976, + "k": 16, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 433, + 18480, + 1116, + 444, + 7666, + 3481, + 1927, + 6553, + 1335, + 15841, + 1928, + 1064, + 4110, + 993, + 1500, + 988 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e6077c3f34f541ee919b05e1f17e8231cb1d2f37 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8026117645813822, + "test_acc": 0.8080808080808081, + "test_auc": 0.8044897959183674, + "val_auc": 0.7692307692307692, + "k": 1, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629 + ] + }, + { + "test_f1": 0.7373737373737373, + "test_acc": 0.7373737373737373, + "test_auc": 0.8257142857142856, + "val_auc": 0.821514423076923, + "k": 2, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625 + ] + }, + { + "test_f1": 0.7373737373737373, + "test_acc": 0.7373737373737373, + "test_auc": 0.8438775510204082, + "val_auc": 0.8545673076923078, + "k": 5, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 794, + 1874, + 530 + ] + }, + { + "test_f1": 0.8078454447386487, + "test_acc": 0.8080808080808081, + "test_auc": 0.8953061224489796, + "val_auc": 0.8826121794871794, + "k": 10, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 794, + 1874, + 530, + 2046, + 1345, + 6400, + 777, + 765 + ] + }, + { + "test_f1": 0.8178098497247434, + "test_acc": 0.8181818181818182, + "test_auc": 0.9061224489795919, + "val_auc": 0.889823717948718, + "k": 16, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 794, + 1874, + 530, + 2046, + 1345, + 6400, + 777, + 765, + 7669, + 25757, + 567, + 920, + 680, + 1335 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f92a280e2851e2c50870086a781deca58873ae4e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8997612596158971, + "test_acc": 0.8997867803837953, + "test_auc": 0.95033642480451, + "val_auc": 0.9695999999999999, + "k": 1, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655 + ] + }, + { + "test_f1": 0.9082855451621452, + "test_acc": 0.908315565031983, + "test_auc": 0.9630660120021821, + "val_auc": 0.9854, + "k": 2, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188 + ] + }, + { + "test_f1": 0.9488268268198404, + "test_acc": 0.9488272921108742, + "test_auc": 0.9842334969994544, + "val_auc": 0.9956, + "k": 5, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188, + 12485, + 1927, + 1417 + ] + }, + { + "test_f1": 0.9658817591285327, + "test_acc": 0.9658848614072495, + "test_auc": 0.9914166212038553, + "val_auc": 0.9976, + "k": 10, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188, + 12485, + 1927, + 1417, + 1625, + 6096, + 4080, + 727, + 9661 + ] + }, + { + "test_f1": 0.9616204690831557, + "test_acc": 0.9616204690831557, + "test_auc": 0.9889798145117294, + "val_auc": 0.9952, + "k": 16, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188, + 12485, + 1927, + 1417, + 1625, + 6096, + 4080, + 727, + 9661, + 1472, + 5406, + 984, + 19585, + 567, + 610 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/scr/saebench_scr_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/scr/saebench_scr_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..f5356d76a4d83b3aaa04dfb93936f296186b1249 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/scr/saebench_scr_custom_sae_eval_results.json @@ -0,0 +1,326 @@ +{ + "eval_type_id": "scr", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": true, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "14ca351b-a499-4d9a-b767-6afab12961fb", + "datetime_epoch_millis": 1776879415827, + "eval_result_metrics": { + "scr_metrics": { + "scr_dir1_threshold_2": 0.23953871041979088, + "scr_metric_threshold_2": 0.19488596838148195, + "scr_dir2_threshold_2": 0.18935350035518578, + "scr_dir1_threshold_5": 0.1122623265286712, + "scr_metric_threshold_5": 0.25459886084619754, + "scr_dir2_threshold_5": 0.24073747684139635, + "scr_dir1_threshold_10": -0.05460961375770672, + "scr_metric_threshold_10": 0.2850779031079488, + "scr_dir2_threshold_10": 0.2665653463192416, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results", + "scr_dir1_threshold_2": 0.48718003580863245, + "scr_metric_threshold_2": 0.06217613139950766, + "scr_dir2_threshold_2": 0.06217613139950766, + "scr_dir1_threshold_5": 0.5641028776049328, + "scr_metric_threshold_5": 0.19948190288891024, + "scr_dir2_threshold_5": 0.19948190288891024, + "scr_dir1_threshold_10": 0.5384614208981502, + "scr_metric_threshold_10": 0.22020717707797058, + "scr_dir2_threshold_10": 0.22020717707797058, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results", + "scr_dir1_threshold_2": 0.21212114370450913, + "scr_metric_threshold_2": 0.321100954221084, + "scr_dir2_threshold_2": 0.321100954221084, + "scr_dir1_threshold_5": -0.05303006015100743, + "scr_metric_threshold_5": 0.27522944141155453, + "scr_dir2_threshold_5": 0.27522944141155453, + "scr_dir1_threshold_10": -0.21212114370450913, + "scr_metric_threshold_10": 0.40061172224601954, + "scr_dir2_threshold_10": 0.40061172224601954, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results", + "scr_dir1_threshold_2": 0.3437501746229541, + "scr_metric_threshold_2": 0.08542722071821189, + "scr_dir2_threshold_2": 0.08542722071821189, + "scr_dir1_threshold_5": -0.1250002328306055, + "scr_metric_threshold_5": 0.10301508102112025, + "scr_dir2_threshold_5": 0.10301508102112025, + "scr_dir1_threshold_10": -1.4218750873114772, + "scr_metric_threshold_10": 0.11306535109152112, + "scr_dir2_threshold_10": 0.11306535109152112, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results", + "scr_dir1_threshold_2": 0.17449641072003616, + "scr_metric_threshold_2": 0.05095551557446537, + "scr_dir2_threshold_2": 0.05095551557446537, + "scr_dir1_threshold_5": -0.45637599732779155, + "scr_metric_threshold_5": 0.04140130894832733, + "scr_dir2_threshold_5": 0.04140130894832733, + "scr_dir1_threshold_10": -0.5771814201507718, + "scr_metric_threshold_10": -0.1624203737021279, + "scr_dir2_threshold_10": -0.1624203737021279, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results", + "scr_dir1_threshold_2": 0.052631529414422, + "scr_metric_threshold_2": 0.5119047252994956, + "scr_dir2_threshold_2": 0.5119047252994956, + "scr_dir1_threshold_5": 0.052631529414422, + "scr_metric_threshold_5": 0.567460267714699, + "scr_dir2_threshold_5": 0.567460267714699, + "scr_dir1_threshold_10": 0.07368420392192293, + "scr_metric_threshold_10": 0.6190476077844602, + "scr_dir2_threshold_10": 0.6190476077844602, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results", + "scr_dir1_threshold_2": 0.2786885673060324, + "scr_metric_threshold_2": 0.1183206688641517, + "scr_dir2_threshold_2": 0.1183206688641517, + "scr_dir1_threshold_5": 0.371584539269045, + "scr_metric_threshold_5": 0.18702297804342538, + "scr_dir2_threshold_5": 0.18702297804342538, + "scr_dir1_threshold_10": 0.4644808369405549, + "scr_metric_threshold_10": 0.24809157092576697, + "scr_dir2_threshold_10": 0.24809157092576697, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results", + "scr_dir1_threshold_2": 0.23720926106331344, + "scr_metric_threshold_2": 0.2789699702565123, + "scr_dir2_threshold_2": 0.2789699702565123, + "scr_dir1_threshold_5": 0.2930233189968013, + "scr_metric_threshold_5": 0.41201726948797046, + "scr_dir2_threshold_5": 0.41201726948797046, + "scr_dir1_threshold_10": 0.4093022598196558, + "scr_metric_threshold_10": 0.5536481499151593, + "scr_dir2_threshold_10": 0.5536481499151593, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results", + "scr_dir1_threshold_2": 0.13023256071842734, + "scr_metric_threshold_2": 0.13023256071842734, + "scr_dir2_threshold_2": 0.08597281650805809, + "scr_dir1_threshold_5": 0.2511626372535729, + "scr_metric_threshold_5": 0.2511626372535729, + "scr_dir2_threshold_5": 0.14027156521516332, + "scr_dir1_threshold_10": 0.2883720195248206, + "scr_metric_threshold_10": 0.2883720195248206, + "scr_dir2_threshold_10": 0.14027156521516332, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_scr", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": null +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/tpp/saebench_tpp_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/tpp/saebench_tpp_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..9a83b978563e4fcdd18fe2c6c15d40f9519259a0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-158691-tokens-650000000/tpp/saebench_tpp_custom_sae_eval_results.json @@ -0,0 +1,297 @@ +{ + "eval_type_id": "tpp", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": false, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "e0f691d4-028b-4bfe-b827-533ac75e94d4", + "datetime_epoch_millis": 1776879582374, + "eval_result_metrics": { + "tpp_metrics": { + "tpp_threshold_2_total_metric": 0.04334999322891236, + "tpp_threshold_2_intended_diff_only": 0.051099997758865354, + "tpp_threshold_2_unintended_diff_only": 0.007750004529953003, + "tpp_threshold_5_total_metric": 0.1019499957561493, + "tpp_threshold_5_intended_diff_only": 0.12179999947547912, + "tpp_threshold_5_unintended_diff_only": 0.019850003719329833, + "tpp_threshold_10_total_metric": 0.12859999835491182, + "tpp_threshold_10_intended_diff_only": 0.1802000045776367, + "tpp_threshold_10_unintended_diff_only": 0.05160000622272491, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results", + "tpp_threshold_2_total_metric": 0.07574999630451203, + "tpp_threshold_2_intended_diff_only": 0.08980001211166382, + "tpp_threshold_2_unintended_diff_only": 0.014050015807151794, + "tpp_threshold_5_total_metric": 0.17814999520778657, + "tpp_threshold_5_intended_diff_only": 0.21400001049041747, + "tpp_threshold_5_unintended_diff_only": 0.03585001528263092, + "tpp_threshold_10_total_metric": 0.20945000052452087, + "tpp_threshold_10_intended_diff_only": 0.30620001554489135, + "tpp_threshold_10_unintended_diff_only": 0.09675001502037048, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results", + "tpp_threshold_2_total_metric": 0.010949990153312683, + "tpp_threshold_2_intended_diff_only": 0.012399983406066895, + "tpp_threshold_2_unintended_diff_only": 0.0014499932527542113, + "tpp_threshold_5_total_metric": 0.025749996304512024, + "tpp_threshold_5_intended_diff_only": 0.029599988460540773, + "tpp_threshold_5_unintended_diff_only": 0.0038499921560287476, + "tpp_threshold_10_total_metric": 0.047749996185302734, + "tpp_threshold_10_intended_diff_only": 0.05419999361038208, + "tpp_threshold_10_unintended_diff_only": 0.006449997425079346, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_tpp", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1": { + "0": { + "tpp_threshold_2_total_metric": 0.17099997401237488, + "tpp_threshold_2_intended_diff_only": 0.21799999475479126, + "tpp_threshold_2_unintended_diff_only": 0.04700002074241638, + "tpp_threshold_5_total_metric": 0.21424996852874756, + "tpp_threshold_5_intended_diff_only": 0.3009999990463257, + "tpp_threshold_5_unintended_diff_only": 0.08675003051757812, + "tpp_threshold_10_total_metric": 0.23599998652935028, + "tpp_threshold_10_intended_diff_only": 0.34200000762939453, + "tpp_threshold_10_unintended_diff_only": 0.10600002110004425 + }, + "1": { + "tpp_threshold_2_total_metric": 0.0007499605417251587, + "tpp_threshold_2_intended_diff_only": 0.0009999871253967285, + "tpp_threshold_2_unintended_diff_only": 0.0002500265836715698, + "tpp_threshold_5_total_metric": 0.10799996554851532, + "tpp_threshold_5_intended_diff_only": 0.13999998569488525, + "tpp_threshold_5_unintended_diff_only": 0.032000020146369934, + "tpp_threshold_10_total_metric": 0.13725002110004425, + "tpp_threshold_10_intended_diff_only": 0.1730000376701355, + "tpp_threshold_10_unintended_diff_only": 0.03575001657009125 + }, + "2": { + "tpp_threshold_2_total_metric": 0.14925001561641693, + "tpp_threshold_2_intended_diff_only": 0.16500002145767212, + "tpp_threshold_2_unintended_diff_only": 0.015750005841255188, + "tpp_threshold_5_total_metric": 0.1937500387430191, + "tpp_threshold_5_intended_diff_only": 0.21400004625320435, + "tpp_threshold_5_unintended_diff_only": 0.02025000751018524, + "tpp_threshold_10_total_metric": 0.1757500022649765, + "tpp_threshold_10_intended_diff_only": 0.3140000104904175, + "tpp_threshold_10_unintended_diff_only": 0.13825000822544098 + }, + "6": { + "tpp_threshold_2_total_metric": 0.0495000034570694, + "tpp_threshold_2_intended_diff_only": 0.05500000715255737, + "tpp_threshold_2_unintended_diff_only": 0.005500003695487976, + "tpp_threshold_5_total_metric": 0.2317500114440918, + "tpp_threshold_5_intended_diff_only": 0.24900001287460327, + "tpp_threshold_5_unintended_diff_only": 0.017250001430511475, + "tpp_threshold_10_total_metric": 0.2957499921321869, + "tpp_threshold_10_intended_diff_only": 0.36500000953674316, + "tpp_threshold_10_unintended_diff_only": 0.06925001740455627 + }, + "9": { + "tpp_threshold_2_total_metric": 0.008250027894973755, + "tpp_threshold_2_intended_diff_only": 0.010000050067901611, + "tpp_threshold_2_unintended_diff_only": 0.0017500221729278564, + "tpp_threshold_5_total_metric": 0.14299999177455902, + "tpp_threshold_5_intended_diff_only": 0.16600000858306885, + "tpp_threshold_5_unintended_diff_only": 0.023000016808509827, + "tpp_threshold_10_total_metric": 0.20250000059604645, + "tpp_threshold_10_intended_diff_only": 0.3370000123977661, + "tpp_threshold_10_unintended_diff_only": 0.13450001180171967 + } + }, + "canrager/amazon_reviews_mcauley_1and5": { + "1": { + "tpp_threshold_2_total_metric": 0.010749980807304382, + "tpp_threshold_2_intended_diff_only": 0.0059999823570251465, + "tpp_threshold_2_unintended_diff_only": -0.004749998450279236, + "tpp_threshold_5_total_metric": 0.010749980807304382, + "tpp_threshold_5_intended_diff_only": 0.010999977588653564, + "tpp_threshold_5_unintended_diff_only": 0.00024999678134918213, + "tpp_threshold_10_total_metric": 0.015499994158744812, + "tpp_threshold_10_intended_diff_only": 0.02399998903274536, + "tpp_threshold_10_unintended_diff_only": 0.00849999487400055 + }, + "2": { + "tpp_threshold_2_total_metric": 0.005250036716461182, + "tpp_threshold_2_intended_diff_only": 0.012000024318695068, + "tpp_threshold_2_unintended_diff_only": 0.006749987602233887, + "tpp_threshold_5_total_metric": 0.027500048279762268, + "tpp_threshold_5_intended_diff_only": 0.03400003910064697, + "tpp_threshold_5_unintended_diff_only": 0.006499990820884705, + "tpp_threshold_10_total_metric": 0.0690000057220459, + "tpp_threshold_10_intended_diff_only": 0.078000009059906, + "tpp_threshold_10_unintended_diff_only": 0.009000003337860107 + }, + "3": { + "tpp_threshold_2_total_metric": 0.019249990582466125, + "tpp_threshold_2_intended_diff_only": 0.015999972820281982, + "tpp_threshold_2_unintended_diff_only": -0.003250017762184143, + "tpp_threshold_5_total_metric": 0.022000014781951904, + "tpp_threshold_5_intended_diff_only": 0.023000001907348633, + "tpp_threshold_5_unintended_diff_only": 0.0009999871253967285, + "tpp_threshold_10_total_metric": 0.018000006675720215, + "tpp_threshold_10_intended_diff_only": 0.023000001907348633, + "tpp_threshold_10_unintended_diff_only": 0.004999995231628418 + }, + "5": { + "tpp_threshold_2_total_metric": -0.0007500350475311279, + "tpp_threshold_2_intended_diff_only": 0.007999956607818604, + "tpp_threshold_2_unintended_diff_only": 0.008749991655349731, + "tpp_threshold_5_total_metric": 0.013999968767166138, + "tpp_threshold_5_intended_diff_only": 0.01699995994567871, + "tpp_threshold_5_unintended_diff_only": 0.0029999911785125732, + "tpp_threshold_10_total_metric": 0.023749977350234985, + "tpp_threshold_10_intended_diff_only": 0.02899998426437378, + "tpp_threshold_10_unintended_diff_only": 0.005250006914138794 + }, + "6": { + "tpp_threshold_2_total_metric": 0.020249977707862854, + "tpp_threshold_2_intended_diff_only": 0.019999980926513672, + "tpp_threshold_2_unintended_diff_only": -0.00024999678134918213, + "tpp_threshold_5_total_metric": 0.05449996888637543, + "tpp_threshold_5_intended_diff_only": 0.06299996376037598, + "tpp_threshold_5_unintended_diff_only": 0.00849999487400055, + "tpp_threshold_10_total_metric": 0.11249999701976776, + "tpp_threshold_10_intended_diff_only": 0.11699998378753662, + "tpp_threshold_10_unintended_diff_only": 0.00449998676776886 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6c479b18fe2b460284bae98dd772da7aece16853 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.975652430347823, + "test_acc": 0.9756563245823389, + "test_auc": 0.981118809101978, + "val_auc": 0.9860574860574861, + "k": 1, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431 + ] + }, + { + "test_f1": 0.9780379956472244, + "test_acc": 0.9780429594272076, + "test_auc": 0.9934955926420087, + "val_auc": 0.9935649935649935, + "k": 2, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928 + ] + }, + { + "test_f1": 0.979949745630883, + "test_acc": 0.9799522673031026, + "test_auc": 0.9978081687409319, + "val_auc": 0.996996996996997, + "k": 5, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472 + ] + }, + { + "test_f1": 0.985679760250418, + "test_acc": 0.9856801909307876, + "test_auc": 0.9989136536961292, + "val_auc": 0.9974259974259975, + "k": 10, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472, + 126, + 461, + 1411, + 542, + 5951 + ] + }, + { + "test_f1": 0.993317419389236, + "test_acc": 0.9933174224343675, + "test_auc": 0.9993884745218983, + "val_auc": 0.997854997854998, + "k": 16, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 1472, + 126, + 461, + 1411, + 542, + 5951, + 786, + 1634, + 1848, + 1319, + 5020, + 7026 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..812edfb04b744d55e2801a39ea370df8ddb34581 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9196326149008947, + "test_acc": 0.9196370706416073, + "test_auc": 0.9518356484748289, + "val_auc": 0.947498966515089, + "k": 1, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838 + ] + }, + { + "test_f1": 0.9513922461931187, + "test_acc": 0.9513933895009722, + "test_auc": 0.9648923408802241, + "val_auc": 0.9574204216618438, + "k": 2, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441 + ] + }, + { + "test_f1": 0.9572115893899096, + "test_acc": 0.9572261827608555, + "test_auc": 0.9690664838746531, + "val_auc": 0.9656883009508062, + "k": 5, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 943 + ] + }, + { + "test_f1": 0.9546161022498805, + "test_acc": 0.9546338302009073, + "test_auc": 0.9740243812288731, + "val_auc": 0.9702356345597355, + "k": 10, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 943, + 30737, + 1411, + 1180, + 1697, + 908 + ] + }, + { + "test_f1": 0.962396199007044, + "test_acc": 0.9624108878807518, + "test_auc": 0.9843753150138104, + "val_auc": 0.9842910293509715, + "k": 16, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 943, + 30737, + 1411, + 1180, + 1697, + 908, + 601, + 1417, + 698, + 654, + 1703, + 302 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9f7ce703dfec6253c6b477142251e3f3424f90e2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5644338571444913, + "test_acc": 0.58062893081761, + "test_auc": 0.5879629817151526, + "val_auc": 0.5704281712685074, + "k": 1, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777 + ] + }, + { + "test_f1": 0.5678715794857147, + "test_acc": 0.5874213836477987, + "test_auc": 0.5985059830548465, + "val_auc": 0.6054421768707482, + "k": 2, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921 + ] + }, + { + "test_f1": 0.5765453457366744, + "test_acc": 0.5894339622641509, + "test_auc": 0.617955847819681, + "val_auc": 0.6054421768707484, + "k": 5, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921, + 7284, + 1438, + 2046 + ] + }, + { + "test_f1": 0.5872261916464676, + "test_acc": 0.5879245283018868, + "test_auc": 0.6237158734996795, + "val_auc": 0.6670668267306922, + "k": 10, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921, + 7284, + 1438, + 2046, + 969, + 1380, + 1335, + 559, + 505 + ] + }, + { + "test_f1": 0.5955545964740855, + "test_acc": 0.6025157232704402, + "test_auc": 0.6280748152731184, + "val_auc": 0.6798719487795117, + "k": 16, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921, + 7284, + 1438, + 2046, + 969, + 1380, + 1335, + 559, + 505, + 6411, + 5951, + 1178, + 521, + 1749, + 1310 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cbf3dba1edd0bf029fd9a97ea136f36d4f37baae --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6746673233323355, + "test_acc": 0.6782389937106919, + "test_auc": 0.6957170805406165, + "val_auc": 0.7871148459383753, + "k": 1, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567 + ] + }, + { + "test_f1": 0.7012478149838559, + "test_acc": 0.7026415094339623, + "test_auc": 0.7423192906811783, + "val_auc": 0.8443377350940376, + "k": 2, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4505 + ] + }, + { + "test_f1": 0.7131112303095847, + "test_acc": 0.7139622641509434, + "test_auc": 0.7620759787714714, + "val_auc": 0.8309323729491797, + "k": 5, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4505, + 1164, + 1345, + 4470 + ] + }, + { + "test_f1": 0.7291780020521437, + "test_acc": 0.7303144654088051, + "test_auc": 0.7798965407948446, + "val_auc": 0.8447378951580633, + "k": 10, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4505, + 1164, + 1345, + 4470, + 559, + 1629, + 6411, + 433, + 4196 + ] + }, + { + "test_f1": 0.7365105336927223, + "test_acc": 0.7373584905660377, + "test_auc": 0.7881712772862641, + "val_auc": 0.8559423769507803, + "k": 16, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 4505, + 1164, + 1345, + 4470, + 559, + 1629, + 6411, + 433, + 4196, + 1731, + 993, + 1335, + 794, + 523, + 1928 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..38535a9b23a042607ac4585fd107f0794b57ae18 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7750504905927912, + "test_acc": 0.7750943396226415, + "test_auc": 0.8150500891610356, + "val_auc": 0.8555422168867547, + "k": 1, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411 + ] + }, + { + "test_f1": 0.7517808969868385, + "test_acc": 0.7529559748427673, + "test_auc": 0.8233912787241819, + "val_auc": 0.8673469387755102, + "k": 2, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786 + ] + }, + { + "test_f1": 0.764322377047156, + "test_acc": 0.7645283018867924, + "test_auc": 0.8360750815917143, + "val_auc": 0.8971588635454182, + "k": 5, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417 + ] + }, + { + "test_f1": 0.7914300694219006, + "test_acc": 0.7916981132075471, + "test_auc": 0.8890643812548161, + "val_auc": 0.9297719087635055, + "k": 10, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417, + 1758, + 1335, + 181, + 1360, + 5054 + ] + }, + { + "test_f1": 0.8080293928712197, + "test_acc": 0.8080503144654088, + "test_auc": 0.9015562929666574, + "val_auc": 0.9217687074829932, + "k": 16, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417, + 1758, + 1335, + 181, + 1360, + 5054, + 1183, + 19471, + 191, + 1108, + 126, + 12220 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..60e5101e14d4f7f6ce42a44fc8bea96279d18409 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8390296096101977, + "test_acc": 0.8392452830188679, + "test_auc": 0.8892947519034692, + "val_auc": 0.9451780712284914, + "k": 1, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8734490278335048, + "test_acc": 0.8734591194968554, + "test_auc": 0.933847929043815, + "val_auc": 0.961984793917567, + "k": 2, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.873204176776771, + "test_acc": 0.8732075471698113, + "test_auc": 0.9417483765198134, + "val_auc": 0.9663865546218487, + "k": 5, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 821, + 7867 + ] + }, + { + "test_f1": 0.8732073866791302, + "test_acc": 0.8732075471698113, + "test_auc": 0.9444642186283276, + "val_auc": 0.9679871948779513, + "k": 10, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 821, + 7867, + 505, + 7284, + 100, + 1697, + 6745 + ] + }, + { + "test_f1": 0.8835220125786164, + "test_acc": 0.8835220125786164, + "test_auc": 0.9461284566989254, + "val_auc": 0.9675870348139255, + "k": 16, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 821, + 7867, + 505, + 7284, + 100, + 1697, + 6745, + 1132, + 1438, + 5256, + 4675, + 3654, + 22141 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..de544309fc1572f07d7385991396cfa51b26c439 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5138526198657712, + "test_acc": 0.5251282051282051, + "test_auc": 0.5322819537482748, + "val_auc": 0.5634434924520604, + "k": 1, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264 + ] + }, + { + "test_f1": 0.5536947646203767, + "test_acc": 0.5558974358974359, + "test_auc": 0.5734191436361801, + "val_auc": 0.5915952672378622, + "k": 2, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945 + ] + }, + { + "test_f1": 0.5547931025416486, + "test_acc": 0.5620512820512821, + "test_auc": 0.5978094388527956, + "val_auc": 0.6258669930640556, + "k": 5, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 1851 + ] + }, + { + "test_f1": 0.5597773673070316, + "test_acc": 0.5651282051282052, + "test_auc": 0.6055475140539267, + "val_auc": 0.6546307629538964, + "k": 10, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 1851, + 470, + 1523, + 551, + 3301, + 1447 + ] + }, + { + "test_f1": 0.5703090477524809, + "test_acc": 0.5723076923076923, + "test_auc": 0.6312863971454539, + "val_auc": 0.7156262749898001, + "k": 16, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 1851, + 470, + 1523, + 551, + 3301, + 1447, + 4531, + 386, + 913, + 1924, + 789, + 1588 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2679464834edcd32d305d9c7f4b454620f51fa89 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.49806235262209475, + "test_acc": 0.49846153846153846, + "test_auc": 0.500997239707813, + "val_auc": 0.5026519787841697, + "k": 1, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239 + ] + }, + { + "test_f1": 0.4724497244002164, + "test_acc": 0.518974358974359, + "test_auc": 0.5093265055374154, + "val_auc": 0.5324357405140758, + "k": 2, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 6341 + ] + }, + { + "test_f1": 0.5230106894055988, + "test_acc": 0.5230769230769231, + "test_auc": 0.527015939004275, + "val_auc": 0.5152998776009792, + "k": 5, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 6341, + 1409, + 456, + 238 + ] + }, + { + "test_f1": 0.5164184309636769, + "test_acc": 0.5169230769230769, + "test_auc": 0.5234582758272461, + "val_auc": 0.5177478580171359, + "k": 10, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 6341, + 1409, + 456, + 238, + 3301, + 347, + 538, + 827, + 1264 + ] + }, + { + "test_f1": 0.5093812763956022, + "test_acc": 0.5097435897435898, + "test_auc": 0.5234498603022857, + "val_auc": 0.5711954304365565, + "k": 16, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 6341, + 1409, + 456, + 238, + 3301, + 347, + 538, + 827, + 1264, + 326, + 2386, + 4201, + 400, + 6129, + 26171 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7af270e51cf8cbd8565cf955784f843c219d4d1a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5301911752296812, + "test_acc": 0.5302564102564102, + "test_auc": 0.5451198370754368, + "val_auc": 0.6034271725826194, + "k": 1, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815 + ] + }, + { + "test_f1": 0.5418865668184293, + "test_acc": 0.5425641025641026, + "test_auc": 0.5577010468913051, + "val_auc": 0.6262749898000816, + "k": 2, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156 + ] + }, + { + "test_f1": 0.5361918225918226, + "test_acc": 0.5374358974358975, + "test_auc": 0.5469249671794527, + "val_auc": 0.6234190126478989, + "k": 5, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 1770, + 16809, + 191 + ] + }, + { + "test_f1": 0.5380287488052425, + "test_acc": 0.5405128205128205, + "test_auc": 0.5489678358636012, + "val_auc": 0.6177070583435333, + "k": 10, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 1770, + 16809, + 191, + 1409, + 12594, + 698, + 4347, + 6990 + ] + }, + { + "test_f1": 0.5504029787479593, + "test_acc": 0.5528205128205128, + "test_auc": 0.5685276197529201, + "val_auc": 0.6201550387596899, + "k": 16, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 1770, + 16809, + 191, + 1409, + 12594, + 698, + 4347, + 6990, + 1611, + 400, + 360, + 203, + 4110, + 848 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..69408e9265e359f5bba9479a865e1fa9852d6bc8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8374557718282082, + "test_acc": 0.8383838383838383, + "test_auc": 0.8579591836734693, + "val_auc": 0.873798076923077, + "k": 1, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997 + ] + }, + { + "test_f1": 0.8049805291184601, + "test_acc": 0.8080808080808081, + "test_auc": 0.8579591836734693, + "val_auc": 0.873798076923077, + "k": 2, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369 + ] + }, + { + "test_f1": 0.8171377433067344, + "test_acc": 0.8181818181818182, + "test_auc": 0.8602040816326529, + "val_auc": 0.9266826923076923, + "k": 5, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 2335, + 1499, + 1156 + ] + }, + { + "test_f1": 0.8478331704138156, + "test_acc": 0.8484848484848485, + "test_auc": 0.8955102040816326, + "val_auc": 0.9489182692307692, + "k": 10, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 2335, + 1499, + 1156, + 3526, + 314, + 1279, + 12946, + 461 + ] + }, + { + "test_f1": 0.8279318478603097, + "test_acc": 0.8282828282828283, + "test_auc": 0.8804081632653061, + "val_auc": 0.9453125, + "k": 16, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 2335, + 1499, + 1156, + 3526, + 314, + 1279, + 12946, + 461, + 8141, + 4939, + 337, + 4201, + 1473, + 851 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5fbfff3e9621d11dd358d01a410648c945c3058b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8780918288711562, + "test_acc": 0.8787878787878788, + "test_auc": 0.9016326530612243, + "val_auc": 0.9431089743589742, + "k": 1, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.848175159876744, + "test_acc": 0.8484848484848485, + "test_auc": 0.9179591836734693, + "val_auc": 0.9611378205128205, + "k": 2, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335 + ] + }, + { + "test_f1": 0.8784159103308039, + "test_acc": 0.8787878787878788, + "test_auc": 0.9489795918367346, + "val_auc": 0.9815705128205128, + "k": 5, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4889, + 4201 + ] + }, + { + "test_f1": 0.8784159103308039, + "test_acc": 0.8787878787878788, + "test_auc": 0.9453061224489796, + "val_auc": 0.9837740384615384, + "k": 10, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4889, + 4201, + 1924, + 2832, + 1116, + 7085, + 3526 + ] + }, + { + "test_f1": 0.89867992527567, + "test_acc": 0.898989898989899, + "test_auc": 0.9657142857142856, + "val_auc": 0.9887820512820513, + "k": 16, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 2335, + 789, + 4889, + 4201, + 1924, + 2832, + 1116, + 7085, + 3526, + 2900, + 581, + 2000, + 12946, + 7997, + 7481 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2bc522afeb67e6d7e8679c74fe51492273b4c7e8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8784159103308039, + "test_acc": 0.8787878787878788, + "test_auc": 0.900612244897959, + "val_auc": 0.8381410256410258, + "k": 1, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388 + ] + }, + { + "test_f1": 0.888410991636798, + "test_acc": 0.8888888888888888, + "test_auc": 0.9277551020408161, + "val_auc": 0.8822115384615383, + "k": 2, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335 + ] + }, + { + "test_f1": 0.8482990353199857, + "test_acc": 0.8484848484848485, + "test_auc": 0.9481632653061225, + "val_auc": 0.9471153846153846, + "k": 5, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 1769, + 1596 + ] + }, + { + "test_f1": 0.8585858585858586, + "test_acc": 0.8585858585858586, + "test_auc": 0.956938775510204, + "val_auc": 0.9499198717948718, + "k": 10, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 1769, + 1596, + 5888, + 2832, + 2900, + 2000, + 4201 + ] + }, + { + "test_f1": 0.8686600618824769, + "test_acc": 0.8686868686868687, + "test_auc": 0.9606122448979592, + "val_auc": 0.9567307692307692, + "k": 16, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 337, + 1769, + 1596, + 5888, + 2832, + 2900, + 2000, + 4201, + 7481, + 6747, + 1847, + 5463, + 3825, + 3526 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..431e1774582128cf416841e930c8f896a6ee9023 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6505723404770276, + "test_acc": 0.6718682271348071, + "test_auc": 0.6831684656558656, + "val_auc": 0.7614000000000001, + "k": 1, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.656604458809435, + "test_acc": 0.6809709579540529, + "test_auc": 0.7762283907100911, + "val_auc": 0.8576, + "k": 2, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388 + ] + }, + { + "test_f1": 0.7145894237495565, + "test_acc": 0.7156480277416558, + "test_auc": 0.8172636825642097, + "val_auc": 0.9156, + "k": 5, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 789, + 5352, + 337 + ] + }, + { + "test_f1": 0.7728549501050527, + "test_acc": 0.7728651928912007, + "test_auc": 0.8628068440253065, + "val_auc": 0.9254, + "k": 10, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 789, + 5352, + 337, + 2000, + 12946, + 239, + 7384, + 2659 + ] + }, + { + "test_f1": 0.7877715523893319, + "test_acc": 0.788036410923277, + "test_auc": 0.8854626841890869, + "val_auc": 0.9570000000000001, + "k": 16, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 5388, + 789, + 5352, + 337, + 2000, + 12946, + 239, + 7384, + 2659, + 4889, + 7481, + 1602, + 1596, + 4201, + 2335 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cb91d507cb133742198fd36eb7fd1be099d55a57 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6352343092774233, + "test_acc": 0.6549631556133507, + "test_auc": 0.6725932350390287, + "val_auc": 0.7118, + "k": 1, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.6434039282014004, + "test_acc": 0.6653662765496315, + "test_auc": 0.7496065572292009, + "val_auc": 0.8136, + "k": 2, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 3526 + ] + }, + { + "test_f1": 0.7710456314138053, + "test_acc": 0.7724317295188556, + "test_auc": 0.8655135950072226, + "val_auc": 0.9244, + "k": 5, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 3526, + 7384, + 4201, + 5388 + ] + }, + { + "test_f1": 0.7921891589860001, + "test_acc": 0.7941048981361075, + "test_auc": 0.897783042052907, + "val_auc": 0.9176, + "k": 10, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 3526, + 7384, + 4201, + 5388, + 7481, + 1924, + 2900, + 1091, + 2832 + ] + }, + { + "test_f1": 0.8152486112857276, + "test_acc": 0.8157780667533593, + "test_auc": 0.9129435531752749, + "val_auc": 0.9284, + "k": 16, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 3526, + 7384, + 4201, + 5388, + 7481, + 1924, + 2900, + 1091, + 2832, + 3825, + 3842, + 2659, + 6244, + 1614, + 5352 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..edab549792976cfb8b70c6512fe09709a319d3dc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.831087804663418, + "test_acc": 0.8318162115301257, + "test_auc": 0.8627361971858508, + "val_auc": 0.866, + "k": 1, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8265214059918895, + "test_acc": 0.8292154312960555, + "test_auc": 0.8627361971858508, + "val_auc": 0.866, + "k": 2, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889 + ] + }, + { + "test_f1": 0.8338498182595103, + "test_acc": 0.8361508452535761, + "test_auc": 0.9116418475801954, + "val_auc": 0.9259999999999999, + "k": 5, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 12946, + 1924 + ] + }, + { + "test_f1": 0.8538345823557766, + "test_acc": 0.8547897702644126, + "test_auc": 0.9349913795824623, + "val_auc": 0.9159999999999999, + "k": 10, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 12946, + 1924, + 14154, + 4201, + 1091, + 3526, + 4531 + ] + }, + { + "test_f1": 0.8364510753905378, + "test_acc": 0.8383181621153013, + "test_auc": 0.9249031612205971, + "val_auc": 0.9444, + "k": 16, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 12946, + 1924, + 14154, + 4201, + 1091, + 3526, + 4531, + 5388, + 1158, + 5352, + 470, + 1851, + 10069 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dff80a7257b344c8a3a312831fcc2b2a4fa536e6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9302339534217101, + "test_acc": 0.9302564102564103, + "test_auc": 0.9759189753256808, + "val_auc": 0.9885760913912689, + "k": 1, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958 + ] + }, + { + "test_f1": 0.9353661221629138, + "test_acc": 0.9353846153846154, + "test_auc": 0.9752289022789241, + "val_auc": 0.988984088127295, + "k": 2, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818 + ] + }, + { + "test_f1": 0.9620460118451767, + "test_acc": 0.9620512820512821, + "test_auc": 0.9895626451678056, + "val_auc": 0.9983680130558955, + "k": 5, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 2780 + ] + }, + { + "test_f1": 0.9630475368639546, + "test_acc": 0.963076923076923, + "test_auc": 0.9910332581546437, + "val_auc": 0.9991840065279478, + "k": 10, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 2780, + 1279, + 730, + 4586, + 1499, + 778 + ] + }, + { + "test_f1": 0.9764102067800808, + "test_acc": 0.9764102564102564, + "test_auc": 0.9941722489648904, + "val_auc": 1.0, + "k": 16, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 508, + 5888, + 2780, + 1279, + 730, + 4586, + 1499, + 778, + 1596, + 3475, + 1158, + 5463, + 779, + 386 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..513decf8a3d107419d284f3519b7ac15730c1770 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9341299108923368, + "test_acc": 0.9343589743589743, + "test_auc": 0.9367026290099977, + "val_auc": 0.9767441860465116, + "k": 1, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065 + ] + }, + { + "test_f1": 0.9351348940914158, + "test_acc": 0.9353846153846154, + "test_auc": 0.9793651327969839, + "val_auc": 0.993064055487556, + "k": 2, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508 + ] + }, + { + "test_f1": 0.9651142602854548, + "test_acc": 0.9651282051282051, + "test_auc": 0.9913719830343016, + "val_auc": 1.0, + "k": 5, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 5406 + ] + }, + { + "test_f1": 0.9743581651440885, + "test_acc": 0.9743589743589743, + "test_auc": 0.9979844817719729, + "val_auc": 1.0, + "k": 10, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 5406, + 3958, + 992, + 1279, + 1499, + 581 + ] + }, + { + "test_f1": 0.9774349004695236, + "test_acc": 0.9774358974358974, + "test_auc": 0.9979886895344532, + "val_auc": 1.0, + "k": 16, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 3291, + 12813, + 5406, + 3958, + 992, + 1279, + 1499, + 581, + 3475, + 778, + 2708, + 1482, + 305, + 343 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e54862248edd0bf08a90d85d8d8221fed48b24fc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9620345086959836, + "test_acc": 0.9620512820512821, + "test_auc": 0.983943178375467, + "val_auc": 0.9987760097919216, + "k": 1, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037 + ] + }, + { + "test_f1": 0.9640896451189189, + "test_acc": 0.9641025641025641, + "test_auc": 0.9834003770155182, + "val_auc": 0.9987760097919216, + "k": 2, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508 + ] + }, + { + "test_f1": 0.9764050441890869, + "test_acc": 0.9764102564102564, + "test_auc": 0.9895205675430033, + "val_auc": 1.0, + "k": 5, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 1253, + 5472 + ] + }, + { + "test_f1": 0.9784590459118746, + "test_acc": 0.9784615384615385, + "test_auc": 0.9968736324771941, + "val_auc": 1.0, + "k": 10, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 1253, + 5472, + 26034, + 1079, + 3958, + 493, + 1279 + ] + }, + { + "test_f1": 0.9764092141414832, + "test_acc": 0.9764102564102564, + "test_auc": 0.9975005890867473, + "val_auc": 1.0, + "k": 16, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 1253, + 5472, + 26034, + 1079, + 3958, + 493, + 1279, + 1039, + 778, + 581, + 1274, + 343, + 12813 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3c2cd756ca4c50184d0b94e73393c4be16dd9f88 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5673879534801727, + "test_acc": 0.5873428695275249, + "test_auc": 0.5997864060449644, + "val_auc": 0.5592, + "k": 1, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894 + ] + }, + { + "test_f1": 0.6328357465668673, + "test_acc": 0.6393584742089293, + "test_auc": 0.6725939866011504, + "val_auc": 0.6582, + "k": 2, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351 + ] + }, + { + "test_f1": 0.6750241919406703, + "test_acc": 0.6775032509752926, + "test_auc": 0.7277199408971549, + "val_auc": 0.7898000000000001, + "k": 5, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351, + 1441, + 3078, + 559 + ] + }, + { + "test_f1": 0.7047908082365567, + "test_acc": 0.70567837017772, + "test_auc": 0.7712447822799688, + "val_auc": 0.8468, + "k": 10, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351, + 1441, + 3078, + 559, + 1602, + 3274, + 654, + 619, + 1205 + ] + }, + { + "test_f1": 0.7112827261002995, + "test_acc": 0.7117468573905505, + "test_auc": 0.7828676904946933, + "val_auc": 0.8664, + "k": 16, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1894, + 1351, + 1441, + 3078, + 559, + 1602, + 3274, + 654, + 619, + 1205, + 343, + 1158, + 4510, + 2046, + 19018, + 1960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d84c15bc86fcf7f4ef70c99faf482ba2d646c080 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6252293945802673, + "test_acc": 0.6402254009536195, + "test_auc": 0.6521124156559409, + "val_auc": 0.6776, + "k": 1, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274 + ] + }, + { + "test_f1": 0.6252293945802673, + "test_acc": 0.6402254009536195, + "test_auc": 0.6525231443555429, + "val_auc": 0.6960000000000001, + "k": 2, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121 + ] + }, + { + "test_f1": 0.6745123401373795, + "test_acc": 0.6770697876029476, + "test_auc": 0.7183644956041132, + "val_auc": 0.718, + "k": 5, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 698, + 3078 + ] + }, + { + "test_f1": 0.7382187582513743, + "test_acc": 0.7386215864759428, + "test_auc": 0.8112560707430394, + "val_auc": 0.8026000000000001, + "k": 10, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 698, + 3078, + 1524, + 4510, + 52, + 1602, + 1895 + ] + }, + { + "test_f1": 0.7555194920319587, + "test_acc": 0.7555266579973993, + "test_auc": 0.8445307321267254, + "val_auc": 0.8604, + "k": 16, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 567, + 698, + 3078, + 1524, + 4510, + 52, + 1602, + 1895, + 1158, + 2046, + 1894, + 19018, + 1499, + 649 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7146ed94288fe7ed7c18284bf054313e41700217 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5442807509061387, + "test_acc": 0.5730385782401387, + "test_auc": 0.5792567351239551, + "val_auc": 0.545, + "k": 1, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.547226762042519, + "test_acc": 0.5830082358040746, + "test_auc": 0.6517494111510774, + "val_auc": 0.6958, + "k": 2, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508 + ] + }, + { + "test_f1": 0.637696587901129, + "test_acc": 0.6428261811876896, + "test_auc": 0.7050795077568727, + "val_auc": 0.7288, + "k": 5, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121 + ] + }, + { + "test_f1": 0.6525157455170013, + "test_acc": 0.6527958387516255, + "test_auc": 0.7228299019512057, + "val_auc": 0.7436, + "k": 10, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 3274, + 925, + 1441, + 51, + 567 + ] + }, + { + "test_f1": 0.6634143626705191, + "test_acc": 0.6640658864325965, + "test_auc": 0.7340450877148152, + "val_auc": 0.8200000000000001, + "k": 16, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 3274, + 925, + 1441, + 51, + 567, + 52, + 3078, + 7893, + 4510, + 957, + 1731 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..08036afa0113386a0639d202d232f24c9824397c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6080120178567012, + "test_acc": 0.6226012793176973, + "test_auc": 0.6370612838697945, + "val_auc": 0.6087999999999999, + "k": 1, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504 + ] + }, + { + "test_f1": 0.6161376529031543, + "test_acc": 0.6183368869936035, + "test_auc": 0.6828150572831424, + "val_auc": 0.6938000000000001, + "k": 2, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 1048 + ] + }, + { + "test_f1": 0.6978187050040029, + "test_acc": 0.6993603411513859, + "test_auc": 0.7778868885251864, + "val_auc": 0.779, + "k": 5, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 1048, + 696, + 927, + 1499 + ] + }, + { + "test_f1": 0.6900536501704488, + "test_acc": 0.6908315565031983, + "test_auc": 0.7677395890161848, + "val_auc": 0.7972, + "k": 10, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 1048, + 696, + 927, + 1499, + 1902, + 499, + 1818, + 7376, + 246 + ] + }, + { + "test_f1": 0.7180116182933727, + "test_acc": 0.7185501066098081, + "test_auc": 0.7981269321694853, + "val_auc": 0.8196, + "k": 16, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 1048, + 696, + 927, + 1499, + 1902, + 499, + 1818, + 7376, + 246, + 10841, + 29085, + 7129, + 1927, + 741, + 437 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..457d4c150ed30f50b060fdfca85c34ad2859a787 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9393815708101421, + "test_acc": 0.9393939393939394, + "test_auc": 0.9706122448979593, + "val_auc": 0.9520933977455718, + "k": 1, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299 + ] + }, + { + "test_f1": 0.9494949494949495, + "test_acc": 0.9494949494949495, + "test_auc": 0.9726530612244899, + "val_auc": 0.9551127214170694, + "k": 2, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114 + ] + }, + { + "test_f1": 0.9797856011091305, + "test_acc": 0.9797979797979798, + "test_auc": 0.9983673469387755, + "val_auc": 0.9738325281803542, + "k": 5, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 432, + 23633, + 2535 + ] + }, + { + "test_f1": 0.9595959595959596, + "test_acc": 0.9595959595959596, + "test_auc": 0.9951020408163265, + "val_auc": 0.9967793880837359, + "k": 10, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 432, + 23633, + 2535, + 6877, + 1232, + 6055, + 373, + 177 + ] + }, + { + "test_f1": 0.9696907835113407, + "test_acc": 0.9696969696969697, + "test_auc": 0.9975510204081632, + "val_auc": 0.999194847020934, + "k": 16, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 432, + 23633, + 2535, + 6877, + 1232, + 6055, + 373, + 177, + 1869, + 1011, + 1244, + 1960, + 1610, + 13324 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0c19843d2cb10c41fe725f55de24f652a7f427b6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8026117645813822, + "test_acc": 0.8080808080808081, + "test_auc": 0.8118367346938776, + "val_auc": 0.7286634460547504, + "k": 1, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791 + ] + }, + { + "test_f1": 0.8353300446323702, + "test_acc": 0.8383838383838383, + "test_auc": 0.8457142857142858, + "val_auc": 0.7908615136876007, + "k": 2, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 278 + ] + }, + { + "test_f1": 0.898989898989899, + "test_acc": 0.898989898989899, + "test_auc": 0.9593877551020408, + "val_auc": 0.9102254428341385, + "k": 5, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 278, + 1657, + 4580, + 114 + ] + }, + { + "test_f1": 0.8989280055456527, + "test_acc": 0.898989898989899, + "test_auc": 0.9755102040816327, + "val_auc": 0.9506843800322061, + "k": 10, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 278, + 1657, + 4580, + 114, + 1396, + 521, + 23557, + 1299, + 1658 + ] + }, + { + "test_f1": 0.9290757032692516, + "test_acc": 0.9292929292929293, + "test_auc": 0.9881632653061225, + "val_auc": 0.9810789049919485, + "k": 16, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 278, + 1657, + 4580, + 114, + 1396, + 521, + 23557, + 1299, + 1658, + 429, + 400, + 1596, + 16326, + 5387, + 1244 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..19b9b081a1b1eb0dba5a13d0ded8d8341330f6d9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9292784948597953, + "test_acc": 0.9292929292929293, + "test_auc": 0.953469387755102, + "val_auc": 0.9533011272141707, + "k": 1, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9595132999388319, + "test_acc": 0.9595959595959596, + "test_auc": 0.996734693877551, + "val_auc": 0.9871175523349437, + "k": 2, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244 + ] + }, + { + "test_f1": 0.9696784111400831, + "test_acc": 0.9696969696969697, + "test_auc": 0.9975510204081632, + "val_auc": 0.9897342995169082, + "k": 5, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244, + 1610, + 177, + 1501 + ] + }, + { + "test_f1": 0.979793856936714, + "test_acc": 0.9797979797979798, + "test_auc": 1.0, + "val_auc": 0.9927536231884059, + "k": 10, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244, + 1610, + 177, + 1501, + 15752, + 27804, + 1550, + 635, + 1360 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 0.9987755102040816, + "val_auc": 1.0, + "k": 16, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244, + 1610, + 177, + 1501, + 15752, + 27804, + 1550, + 635, + 1360, + 1960, + 3644, + 373, + 907, + 114, + 5948 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..53b17435a43a7fede544b9ef2733c1a7b507a50d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8977171585390763, + "test_acc": 0.8977777777777778, + "test_auc": 0.9354456384323642, + "val_auc": 0.9279626623376623, + "k": 1, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752 + ] + }, + { + "test_f1": 0.9155555555555556, + "test_acc": 0.9155555555555556, + "test_auc": 0.9468236409608092, + "val_auc": 0.9342532467532467, + "k": 2, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309 + ] + }, + { + "test_f1": 0.9466645596291614, + "test_acc": 0.9466666666666667, + "test_auc": 0.9618362831858407, + "val_auc": 0.9803165584415583, + "k": 5, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580 + ] + }, + { + "test_f1": 0.9466645596291614, + "test_acc": 0.9466666666666667, + "test_auc": 0.9619548040455119, + "val_auc": 0.981737012987013, + "k": 10, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 5563, + 3301, + 943, + 969, + 1327 + ] + }, + { + "test_f1": 0.9599984196282176, + "test_acc": 0.96, + "test_auc": 0.964838811630847, + "val_auc": 0.987012987012987, + "k": 16, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 5563, + 3301, + 943, + 969, + 1327, + 1295, + 1360, + 1182, + 1073, + 1915, + 1232 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cb215fbdad4355d18f47b5900237a8fe07f92d22 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8044367186402585, + "test_acc": 0.8044444444444444, + "test_auc": 0.8363226927939317, + "val_auc": 0.8636363636363638, + "k": 1, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406 + ] + }, + { + "test_f1": 0.8167937510064612, + "test_acc": 0.8177777777777778, + "test_auc": 0.8733407079646018, + "val_auc": 0.8855519480519481, + "k": 2, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854 + ] + }, + { + "test_f1": 0.8128454511853846, + "test_acc": 0.8133333333333334, + "test_auc": 0.9006005056890013, + "val_auc": 0.9034090909090909, + "k": 5, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1804, + 1232 + ] + }, + { + "test_f1": 0.8346675313960747, + "test_acc": 0.8355555555555556, + "test_auc": 0.9088969658659924, + "val_auc": 0.8952922077922079, + "k": 10, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1804, + 1232, + 1182, + 1570, + 6662, + 311, + 800 + ] + }, + { + "test_f1": 0.8397150997150998, + "test_acc": 0.84, + "test_auc": 0.9192477876106194, + "val_auc": 0.8924512987012988, + "k": 16, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1804, + 1232, + 1182, + 1570, + 6662, + 311, + 800, + 3481, + 920, + 15752, + 1411, + 629, + 748 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f0c08f06db57c0bbf559a5b0fe3362d6a5341d43 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9244444444444444, + "test_acc": 0.9244444444444444, + "test_auc": 0.9583596713021492, + "val_auc": 0.935064935064935, + "k": 1, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9155221842403964, + "test_acc": 0.9155555555555556, + "test_auc": 0.9694216182048041, + "val_auc": 0.9557629870129871, + "k": 2, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16 + ] + }, + { + "test_f1": 0.9644360149623308, + "test_acc": 0.9644444444444444, + "test_auc": 0.9892541087231352, + "val_auc": 0.9646915584415584, + "k": 5, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 21467 + ] + }, + { + "test_f1": 0.955553799690968, + "test_acc": 0.9555555555555556, + "test_auc": 0.9886219974715549, + "val_auc": 0.9784902597402597, + "k": 10, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 21467, + 20486, + 1960, + 7033, + 15752, + 1835 + ] + }, + { + "test_f1": 0.955553799690968, + "test_acc": 0.9555555555555556, + "test_auc": 0.989965233881163, + "val_auc": 0.9955357142857143, + "k": 16, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 114, + 21467, + 20486, + 1960, + 7033, + 15752, + 1835, + 969, + 2024, + 23879, + 79, + 4678, + 1869 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..da501a404562a46dc0184edfef741b364230a0ab --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6492383955814012, + "test_acc": 0.6562635457303858, + "test_auc": 0.6733387846639239, + "val_auc": 0.7214, + "k": 1, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649 + ] + }, + { + "test_f1": 0.7032757847498106, + "test_acc": 0.7091460771564803, + "test_auc": 0.7607702609874624, + "val_auc": 0.83, + "k": 2, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394 + ] + }, + { + "test_f1": 0.7032757847498106, + "test_acc": 0.7091460771564803, + "test_auc": 0.760767254738975, + "val_auc": 0.83, + "k": 5, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 7614, + 943, + 28053 + ] + }, + { + "test_f1": 0.732057764776827, + "test_acc": 0.7334200260078023, + "test_auc": 0.7914445174294771, + "val_auc": 0.8534, + "k": 10, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 7614, + 943, + 28053, + 7246, + 1928, + 1737, + 28397, + 584 + ] + }, + { + "test_f1": 0.7350380994315998, + "test_acc": 0.7360208062418726, + "test_auc": 0.8151213547358185, + "val_auc": 0.8648, + "k": 16, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 5394, + 7614, + 943, + 28053, + 7246, + 1928, + 1737, + 28397, + 584, + 1560, + 811, + 17225, + 253, + 5133, + 1551 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4ac9507f612003c9b1e71e61b2d9f4a78e068bf3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6035104106012152, + "test_acc": 0.6042479410489814, + "test_auc": 0.6251553854686966, + "val_auc": 0.6384000000000001, + "k": 1, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943 + ] + }, + { + "test_f1": 0.6193152995403572, + "test_acc": 0.6202860858257477, + "test_auc": 0.6594649478941981, + "val_auc": 0.6961999999999999, + "k": 2, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253 + ] + }, + { + "test_f1": 0.6402670316342568, + "test_acc": 0.6410923276983095, + "test_auc": 0.7016914657114813, + "val_auc": 0.75, + "k": 5, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 1011, + 7649, + 28053 + ] + }, + { + "test_f1": 0.6394856960152498, + "test_acc": 0.6402254009536195, + "test_auc": 0.7045917439397789, + "val_auc": 0.7552000000000001, + "k": 10, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 1011, + 7649, + 28053, + 1551, + 7614, + 5394, + 32305, + 1928 + ] + }, + { + "test_f1": 0.6588558650436412, + "test_acc": 0.658864325964456, + "test_auc": 0.72733626843394, + "val_auc": 0.76, + "k": 16, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 1011, + 7649, + 28053, + 1551, + 7614, + 5394, + 32305, + 1928, + 6718, + 1848, + 319, + 28397, + 1411, + 1417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d8caad6758af3676fbd009d99f918068d11ff365 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6126719959172021, + "test_acc": 0.6250541829215431, + "test_auc": 0.6446704475251812, + "val_auc": 0.6960000000000002, + "k": 1, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253 + ] + }, + { + "test_f1": 0.6241047363812928, + "test_acc": 0.6380580840918942, + "test_auc": 0.7078046720107742, + "val_auc": 0.7631999999999999, + "k": 2, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394 + ] + }, + { + "test_f1": 0.7080069703953186, + "test_acc": 0.7113133940182055, + "test_auc": 0.7790106736852549, + "val_auc": 0.7694, + "k": 5, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 1541, + 1099, + 844 + ] + }, + { + "test_f1": 0.7215666745562475, + "test_acc": 0.7230169050715215, + "test_auc": 0.8138034905551188, + "val_auc": 0.7836, + "k": 10, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 1541, + 1099, + 844, + 7649, + 7246, + 100, + 794, + 7614 + ] + }, + { + "test_f1": 0.7316387086851377, + "test_acc": 0.7329865626354573, + "test_auc": 0.8199351101263977, + "val_auc": 0.7964, + "k": 16, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 1541, + 1099, + 844, + 7649, + 7246, + 100, + 794, + 7614, + 1960, + 28397, + 28053, + 1737, + 5744, + 1857 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..eb385457d6d28fc31d0ab4ff364a411b5b7efe8c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8139573992227852, + "test_acc": 0.816271186440678, + "test_auc": 0.8337727474968101, + "val_auc": 0.8091787439613528, + "k": 1, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.8132993936383768, + "test_acc": 0.8155932203389831, + "test_auc": 0.8499722378499227, + "val_auc": 0.8435990338164252, + "k": 2, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 508 + ] + }, + { + "test_f1": 0.815988652177892, + "test_acc": 0.8183050847457627, + "test_auc": 0.9089162465573096, + "val_auc": 0.8929146537842191, + "k": 5, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 508, + 266, + 1847, + 5020 + ] + }, + { + "test_f1": 0.8592825280640672, + "test_acc": 0.8596610169491525, + "test_auc": 0.9371306806690862, + "val_auc": 0.931159420289855, + "k": 10, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 508, + 266, + 1847, + 5020, + 776, + 4413, + 6838, + 1528, + 1433 + ] + }, + { + "test_f1": 0.8711031423614813, + "test_acc": 0.8711864406779661, + "test_auc": 0.9503940019047409, + "val_auc": 0.9472624798711755, + "k": 16, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 508, + 266, + 1847, + 5020, + 776, + 4413, + 6838, + 1528, + 1433, + 1697, + 1180, + 27673, + 1628, + 1634, + 4674 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c9f9fb785239a44599c94dd84a1cd4b8ee93671c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8269930414650088, + "test_acc": 0.8284745762711865, + "test_auc": 0.8470572120917954, + "val_auc": 0.7765700483091788, + "k": 1, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528 + ] + }, + { + "test_f1": 0.827698588853815, + "test_acc": 0.8291525423728814, + "test_auc": 0.847514092508632, + "val_auc": 0.789049919484702, + "k": 2, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791 + ] + }, + { + "test_f1": 0.8294628692538736, + "test_acc": 0.8305084745762712, + "test_auc": 0.8959149191220541, + "val_auc": 0.8425925925925927, + "k": 5, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1628, + 4185, + 1815 + ] + }, + { + "test_f1": 0.831822580162747, + "test_acc": 0.8325423728813559, + "test_auc": 0.9035439947343843, + "val_auc": 0.8627214170692431, + "k": 10, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1628, + 4185, + 1815, + 203, + 1103, + 16223, + 416, + 1180 + ] + }, + { + "test_f1": 0.8464387579307641, + "test_acc": 0.8467796610169491, + "test_auc": 0.9217658933712811, + "val_auc": 0.8454106280193237, + "k": 16, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1628, + 4185, + 1815, + 203, + 1103, + 16223, + 416, + 1180, + 1866, + 2023, + 4558, + 461, + 20997, + 31149 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f0e84dae9a343f428758bd983a7cde2d75bb9af1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6925391877517628, + "test_acc": 0.6996610169491525, + "test_auc": 0.720378153578008, + "val_auc": 0.7727455716586153, + "k": 1, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 266 + ] + }, + { + "test_f1": 0.6912391077283934, + "test_acc": 0.7010169491525424, + "test_auc": 0.7892439134703424, + "val_auc": 0.8289049919484702, + "k": 2, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 266, + 5020 + ] + }, + { + "test_f1": 0.7440700996009582, + "test_acc": 0.744406779661017, + "test_auc": 0.8206031924634036, + "val_auc": 0.8913043478260869, + "k": 5, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 266, + 5020, + 1866, + 431, + 373 + ] + }, + { + "test_f1": 0.7825990104221496, + "test_acc": 0.7844067796610169, + "test_auc": 0.8617352998496064, + "val_auc": 0.9142512077294687, + "k": 10, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 266, + 5020, + 1866, + 431, + 373, + 1628, + 32278, + 27673, + 1232, + 4674 + ] + }, + { + "test_f1": 0.8438386762400216, + "test_acc": 0.8440677966101695, + "test_auc": 0.9101113059977277, + "val_auc": 0.9619565217391305, + "k": 16, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 266, + 5020, + 1866, + 431, + 373, + 1628, + 32278, + 27673, + 1232, + 4674, + 16223, + 2893, + 776, + 1528, + 1103, + 1847 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f13f9befb77ae338530d9b923f48689f0d8ef602 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5070039126060121, + "test_acc": 0.5451895043731778, + "test_auc": 0.5505575955392357, + "val_auc": 0.5358143257302921, + "k": 1, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 219 + ] + }, + { + "test_f1": 0.5567192001355702, + "test_acc": 0.5714285714285714, + "test_auc": 0.5991772065823473, + "val_auc": 0.558623449379752, + "k": 2, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 219, + 4381 + ] + }, + { + "test_f1": 0.5680234901434557, + "test_acc": 0.5743440233236151, + "test_auc": 0.6280089759281926, + "val_auc": 0.5874349739895958, + "k": 5, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 219, + 4381, + 1152, + 799, + 1122 + ] + }, + { + "test_f1": 0.6197633077025718, + "test_acc": 0.6209912536443148, + "test_auc": 0.6886644906840744, + "val_auc": 0.6730692276910765, + "k": 10, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 219, + 4381, + 1152, + 799, + 1122, + 1799, + 275, + 1039, + 1411, + 1697 + ] + }, + { + "test_f1": 0.6347241240636998, + "test_acc": 0.6355685131195336, + "test_auc": 0.7054093567251462, + "val_auc": 0.7214885954381752, + "k": 16, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 219, + 4381, + 1152, + 799, + 1122, + 1799, + 275, + 1039, + 1411, + 1697, + 673, + 993, + 801, + 26639, + 13, + 304 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..db01077b0f4ee252935efba347e12ad2370dc54e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7311678957297204, + "test_acc": 0.7463556851311953, + "test_auc": 0.7490140078879369, + "val_auc": 0.7166866746698679, + "k": 1, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7874354163870886, + "test_acc": 0.793002915451895, + "test_auc": 0.8037535699714402, + "val_auc": 0.8131252501000401, + "k": 2, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275 + ] + }, + { + "test_f1": 0.821581052728057, + "test_acc": 0.8221574344023324, + "test_auc": 0.8746770025839794, + "val_auc": 0.8785514205682272, + "k": 5, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 191, + 8006, + 534 + ] + }, + { + "test_f1": 0.827539707734179, + "test_acc": 0.8279883381924198, + "test_auc": 0.8974228206174351, + "val_auc": 0.9019607843137255, + "k": 10, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 191, + 8006, + 534, + 1928, + 433, + 1122, + 799, + 1799 + ] + }, + { + "test_f1": 0.8949454982553131, + "test_acc": 0.8950437317784257, + "test_auc": 0.9498504011967904, + "val_auc": 0.9367747098839536, + "k": 16, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 191, + 8006, + 534, + 1928, + 433, + 1122, + 799, + 1799, + 801, + 142, + 1411, + 26639, + 1239, + 635 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c3f88517e55bc9be69ed0d463142e1eae6b473f5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6892365902921849, + "test_acc": 0.6967930029154519, + "test_auc": 0.6931184550523596, + "val_auc": 0.696078431372549, + "k": 1, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7024822990420657, + "test_acc": 0.7026239067055393, + "test_auc": 0.7438460492316062, + "val_auc": 0.7458983593437375, + "k": 2, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191 + ] + }, + { + "test_f1": 0.7434358717550085, + "test_acc": 0.7434402332361516, + "test_auc": 0.8234394124847001, + "val_auc": 0.8011204481792717, + "k": 5, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 534, + 1239, + 433 + ] + }, + { + "test_f1": 0.7987106405626053, + "test_acc": 0.7988338192419825, + "test_auc": 0.8762409900720795, + "val_auc": 0.8613445378151261, + "k": 10, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 534, + 1239, + 433, + 275, + 635, + 5897, + 8006, + 1464 + ] + }, + { + "test_f1": 0.8104762950659483, + "test_acc": 0.8104956268221575, + "test_auc": 0.8778049775601795, + "val_auc": 0.874749899959984, + "k": 16, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 191, + 534, + 1239, + 433, + 275, + 635, + 5897, + 8006, + 1464, + 1195, + 3233, + 4003, + 1596, + 142, + 1411 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..307759b2678132574f5b7836413d0138e0242baa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320 + ] + }, + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 2, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 1073 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 5, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 1073, + 2779, + 1600, + 987 + ] + }, + { + "test_f1": 0.4724853125181263, + "test_acc": 0.47474747474747475, + "test_auc": 0.5055102040816326, + "val_auc": 0.5256565656565657, + "k": 10, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 1073, + 2779, + 1600, + 987, + 21417, + 1224, + 13612, + 6534, + 15517 + ] + }, + { + "test_f1": 0.46431859680813437, + "test_acc": 0.46464646464646464, + "test_auc": 0.47428571428571425, + "val_auc": 0.5638383838383838, + "k": 16, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 320, + 1073, + 2779, + 1600, + 987, + 21417, + 1224, + 13612, + 6534, + 15517, + 1116, + 1377, + 10814, + 1030, + 5883, + 1408 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7b7d5aabc3827830cc6036f3ea60bd822a3c7a05 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.46363694784747417, + "test_acc": 0.5252525252525253, + "test_auc": 0.5216326530612245, + "val_auc": 0.5719191919191919, + "k": 1, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6534 + ] + }, + { + "test_f1": 0.47159529670638245, + "test_acc": 0.5252525252525253, + "test_auc": 0.5571428571428572, + "val_auc": 0.6145454545454545, + "k": 2, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6534, + 4520 + ] + }, + { + "test_f1": 0.47665211436397875, + "test_acc": 0.494949494949495, + "test_auc": 0.5461224489795918, + "val_auc": 0.613939393939394, + "k": 5, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6534, + 4520, + 987, + 2046, + 320 + ] + }, + { + "test_f1": 0.491208380097269, + "test_acc": 0.494949494949495, + "test_auc": 0.5697959183673469, + "val_auc": 0.57010101010101, + "k": 10, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6534, + 4520, + 987, + 2046, + 320, + 5883, + 31245, + 7666, + 19180, + 21417 + ] + }, + { + "test_f1": 0.5440580924451892, + "test_acc": 0.5454545454545454, + "test_auc": 0.5644897959183673, + "val_auc": 0.6016161616161616, + "k": 16, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6534, + 4520, + 987, + 2046, + 320, + 5883, + 31245, + 7666, + 19180, + 21417, + 8700, + 7570, + 1791, + 1030, + 347, + 680 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ca7c428f5d707820c165b411ea4a273f2981732c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5958771387342816, + "test_acc": 0.5959595959595959, + "test_auc": 0.6018367346938776, + "val_auc": 0.5854545454545454, + "k": 1, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666 + ] + }, + { + "test_f1": 0.5858585858585859, + "test_acc": 0.5858585858585859, + "test_auc": 0.6085714285714285, + "val_auc": 0.6121212121212121, + "k": 2, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 1030 + ] + }, + { + "test_f1": 0.5754976232917409, + "test_acc": 0.5757575757575758, + "test_auc": 0.5922448979591837, + "val_auc": 0.6248484848484848, + "k": 5, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 1030, + 31245, + 7026, + 2779 + ] + }, + { + "test_f1": 0.5454545454545454, + "test_acc": 0.5454545454545454, + "test_auc": 0.5979591836734695, + "val_auc": 0.6004040404040404, + "k": 10, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 1030, + 31245, + 7026, + 2779, + 1116, + 111, + 2595, + 27174, + 10308 + ] + }, + { + "test_f1": 0.5530033725275726, + "test_acc": 0.5555555555555556, + "test_auc": 0.5832653061224491, + "val_auc": 0.6193939393939394, + "k": 16, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 1030, + 31245, + 7026, + 2779, + 1116, + 111, + 2595, + 27174, + 10308, + 4985, + 373, + 5883, + 680, + 1011, + 1847 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3d3c0d618ae4b727ee7e0ee4573c9bb6d2388dde --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.49984919161525415, + "test_acc": 0.5471186440677966, + "test_auc": 0.549449721091512, + "val_auc": 0.5813204508856683, + "k": 1, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 539 + ] + }, + { + "test_f1": 0.5365464592523799, + "test_acc": 0.5627118644067797, + "test_auc": 0.5700874783510386, + "val_auc": 0.6368760064412239, + "k": 2, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 539, + 217 + ] + }, + { + "test_f1": 0.5891546068544524, + "test_acc": 0.5938983050847457, + "test_auc": 0.619744955929885, + "val_auc": 0.6030595813204508, + "k": 5, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 539, + 217, + 1960, + 1319, + 1541 + ] + }, + { + "test_f1": 0.5679642340315864, + "test_acc": 0.5708474576271186, + "test_auc": 0.6054364173221107, + "val_auc": 0.605475040257649, + "k": 10, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 539, + 217, + 1960, + 1319, + 1541, + 186, + 311, + 20, + 836, + 1255 + ] + }, + { + "test_f1": 0.5763653886385651, + "test_acc": 0.5776271186440678, + "test_auc": 0.6116093589701161, + "val_auc": 0.6195652173913043, + "k": 16, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 539, + 217, + 1960, + 1319, + 1541, + 186, + 311, + 20, + 836, + 1255, + 373, + 723, + 443, + 26270, + 3780, + 1812 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c6d3288e513e909f32e532307718d3eac53e6a3d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.62642178986993, + "test_acc": 0.6264406779661017, + "test_auc": 0.6507815688740334, + "val_auc": 0.6614331723027376, + "k": 1, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319 + ] + }, + { + "test_f1": 0.6580212003818958, + "test_acc": 0.6583050847457627, + "test_auc": 0.7010071593253246, + "val_auc": 0.6807568438003221, + "k": 2, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541 + ] + }, + { + "test_f1": 0.6643178881734366, + "test_acc": 0.6650847457627118, + "test_auc": 0.720129948924998, + "val_auc": 0.7073268921095008, + "k": 5, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1960, + 1874 + ] + }, + { + "test_f1": 0.6724721095133531, + "test_acc": 0.6732203389830509, + "test_auc": 0.7321099601769424, + "val_auc": 0.7558373590982286, + "k": 10, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1960, + 1874, + 1791, + 1255, + 186, + 1335, + 908 + ] + }, + { + "test_f1": 0.6745316574892674, + "test_acc": 0.6752542372881356, + "test_auc": 0.7346655488264516, + "val_auc": 0.7342995169082125, + "k": 16, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1960, + 1874, + 1791, + 1255, + 186, + 1335, + 908, + 311, + 821, + 765, + 253, + 217, + 1108 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..97599232f506ae7daff56c18e97a750bf3544d0b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5878256486931803, + "test_acc": 0.5972881355932204, + "test_auc": 0.6055108787180137, + "val_auc": 0.6489533011272143, + "k": 1, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.5816045661026434, + "test_acc": 0.5925423728813559, + "test_auc": 0.6055108787180137, + "val_auc": 0.6489533011272143, + "k": 2, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255 + ] + }, + { + "test_f1": 0.6106966256221456, + "test_acc": 0.6128813559322034, + "test_auc": 0.6465859909616736, + "val_auc": 0.7033011272141707, + "k": 5, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 12485, + 186, + 1791 + ] + }, + { + "test_f1": 0.6078886269090481, + "test_acc": 0.6101694915254238, + "test_auc": 0.6362974484561671, + "val_auc": 0.7210144927536232, + "k": 10, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 12485, + 186, + 1791, + 20, + 4511, + 1942, + 1319, + 279 + ] + }, + { + "test_f1": 0.6171703635353333, + "test_acc": 0.6189830508474576, + "test_auc": 0.6522248329674614, + "val_auc": 0.7041062801932367, + "k": 16, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 12485, + 186, + 1791, + 20, + 4511, + 1942, + 1319, + 279, + 521, + 567, + 311, + 6212, + 908, + 1596 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a0cf8c7cdf558b7a5539dfe7d4e2415132f05fef --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5979109010580294, + "test_acc": 0.598014888337469, + "test_auc": 0.5884685483473721, + "val_auc": 0.5866346538615446, + "k": 1, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067 + ] + }, + { + "test_f1": 0.6001121989050935, + "test_acc": 0.6004962779156328, + "test_auc": 0.6291069405448007, + "val_auc": 0.6096438575430172, + "k": 2, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268 + ] + }, + { + "test_f1": 0.6129032258064516, + "test_acc": 0.6129032258064516, + "test_auc": 0.6445987882370326, + "val_auc": 0.6214485794317727, + "k": 5, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 24630, + 5731, + 1441 + ] + }, + { + "test_f1": 0.5813165238936575, + "test_acc": 0.5831265508684863, + "test_auc": 0.6456209053741194, + "val_auc": 0.678671468587435, + "k": 10, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 24630, + 5731, + 1441, + 1674, + 1619, + 1551, + 8013, + 4802 + ] + }, + { + "test_f1": 0.5813165238936575, + "test_acc": 0.5831265508684863, + "test_auc": 0.648785774099798, + "val_auc": 0.682673069227691, + "k": 16, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 24630, + 5731, + 1441, + 1674, + 1619, + 1551, + 8013, + 4802, + 1070, + 32279, + 2, + 8006, + 898, + 710 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d5b34575f6af99d107f91a8871f267d5e92b9ea3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5796594504908162, + "test_acc": 0.607940446650124, + "test_auc": 0.6169769962070835, + "val_auc": 0.6316526610644257, + "k": 1, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575 + ] + }, + { + "test_f1": 0.6323830573178306, + "test_acc": 0.6352357320099256, + "test_auc": 0.6527264666765183, + "val_auc": 0.6514605842336935, + "k": 2, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711 + ] + }, + { + "test_f1": 0.6093103024215099, + "test_acc": 0.6104218362282878, + "test_auc": 0.656630215260332, + "val_auc": 0.6914765906362546, + "k": 5, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 583, + 1625, + 1381 + ] + }, + { + "test_f1": 0.6093103024215099, + "test_acc": 0.6104218362282878, + "test_auc": 0.6365696271119651, + "val_auc": 0.7350940376150461, + "k": 10, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 583, + 1625, + 1381, + 434, + 213, + 1619, + 10, + 1799 + ] + }, + { + "test_f1": 0.6047142038866811, + "test_acc": 0.6054590570719603, + "test_auc": 0.6570119698537018, + "val_auc": 0.737094837935174, + "k": 16, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 28711, + 583, + 1625, + 1381, + 434, + 213, + 1619, + 10, + 1799, + 1551, + 765, + 433, + 680, + 32279, + 1653 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b5461f4f63f85eaf05de56a38529f17fc52c60da --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5527715419021669, + "test_acc": 0.5533498759305211, + "test_auc": 0.5810304911088124, + "val_auc": 0.5598239295718288, + "k": 1, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619 + ] + }, + { + "test_f1": 0.5827257110135522, + "test_acc": 0.5831265508684863, + "test_auc": 0.620092606275553, + "val_auc": 0.6654661864745899, + "k": 2, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042 + ] + }, + { + "test_f1": 0.611222486472682, + "test_acc": 0.6129032258064516, + "test_auc": 0.6822570316733165, + "val_auc": 0.7216886754701881, + "k": 5, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 32279, + 125 + ] + }, + { + "test_f1": 0.6597730835637167, + "test_acc": 0.6600496277915633, + "test_auc": 0.7190286192798384, + "val_auc": 0.7386954781912765, + "k": 10, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 32279, + 125, + 7319, + 1067, + 2, + 5010, + 772 + ] + }, + { + "test_f1": 0.6897193100469874, + "test_acc": 0.6898263027295285, + "test_auc": 0.7349637948869514, + "val_auc": 0.7246898759503801, + "k": 16, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 575, + 32279, + 125, + 7319, + 1067, + 2, + 5010, + 772, + 10, + 1662, + 1625, + 1872, + 2536, + 1094 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..731390750be1f8828b71e7d524c96646c0081f20 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8299429695469248, + "test_acc": 0.8326639892904953, + "test_auc": 0.8443319809034996, + "val_auc": 0.8434343434343434, + "k": 1, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731 + ] + }, + { + "test_f1": 0.8390641347249308, + "test_acc": 0.8420348058902276, + "test_auc": 0.9071984631044716, + "val_auc": 0.8929292929292929, + "k": 2, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648 + ] + }, + { + "test_f1": 0.8616837207066704, + "test_acc": 0.8634538152610441, + "test_auc": 0.951373457011369, + "val_auc": 0.962020202020202, + "k": 5, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648, + 23226, + 1158, + 6567 + ] + }, + { + "test_f1": 0.8676959558983229, + "test_acc": 0.8688085676037484, + "test_auc": 0.9539433126406789, + "val_auc": 0.9652525252525253, + "k": 10, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648, + 23226, + 1158, + 6567, + 1274, + 23697, + 737, + 5872, + 13701 + ] + }, + { + "test_f1": 0.873792365841522, + "test_acc": 0.8741633199464525, + "test_auc": 0.956606356898109, + "val_auc": 0.9688888888888889, + "k": 16, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648, + 23226, + 1158, + 6567, + 1274, + 23697, + 737, + 5872, + 13701, + 805, + 2046, + 1447, + 23003, + 3733, + 508 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..665756de85e432d8b36f03bac942db728d218463 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8755698265196111, + "test_acc": 0.8768406961178046, + "test_auc": 0.8837364338862526, + "val_auc": 0.8864646464646464, + "k": 1, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226 + ] + }, + { + "test_f1": 0.8768193105801836, + "test_acc": 0.8781793842034806, + "test_auc": 0.9312626342274662, + "val_auc": 0.9313131313131313, + "k": 2, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648 + ] + }, + { + "test_f1": 0.8870543439976374, + "test_acc": 0.8875502008032129, + "test_auc": 0.9670972459176211, + "val_auc": 0.9692929292929293, + "k": 5, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648, + 4731, + 32, + 4505 + ] + }, + { + "test_f1": 0.8845500892025715, + "test_acc": 0.8848728246318608, + "test_auc": 0.9670112256455105, + "val_auc": 0.9652525252525253, + "k": 10, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648, + 4731, + 32, + 4505, + 1065, + 1380, + 1447, + 6567, + 1682 + ] + }, + { + "test_f1": 0.8784042580820816, + "test_acc": 0.8795180722891566, + "test_auc": 0.9690255336841048, + "val_auc": 0.9648484848484848, + "k": 16, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648, + 4731, + 32, + 4505, + 1065, + 1380, + 1447, + 6567, + 1682, + 22848, + 1158, + 2046, + 1274, + 1473, + 805 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0fbbf1466a9860f6e6d5f9f9e8e95f025524d36c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8556846159306413, + "test_acc": 0.856760374832664, + "test_auc": 0.8800949090335621, + "val_auc": 0.8632323232323232, + "k": 1, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648 + ] + }, + { + "test_f1": 0.8637770673737828, + "test_acc": 0.8647925033467202, + "test_auc": 0.9081160126736534, + "val_auc": 0.8751515151515152, + "k": 2, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505 + ] + }, + { + "test_f1": 0.8940784885736239, + "test_acc": 0.8942436412315931, + "test_auc": 0.9759214921649868, + "val_auc": 0.971919191919192, + "k": 5, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32 + ] + }, + { + "test_f1": 0.8915196062410404, + "test_acc": 0.891566265060241, + "test_auc": 0.9729609611331738, + "val_auc": 0.9733333333333334, + "k": 10, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32, + 6828, + 1305, + 2832, + 1682, + 1101 + ] + }, + { + "test_f1": 0.8928750040076172, + "test_acc": 0.892904953145917, + "test_auc": 0.9737566486501986, + "val_auc": 0.9717171717171718, + "k": 16, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 4505, + 23226, + 4731, + 32, + 6828, + 1305, + 2832, + 1682, + 1101, + 1065, + 1380, + 22848, + 152, + 23, + 31914 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d753947e083f9d19bc622660a81ca446ede68f7e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7496858545459487, + "test_acc": 0.7610256410256411, + "test_auc": 0.7658590567879625, + "val_auc": 0.7401060791513667, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518 + ] + }, + { + "test_f1": 0.7615293975589982, + "test_acc": 0.7702564102564102, + "test_auc": 0.795105109906756, + "val_auc": 0.7649938800489596, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541 + ] + }, + { + "test_f1": 0.7736948079620621, + "test_acc": 0.7815384615384615, + "test_auc": 0.8611143838152624, + "val_auc": 0.8433292533659731, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 1791, + 186, + 821 + ] + }, + { + "test_f1": 0.7960623991393222, + "test_acc": 0.798974358974359, + "test_auc": 0.8936572188373111, + "val_auc": 0.8598531211750305, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 1791, + 186, + 821, + 1335, + 2581, + 253, + 374, + 1361 + ] + }, + { + "test_f1": 0.809997721704342, + "test_acc": 0.8123076923076923, + "test_auc": 0.9029647894435655, + "val_auc": 0.8906568747450021, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 1791, + 186, + 821, + 1335, + 2581, + 253, + 374, + 1361, + 1030, + 549, + 7813, + 598, + 1438, + 2203 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..439dc145c88840af59b184b0f639efad430a4148 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8458474078655818, + "test_acc": 0.8487394957983193, + "test_auc": 0.8517623164197237, + "val_auc": 0.8584905660377358, + "k": 1, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852 + ] + }, + { + "test_f1": 0.890990949530869, + "test_acc": 0.8914285714285715, + "test_auc": 0.9078740662805244, + "val_auc": 0.9433962264150944, + "k": 2, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782 + ] + }, + { + "test_f1": 0.9037557799119296, + "test_acc": 0.9042016806722689, + "test_auc": 0.9558216460217946, + "val_auc": 0.9889602569249297, + "k": 5, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 173 + ] + }, + { + "test_f1": 0.9040890589315422, + "test_acc": 0.9045378151260505, + "test_auc": 0.9674612773065492, + "val_auc": 0.9871537535126456, + "k": 10, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 173, + 1275, + 8815, + 11077, + 1688, + 6919 + ] + }, + { + "test_f1": 0.9088419595924839, + "test_acc": 0.9092436974789916, + "test_auc": 0.9734233428061118, + "val_auc": 0.9901645925331193, + "k": 16, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 60, + 173, + 1275, + 8815, + 11077, + 1688, + 6919, + 1152, + 1283, + 1521, + 42, + 3683, + 23720 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..74cd08ae2e5225a77afba765a88d3d177e53c82e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.756189061557645, + "test_acc": 0.7616806722689076, + "test_auc": 0.7711311654409904, + "val_auc": 0.769570453633079, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021 + ] + }, + { + "test_f1": 0.7727162586590951, + "test_acc": 0.7734453781512605, + "test_auc": 0.8820044778763622, + "val_auc": 0.8831794460056203, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914 + ] + }, + { + "test_f1": 0.7792358001586696, + "test_acc": 0.7805042016806722, + "test_auc": 0.894696238366922, + "val_auc": 0.9152950622240064, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1180, + 1493, + 4790 + ] + }, + { + "test_f1": 0.8224913637565298, + "test_acc": 0.8235294117647058, + "test_auc": 0.9192617379294385, + "val_auc": 0.9355680449618627, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1180, + 1493, + 4790, + 6919, + 1275, + 1794, + 1067, + 98 + ] + }, + { + "test_f1": 0.8346106366614442, + "test_acc": 0.8346218487394959, + "test_auc": 0.9300532934175036, + "val_auc": 0.965074267362505, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1180, + 1493, + 4790, + 6919, + 1275, + 1794, + 1067, + 98, + 1091, + 27957, + 2962, + 727, + 964, + 347 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..53b675641910e9135562bab60b03bc715f8fa642 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7649637989079743, + "test_acc": 0.772436974789916, + "test_auc": 0.7739273524668995, + "val_auc": 0.7800080289040545, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235 + ] + }, + { + "test_f1": 0.7866836813853826, + "test_acc": 0.7915966386554621, + "test_auc": 0.8204329999783064, + "val_auc": 0.8012846246487355, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 3494 + ] + }, + { + "test_f1": 0.8058836677533933, + "test_acc": 0.8090756302521008, + "test_auc": 0.8509867778814241, + "val_auc": 0.8366118024889603, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 3494, + 5189, + 17338, + 692 + ] + }, + { + "test_f1": 0.8255288415219044, + "test_acc": 0.8272268907563025, + "test_auc": 0.9147187814102147, + "val_auc": 0.9104777197912486, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 3494, + 5189, + 17338, + 692, + 42, + 764, + 8189, + 2046, + 32655 + ] + }, + { + "test_f1": 0.8273208252804535, + "test_acc": 0.8282352941176471, + "test_auc": 0.9283471990223513, + "val_auc": 0.9291449217181855, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4235, + 3494, + 5189, + 17338, + 692, + 42, + 764, + 8189, + 2046, + 32655, + 30916, + 576, + 1980, + 1021, + 60, + 1521 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c79b9b5d44aa0c4a4bf919e8e67df358c88ed150 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9475625986304088, + "test_acc": 0.9475630252100841, + "test_auc": 0.9871719779305956, + "val_auc": 0.9871537535126456, + "k": 1, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9872265443825546, + "test_acc": 0.9872268907563025, + "test_auc": 0.999508283257768, + "val_auc": 0.9995985547972703, + "k": 2, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847 + ] + }, + { + "test_f1": 0.9875625446111328, + "test_acc": 0.987563025210084, + "test_auc": 0.9995507661380711, + "val_auc": 0.9995985547972702, + "k": 5, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1758, + 431 + ] + }, + { + "test_f1": 0.9983193239332243, + "test_acc": 0.9983193277310924, + "test_auc": 0.9997631805395867, + "val_auc": 1.0, + "k": 10, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1758, + 431, + 1411, + 786, + 279, + 1697, + 461 + ] + }, + { + "test_f1": 0.9989915959550402, + "test_acc": 0.9989915966386554, + "test_auc": 0.9999914130348324, + "val_auc": 1.0, + "k": 16, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1758, + 431, + 1411, + 786, + 279, + 1697, + 461, + 1634, + 4674, + 27673, + 1360, + 4148, + 126 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..92731ab24fa07ea640cd800fca6858fabe55198b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9435278804281362, + "test_acc": 0.9435294117647058, + "test_auc": 0.9881581230882704, + "val_auc": 0.9899638699317543, + "k": 1, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9620083075288889, + "test_acc": 0.9620168067226891, + "test_auc": 0.9958601788981206, + "val_auc": 0.997591328783621, + "k": 2, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9623323852088321, + "test_acc": 0.9623529411764706, + "test_auc": 0.9966836236631451, + "val_auc": 0.9975913287836211, + "k": 5, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758 + ] + }, + { + "test_f1": 0.9929411190465116, + "test_acc": 0.9929411764705882, + "test_auc": 0.9998698396858797, + "val_auc": 0.9995985547972703, + "k": 10, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758, + 431, + 1847, + 279, + 1634, + 1360 + ] + }, + { + "test_f1": 0.9983193254523717, + "test_acc": 0.9983193277310924, + "test_auc": 0.9999914130348324, + "val_auc": 1.0, + "k": 16, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 1758, + 431, + 1847, + 279, + 1634, + 1360, + 6918, + 126, + 1914, + 1551, + 1472, + 1628 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..03ed8b2b82071e056e4a6e0754dbb7ca2f30fad4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9334450623193737, + "test_acc": 0.9334453781512605, + "test_auc": 0.9814654876311545, + "val_auc": 0.9811320754716981, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9680646939036484, + "test_acc": 0.9680672268907563, + "test_auc": 0.9960798244281985, + "val_auc": 0.9967884383781613, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9757898943111762, + "test_acc": 0.9757983193277311, + "test_auc": 0.9990820986181314, + "val_auc": 1.0, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 1847, + 786, + 1758 + ] + }, + { + "test_f1": 0.9959663455373453, + "test_acc": 0.9959663865546219, + "test_auc": 0.9997780947422463, + "val_auc": 1.0, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 1847, + 786, + 1758, + 461, + 1411, + 279, + 431, + 1360 + ] + }, + { + "test_f1": 0.9969747878651205, + "test_acc": 0.9969747899159664, + "test_auc": 0.9999837299607349, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 1847, + 786, + 1758, + 461, + 1411, + 279, + 431, + 1360, + 4148, + 27673, + 1634, + 1697, + 1914, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..430a219d42f4ae356c7b031f2cc34b433778ab9b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7460876554794171, + "test_acc": 0.7471698113207547, + "test_auc": 0.8124543688907475, + "val_auc": 0.8179271708683473, + "k": 1, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.8064515249737763, + "test_acc": 0.8072955974842767, + "test_auc": 0.8761862822632829, + "val_auc": 0.9359743897559023, + "k": 2, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585 + ] + }, + { + "test_f1": 0.8217539922326704, + "test_acc": 0.8218867924528301, + "test_auc": 0.9009117614595474, + "val_auc": 0.9683873549419768, + "k": 5, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 32700, + 20310, + 776 + ] + }, + { + "test_f1": 0.8609632830494558, + "test_acc": 0.8611320754716981, + "test_auc": 0.9284329530276778, + "val_auc": 0.9695878351340537, + "k": 10, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 32700, + 20310, + 776, + 508, + 1866, + 10363, + 1327, + 1499 + ] + }, + { + "test_f1": 0.8660362944494613, + "test_acc": 0.8661635220125786, + "test_auc": 0.9322530553223721, + "val_auc": 0.9683873549419768, + "k": 16, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 32700, + 20310, + 776, + 508, + 1866, + 10363, + 1327, + 1499, + 431, + 1608, + 7976, + 31906, + 3832, + 28521 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..90e74212210b27ddb45e9241367beb20efb2e452 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6861083015173717, + "test_acc": 0.6862123613312203, + "test_auc": 0.6848201727948563, + "val_auc": 0.7114845938375349, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.6764550005076133, + "test_acc": 0.6767036450079239, + "test_auc": 0.7422041390395822, + "val_auc": 0.7767106842737095, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 28521 + ] + }, + { + "test_f1": 0.757483880114319, + "test_acc": 0.757527733755943, + "test_auc": 0.8229053646775166, + "val_auc": 0.8751500600240096, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 28521, + 1327, + 431, + 1288 + ] + }, + { + "test_f1": 0.8034717218052289, + "test_acc": 0.803486529318542, + "test_auc": 0.8887181032750652, + "val_auc": 0.9279711884753902, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 28521, + 1327, + 431, + 1288, + 508, + 913, + 1042, + 5186, + 21257 + ] + }, + { + "test_f1": 0.8272531137317297, + "test_acc": 0.8272583201267829, + "test_auc": 0.9149839260598753, + "val_auc": 0.9415766306522609, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 28521, + 1327, + 431, + 1288, + 508, + 913, + 1042, + 5186, + 21257, + 1608, + 1866, + 272, + 468, + 559, + 1681 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..efca366b129dbc0ea6d64abefc8f1728344ea2de --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7030011297753292, + "test_acc": 0.7034813925570228, + "test_auc": 0.7126222099243682, + "val_auc": 0.7349033816425121, + "k": 1, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.7346915834580011, + "test_acc": 0.7346938775510204, + "test_auc": 0.7992644346061613, + "val_auc": 0.8043478260869565, + "k": 2, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501 + ] + }, + { + "test_f1": 0.7622857375395679, + "test_acc": 0.7623049219687875, + "test_auc": 0.8442111695259177, + "val_auc": 0.8079710144927537, + "k": 5, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501, + 1447, + 1528, + 32700 + ] + }, + { + "test_f1": 0.7886679736762375, + "test_acc": 0.7887154861944778, + "test_auc": 0.8533711492344587, + "val_auc": 0.8357487922705314, + "k": 10, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501, + 1447, + 1528, + 32700, + 347, + 1417, + 5186, + 776, + 559 + ] + }, + { + "test_f1": 0.7898887137736921, + "test_acc": 0.7899159663865546, + "test_auc": 0.8587495388304741, + "val_auc": 0.8349436392914654, + "k": 16, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 6501, + 1447, + 1528, + 32700, + 347, + 1417, + 5186, + 776, + 559, + 1065, + 2837, + 30052, + 1804, + 6472, + 3507 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bba54429f8382ece2520fd80444766d1ca741b89 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7282395601631866, + "test_acc": 0.7355371900826446, + "test_auc": 0.7388743852832251, + "val_auc": 0.7214885954381752, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6501 + ] + }, + { + "test_f1": 0.7340114431023521, + "test_acc": 0.7355371900826446, + "test_auc": 0.8070700018213831, + "val_auc": 0.7829131652661064, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6501, + 1866 + ] + }, + { + "test_f1": 0.762653152412818, + "test_acc": 0.7630853994490359, + "test_auc": 0.8066601906380912, + "val_auc": 0.8119247699079633, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6501, + 1866, + 5186, + 3507, + 1447 + ] + }, + { + "test_f1": 0.7657122107874651, + "test_acc": 0.7658402203856749, + "test_auc": 0.8099690364883736, + "val_auc": 0.8147258903561425, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6501, + 1866, + 5186, + 3507, + 1447, + 1681, + 1377, + 508, + 2247, + 1528 + ] + }, + { + "test_f1": 0.7711588414784238, + "test_acc": 0.7713498622589532, + "test_auc": 0.8341023617266711, + "val_auc": 0.8227290916366546, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6501, + 1866, + 5186, + 3507, + 1447, + 1681, + 1377, + 508, + 2247, + 1528, + 12612, + 1688, + 601, + 16809, + 20676, + 1585 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..80973f01b8430ad14fd42080553fdcbd6e6fa7d0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6202549802137372, + "test_acc": 0.6266666666666667, + "test_auc": 0.6352523799060088, + "val_auc": 0.6654661864745899, + "k": 1, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6440182631677188, + "test_acc": 0.6440251572327044, + "test_auc": 0.692346580742634, + "val_auc": 0.7230892356942777, + "k": 2, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585 + ] + }, + { + "test_f1": 0.6743416228990086, + "test_acc": 0.6744654088050315, + "test_auc": 0.7338244363007436, + "val_auc": 0.7815126050420168, + "k": 5, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 27673, + 4605, + 776 + ] + }, + { + "test_f1": 0.6807543128887544, + "test_acc": 0.6807547169811321, + "test_auc": 0.7426492776487814, + "val_auc": 0.7665066026410564, + "k": 10, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 27673, + 4605, + 776, + 32700, + 30052, + 2837, + 126, + 3832 + ] + }, + { + "test_f1": 0.6711716184493965, + "test_acc": 0.6711949685534592, + "test_auc": 0.7340727809230825, + "val_auc": 0.7867146858743497, + "k": 16, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 27673, + 4605, + 776, + 32700, + 30052, + 2837, + 126, + 3832, + 1804, + 957, + 7246, + 1447, + 1499, + 1411 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9f18042d84c36bf1ef94fa0498e9a510aaf5daa0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8013914284521216, + "test_acc": 0.8022641509433962, + "test_auc": 0.8417619456041736, + "val_auc": 0.8299319727891156, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293 + ] + }, + { + "test_f1": 0.8087431541092265, + "test_acc": 0.8088050314465409, + "test_auc": 0.8913915045380486, + "val_auc": 0.8887555022008803, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062 + ] + }, + { + "test_f1": 0.8313161888340864, + "test_acc": 0.8314465408805032, + "test_auc": 0.9147375445425446, + "val_auc": 0.9329731892757103, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1441 + ] + }, + { + "test_f1": 0.8391809576207353, + "test_acc": 0.8392452830188679, + "test_auc": 0.9211213430558185, + "val_auc": 0.9387755102040816, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1441, + 1313, + 1166, + 1305, + 409, + 1116 + ] + }, + { + "test_f1": 0.8444922804983549, + "test_acc": 0.8445283018867924, + "test_auc": 0.9247303144483408, + "val_auc": 0.9435774309723889, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1441, + 1313, + 1166, + 1305, + 409, + 1116, + 1928, + 765, + 14876, + 524, + 14, + 1602 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3966076084c502fcf0ed58325d804d1ecd60d492 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6202252409148961, + "test_acc": 0.6262626262626263, + "test_auc": 0.6512244897959184, + "val_auc": 0.6130068245684465, + "k": 1, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377 + ] + }, + { + "test_f1": 0.6387350864185846, + "test_acc": 0.6464646464646465, + "test_auc": 0.6914285714285713, + "val_auc": 0.6507426736250501, + "k": 2, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027 + ] + }, + { + "test_f1": 0.6037949589351458, + "test_acc": 0.6060606060606061, + "test_auc": 0.7338775510204082, + "val_auc": 0.7226013649136892, + "k": 5, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 100 + ] + }, + { + "test_f1": 0.6133183688739244, + "test_acc": 0.6161616161616161, + "test_auc": 0.7104081632653061, + "val_auc": 0.7452830188679245, + "k": 10, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 100, + 1339, + 681, + 1551, + 524, + 1790 + ] + }, + { + "test_f1": 0.6730945821854911, + "test_acc": 0.6767676767676768, + "test_auc": 0.7775510204081633, + "val_auc": 0.771979124849458, + "k": 16, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 100, + 1339, + 681, + 1551, + 524, + 1790, + 6086, + 272, + 11386, + 495, + 1305, + 334 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a126ea0d06180ce8a2795a85a3f855b2a6cb85a4 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6790843611956715, + "test_acc": 0.6805671392827356, + "test_auc": 0.7065831942125766, + "val_auc": 0.8043478260869564, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.7007447583387112, + "test_acc": 0.7039199332777314, + "test_auc": 0.7299123539232053, + "val_auc": 0.8619162640901771, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114 + ] + }, + { + "test_f1": 0.7338820804484603, + "test_acc": 0.737281067556297, + "test_auc": 0.7921313299944353, + "val_auc": 0.8683574879227054, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 30704 + ] + }, + { + "test_f1": 0.7612521978398596, + "test_acc": 0.7631359466221852, + "test_auc": 0.8348469671675013, + "val_auc": 0.8743961352657005, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 30704, + 238, + 7400, + 4481, + 3276, + 363 + ] + }, + { + "test_f1": 0.7835322751753712, + "test_acc": 0.7848206839032527, + "test_auc": 0.869449081803005, + "val_auc": 0.9074074074074074, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 30704, + 238, + 7400, + 4481, + 3276, + 363, + 6086, + 2744, + 6719, + 1670, + 801, + 1335 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6d411b55697f2b3607ef40600d3673dafd2a4bac --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 2, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1960 + ] + }, + { + "test_f1": 0.6201923600325729, + "test_acc": 0.6218867924528302, + "test_auc": 0.6769016211005338, + "val_auc": 0.5716286514605842, + "k": 5, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1960, + 3727, + 486, + 1739 + ] + }, + { + "test_f1": 0.6362505059692949, + "test_acc": 0.6374842767295598, + "test_auc": 0.6957947989902171, + "val_auc": 0.6018407362945177, + "k": 10, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1960, + 3727, + 486, + 1739, + 588, + 7143, + 293, + 1928, + 1438 + ] + }, + { + "test_f1": 0.6405886571542188, + "test_acc": 0.641006289308176, + "test_auc": 0.6977828217417236, + "val_auc": 0.6010404161664665, + "k": 16, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1960, + 3727, + 486, + 1739, + 588, + 7143, + 293, + 1928, + 1438, + 140, + 30704, + 2138, + 765, + 1848, + 2034 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8db3d1f4671673a73970a5e4428446ecebd3029b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5157164333003673, + "test_acc": 0.5350089766606823, + "test_auc": 0.5441681230839756, + "val_auc": 0.5379365716579687, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498 + ] + }, + { + "test_f1": 0.5309486935486609, + "test_acc": 0.542788749251945, + "test_auc": 0.5738575480617711, + "val_auc": 0.5343235648334002, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727 + ] + }, + { + "test_f1": 0.5403986616364274, + "test_acc": 0.542788749251945, + "test_auc": 0.579669369395181, + "val_auc": 0.5459654757125653, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 293, + 22608, + 100 + ] + }, + { + "test_f1": 0.5480339216017069, + "test_acc": 0.5481747456612807, + "test_auc": 0.5727817093086555, + "val_auc": 0.5142513046969088, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 293, + 22608, + 100, + 150, + 26677, + 6417, + 860, + 3757 + ] + }, + { + "test_f1": 0.5493222439315325, + "test_acc": 0.5493716337522442, + "test_auc": 0.5762069163109189, + "val_auc": 0.5122440786832597, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 293, + 22608, + 100, + 150, + 26677, + 6417, + 860, + 3757, + 4913, + 7400, + 19092, + 1641, + 5060, + 2012 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1bcb0b40d4ad96407cefb7a850fce66f2d67e61a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5516068401664449, + "test_acc": 0.5713207547169812, + "test_auc": 0.5817671504619057, + "val_auc": 0.6634653861544617, + "k": 1, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1345 + ] + }, + { + "test_f1": 0.5614169722480631, + "test_acc": 0.5677987421383648, + "test_auc": 0.5954139532717189, + "val_auc": 0.6904761904761905, + "k": 2, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1345, + 794 + ] + }, + { + "test_f1": 0.6016469910110992, + "test_acc": 0.6188679245283019, + "test_auc": 0.6232001976630797, + "val_auc": 0.7244897959183674, + "k": 5, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1345, + 794, + 6659, + 13203, + 7143 + ] + }, + { + "test_f1": 0.6052175729600169, + "test_acc": 0.6168553459119497, + "test_auc": 0.6652896746356346, + "val_auc": 0.7619047619047619, + "k": 10, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1345, + 794, + 6659, + 13203, + 7143, + 6400, + 3053, + 1377, + 777, + 1327 + ] + }, + { + "test_f1": 0.6163204296690155, + "test_acc": 0.6259119496855345, + "test_auc": 0.7052723993685313, + "val_auc": 0.8215286114445779, + "k": 16, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1345, + 794, + 6659, + 13203, + 7143, + 6400, + 3053, + 1377, + 777, + 1327, + 1596, + 448, + 1550, + 121, + 1224, + 1433 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..97f9f1b818ee079aa967964579eecbcac6728a38 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9937105405642113, + "test_acc": 0.9937106918238994, + "test_auc": 0.9994868557089897, + "val_auc": 1.0, + "k": 1, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9939621036560367, + "test_acc": 0.9939622641509434, + "test_auc": 0.9996825441830651, + "val_auc": 1.0, + "k": 2, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1500 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.9999921522086723, + "val_auc": 1.0, + "k": 5, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1500, + 1625, + 1703, + 1551 + ] + }, + { + "test_f1": 0.9984905620254628, + "test_acc": 0.9984905660377359, + "test_auc": 0.999987342272052, + "val_auc": 1.0, + "k": 10, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1500, + 1625, + 1703, + 1551, + 1525, + 347, + 1232, + 6718, + 1255 + ] + }, + { + "test_f1": 0.9989937094180884, + "test_acc": 0.9989937106918239, + "test_auc": 0.9999964558361745, + "val_auc": 1.0, + "k": 16, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1500, + 1625, + 1703, + 1551, + 1525, + 347, + 1232, + 6718, + 1255, + 1042, + 549, + 15467, + 1335, + 1758, + 374 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..01e78fad8a724ac4f099fd61b499e24e10056162 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6462888604580497, + "test_acc": 0.6462893081761006, + "test_auc": 0.6894051779220871, + "val_auc": 0.7607042817126851, + "k": 1, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6458246549290186, + "test_acc": 0.6460377358490565, + "test_auc": 0.7004492987112407, + "val_auc": 0.7488995598239296, + "k": 2, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928 + ] + }, + { + "test_f1": 0.6811765039963434, + "test_acc": 0.6812578616352202, + "test_auc": 0.7517594241847664, + "val_auc": 0.8359343737494997, + "k": 5, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 5827 + ] + }, + { + "test_f1": 0.6900088231605348, + "test_acc": 0.690062893081761, + "test_auc": 0.7724451895064397, + "val_auc": 0.8287314925970388, + "k": 10, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 5827, + 56, + 6400, + 406, + 152, + 1255 + ] + }, + { + "test_f1": 0.7091491458045378, + "test_acc": 0.7091823899371069, + "test_auc": 0.7842025732654611, + "val_auc": 0.8399359743897559, + "k": 16, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1928, + 7143, + 3780, + 5827, + 56, + 6400, + 406, + 152, + 1255, + 1396, + 1327, + 1691, + 30704, + 618, + 1345 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7185f497ff52960d012d1e5ca5502d1422576b43 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6668388669246382, + "test_acc": 0.6684276729559748, + "test_auc": 0.7205849591762957, + "val_auc": 0.7224889955982393, + "k": 1, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6625157604729465, + "test_acc": 0.6649056603773585, + "test_auc": 0.7205849591762957, + "val_auc": 0.7224889955982393, + "k": 2, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143 + ] + }, + { + "test_f1": 0.6627547242287147, + "test_acc": 0.6651572327044025, + "test_auc": 0.7205849591762957, + "val_auc": 0.7224889955982393, + "k": 5, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 19940, + 1011 + ] + }, + { + "test_f1": 0.7125813961264, + "test_acc": 0.7129559748427673, + "test_auc": 0.7844459813739002, + "val_auc": 0.7551020408163266, + "k": 10, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 19940, + 1011, + 1928, + 6400, + 30704, + 598, + 7068 + ] + }, + { + "test_f1": 0.7253261587890553, + "test_acc": 0.7255345911949685, + "test_auc": 0.8032667570597212, + "val_auc": 0.7923169267707083, + "k": 16, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 19940, + 1011, + 1928, + 6400, + 30704, + 598, + 7068, + 1960, + 836, + 253, + 13614, + 2592, + 320 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6298eca08b35a23d2c7c77a19f2067f9018632fb --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.77511325341667, + "test_acc": 0.7758212877792379, + "test_auc": 0.7998324909531854, + "val_auc": 0.7272, + "k": 1, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062 + ] + }, + { + "test_f1": 0.7860335167744913, + "test_acc": 0.7876478318002628, + "test_auc": 0.8806312242609904, + "val_auc": 0.8136, + "k": 2, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448 + ] + }, + { + "test_f1": 0.8178528530764549, + "test_acc": 0.8181340341655716, + "test_auc": 0.9005557664498967, + "val_auc": 0.8255999999999999, + "k": 5, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 1662 + ] + }, + { + "test_f1": 0.8430923825080424, + "test_acc": 0.8431011826544021, + "test_auc": 0.9142881652910647, + "val_auc": 0.866, + "k": 10, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 1662, + 1248, + 2073, + 140, + 296, + 32507 + ] + }, + { + "test_f1": 0.8515095695381137, + "test_acc": 0.8515111695137977, + "test_auc": 0.9198915542618247, + "val_auc": 0.8892, + "k": 16, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 5448, + 811, + 7319, + 1662, + 1248, + 2073, + 140, + 296, + 32507, + 30127, + 2046, + 3892, + 1219, + 1942, + 1313 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6b4811dadaa1b9d512e1ffd062dd323efa6f916e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5182676842555415, + "test_acc": 0.5217610062893082, + "test_auc": 0.5295056701558116, + "val_auc": 0.6174469787915167, + "k": 1, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011 + ] + }, + { + "test_f1": 0.5164200827047272, + "test_acc": 0.5164779874213836, + "test_auc": 0.5382977279884642, + "val_auc": 0.6554621848739496, + "k": 2, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411 + ] + }, + { + "test_f1": 0.5178860752602445, + "test_acc": 0.5187421383647799, + "test_auc": 0.5376966124882157, + "val_auc": 0.6514605842336935, + "k": 5, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928 + ] + }, + { + "test_f1": 0.5178860752602445, + "test_acc": 0.5187421383647799, + "test_auc": 0.5376966124882157, + "val_auc": 0.6514605842336935, + "k": 10, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928, + 1703, + 786, + 15467, + 6718, + 1202 + ] + }, + { + "test_f1": 0.5178860752602445, + "test_acc": 0.5187421383647799, + "test_auc": 0.5376966124882157, + "val_auc": 0.6514605842336935, + "k": 16, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928, + 1703, + 786, + 15467, + 6718, + 1202, + 1596, + 1417, + 4520, + 4880, + 448, + 7332 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0e45e2890a0fad9df8bb842f528a2e468b1b1902 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9932073485568238, + "test_acc": 0.9932075471698113, + "test_auc": 0.9971526440981064, + "val_auc": 1.0, + "k": 1, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.9934590325633026, + "test_acc": 0.9934591194968554, + "test_auc": 0.9988419444700412, + "val_auc": 1.0, + "k": 2, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988 + ] + }, + { + "test_f1": 0.9939622221174144, + "test_acc": 0.9939622641509434, + "test_auc": 0.99928002843432, + "val_auc": 1.0, + "k": 5, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 1744, + 292, + 6475 + ] + }, + { + "test_f1": 0.99345900689675, + "test_acc": 0.9934591194968554, + "test_auc": 0.9993476207015621, + "val_auc": 1.0, + "k": 10, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 1744, + 292, + 6475, + 487, + 152, + 6935, + 334, + 347 + ] + }, + { + "test_f1": 0.9929558812220183, + "test_acc": 0.9929559748427673, + "test_auc": 0.9991309203990931, + "val_auc": 1.0, + "k": 16, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 26988, + 1744, + 292, + 6475, + 487, + 152, + 6935, + 334, + 347, + 559, + 821, + 1851, + 771, + 1960, + 7517 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..449c2f6c665b3527c761b277da83a0865e19e80b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8532478493778678, + "test_acc": 0.8535849056603774, + "test_auc": 0.8873716632963357, + "val_auc": 0.9019607843137254, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487 + ] + }, + { + "test_f1": 0.8930367973875231, + "test_acc": 0.8930817610062893, + "test_auc": 0.940280839541527, + "val_auc": 0.9693877551020408, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744 + ] + }, + { + "test_f1": 0.9121606196359101, + "test_acc": 0.9122012578616352, + "test_auc": 0.9447235754739812, + "val_auc": 0.9767907162865146, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 6475, + 771 + ] + }, + { + "test_f1": 0.9320284954701494, + "test_acc": 0.9320754716981132, + "test_auc": 0.977152801053933, + "val_auc": 0.9779911964785915, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 6475, + 771, + 1524, + 7368, + 1028, + 559, + 24947 + ] + }, + { + "test_f1": 0.944881339228931, + "test_acc": 0.9449056603773585, + "test_auc": 0.9881662901414525, + "val_auc": 0.9831932773109244, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 26988, + 6475, + 771, + 1524, + 7368, + 1028, + 559, + 24947, + 7517, + 152, + 878, + 3417, + 52, + 6797 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..906ba466d42bff3e3db6b39b9f7bbe2f1311b389 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8546900637654345, + "test_acc": 0.8562025316455696, + "test_auc": 0.8631029291723581, + "val_auc": 0.843937575030012, + "k": 1, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104 + ] + }, + { + "test_f1": 0.8563854463585067, + "test_acc": 0.8577215189873417, + "test_auc": 0.8907861921579726, + "val_auc": 0.8891556622649059, + "k": 2, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292 + ] + }, + { + "test_f1": 0.859451949283236, + "test_acc": 0.8607594936708861, + "test_auc": 0.9327215337853636, + "val_auc": 0.9125650260104042, + "k": 5, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 4555, + 1708, + 337 + ] + }, + { + "test_f1": 0.8629805737261488, + "test_acc": 0.8637974683544304, + "test_auc": 0.9433377839032935, + "val_auc": 0.9335734293717487, + "k": 10, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 4555, + 1708, + 337, + 10639, + 677, + 918, + 1815, + 3203 + ] + }, + { + "test_f1": 0.8727757104667034, + "test_acc": 0.8729113924050633, + "test_auc": 0.9452323525671791, + "val_auc": 0.9419767907162865, + "k": 16, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 292, + 4555, + 1708, + 337, + 10639, + 677, + 918, + 1815, + 3203, + 1457, + 968, + 1380, + 1438, + 3502, + 1851 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3570d6d33c19ae38f9db8a5343c0e5b5cb07fbf6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7457981954788718, + "test_acc": 0.7538461538461538, + "test_auc": 0.7650722052041606, + "val_auc": 0.7886576907384741, + "k": 1, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468 + ] + }, + { + "test_f1": 0.8312353246090208, + "test_acc": 0.8328205128205128, + "test_auc": 0.8663404248157, + "val_auc": 0.8339453284373725, + "k": 2, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351 + ] + }, + { + "test_f1": 0.8609672304203564, + "test_acc": 0.8615384615384616, + "test_auc": 0.8970023900090888, + "val_auc": 0.8918808649530803, + "k": 5, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 1132, + 6475 + ] + }, + { + "test_f1": 0.86205042874373, + "test_acc": 0.8625641025641025, + "test_auc": 0.912255528999899, + "val_auc": 0.898204814361485, + "k": 10, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 1132, + 6475, + 1474, + 343, + 930, + 805, + 3417 + ] + }, + { + "test_f1": 0.8683787450850086, + "test_acc": 0.8687179487179487, + "test_auc": 0.9193287777291548, + "val_auc": 0.8957568339453285, + "k": 16, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 1132, + 6475, + 1474, + 343, + 930, + 805, + 3417, + 21810, + 7595, + 1305, + 27324, + 52, + 1551 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6f188f321c4843f9b91dcebf36105a81911de10b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8931080294702451, + "test_acc": 0.8935849056603774, + "test_auc": 0.9039496161670577, + "val_auc": 0.9503801520608243, + "k": 1, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.8787522554669391, + "test_acc": 0.8794968553459119, + "test_auc": 0.9457259409501803, + "val_auc": 0.9795918367346939, + "k": 2, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279 + ] + }, + { + "test_f1": 0.9028095991210349, + "test_acc": 0.9028930817610062, + "test_auc": 0.9708264686255428, + "val_auc": 0.9795918367346939, + "k": 5, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279, + 1447, + 11381, + 1556 + ] + }, + { + "test_f1": 0.9242676611852982, + "test_acc": 0.9242767295597484, + "test_auc": 0.9791025974670368, + "val_auc": 0.9915966386554622, + "k": 10, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279, + 1447, + 11381, + 1556, + 238, + 7740, + 4877, + 1478, + 3301 + ] + }, + { + "test_f1": 0.9255268613574275, + "test_acc": 0.9255345911949685, + "test_auc": 0.9796382725137944, + "val_auc": 0.9911964785914366, + "k": 16, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279, + 1447, + 11381, + 1556, + 238, + 7740, + 4877, + 1478, + 3301, + 262, + 1101, + 848, + 6828, + 88, + 6475 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f9a21e9ea948a4e6b48ee17ab62224e1bb6bbd10 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.829577908077483, + "test_acc": 0.8328205128205128, + "test_auc": 0.8344939744841283, + "val_auc": 0.857609139126887, + "k": 1, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628 + ] + }, + { + "test_f1": 0.8990857243373654, + "test_acc": 0.8994871794871795, + "test_auc": 0.9099707139731378, + "val_auc": 0.9004487964096287, + "k": 2, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416 + ] + }, + { + "test_f1": 0.9010643322703321, + "test_acc": 0.9015384615384615, + "test_auc": 0.910033830410341, + "val_auc": 0.9188086495308037, + "k": 5, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 238 + ] + }, + { + "test_f1": 0.9001023803761079, + "test_acc": 0.9005128205128206, + "test_auc": 0.9229495573433871, + "val_auc": 0.9184006527947776, + "k": 10, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 238, + 805, + 737, + 4211, + 930, + 1132 + ] + }, + { + "test_f1": 0.9021621251106211, + "test_acc": 0.9025641025641026, + "test_auc": 0.9265156360453766, + "val_auc": 0.919624643002856, + "k": 16, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 238, + 805, + 737, + 4211, + 930, + 1132, + 1524, + 2062, + 334, + 878, + 1158, + 1274 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..72ca2e39d40d96e626647e2f6968eed242f3562e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/61_wikidata_occupation_isactor_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8707175528225678, + "test_acc": 0.8707692307692307, + "test_auc": 0.9134673645942034, + "val_auc": 0.9385964912280701, + "k": 1, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524 + ] + }, + { + "test_f1": 0.8727334952766532, + "test_acc": 0.8728205128205129, + "test_auc": 0.9134673645942034, + "val_auc": 0.9373725010199918, + "k": 2, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 3816 + ] + }, + { + "test_f1": 0.8727334952766532, + "test_acc": 0.8728205128205129, + "test_auc": 0.9134673645942034, + "val_auc": 0.9373725010199918, + "k": 5, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 3816, + 24947, + 930, + 888 + ] + }, + { + "test_f1": 0.8727334952766532, + "test_acc": 0.8728205128205129, + "test_auc": 0.9134673645942034, + "val_auc": 0.9373725010199918, + "k": 10, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 3816, + 24947, + 930, + 888, + 969, + 1474, + 862, + 3203, + 805 + ] + }, + { + "test_f1": 0.8738434996771706, + "test_acc": 0.8738461538461538, + "test_auc": 0.9349690308681455, + "val_auc": 0.9494084047327621, + "k": 16, + "dataset": "61_wikidata_occupation_isactor", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1524, + 3816, + 24947, + 930, + 888, + 969, + 1474, + 862, + 3203, + 805, + 7309, + 6631, + 1682, + 1214, + 1012, + 776 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bc4e376a5fae70ff5c55c4d66e27663c7e242c93 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8054275715424559, + "test_acc": 0.8082051282051282, + "test_auc": 0.8239556333524085, + "val_auc": 0.7853937168502652, + "k": 1, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.7926796967560457, + "test_acc": 0.796923076923077, + "test_auc": 0.8656924293937456, + "val_auc": 0.8049775601795186, + "k": 2, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815 + ] + }, + { + "test_f1": 0.84220453898071, + "test_acc": 0.8430769230769231, + "test_auc": 0.9142731511091662, + "val_auc": 0.9061607507139943, + "k": 5, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 334, + 1524, + 1866 + ] + }, + { + "test_f1": 0.889176910836379, + "test_acc": 0.8892307692307693, + "test_auc": 0.9455263069310264, + "val_auc": 0.9204406364749083, + "k": 10, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 334, + 1524, + 1866, + 487, + 26988, + 930, + 6935, + 1158 + ] + }, + { + "test_f1": 0.8860819470283192, + "test_acc": 0.8861538461538462, + "test_auc": 0.9507565556939441, + "val_auc": 0.9330885352917176, + "k": 16, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 334, + 1524, + 1866, + 487, + 26988, + 930, + 6935, + 1158, + 6758, + 2046, + 913, + 1744, + 1221, + 508 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7bc4fc347a58afaedf3b9f1d500d4807e487782d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8486480862133037, + "test_acc": 0.8492307692307692, + "test_auc": 0.8767588447167335, + "val_auc": 0.8875968992248062, + "k": 1, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52 + ] + }, + { + "test_f1": 0.8486480862133037, + "test_acc": 0.8492307692307692, + "test_auc": 0.8762118355943044, + "val_auc": 0.8875968992248062, + "k": 2, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230 + ] + }, + { + "test_f1": 0.8496583240128098, + "test_acc": 0.8502564102564103, + "test_auc": 0.8840340660450399, + "val_auc": 0.8969808241534067, + "k": 5, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 2884, + 2131, + 1274 + ] + }, + { + "test_f1": 0.8246526404123958, + "test_acc": 0.8266666666666667, + "test_auc": 0.8767588447167335, + "val_auc": 0.8875968992248062, + "k": 10, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 2884, + 2131, + 1274, + 17546, + 2046, + 888, + 1193, + 238 + ] + }, + { + "test_f1": 0.848608762013277, + "test_acc": 0.8492307692307692, + "test_auc": 0.8939601777358871, + "val_auc": 0.9261525907792737, + "k": 16, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 2884, + 2131, + 1274, + 17546, + 2046, + 888, + 1193, + 238, + 821, + 2484, + 1682, + 2454, + 1815, + 862 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..97cd19e8e4b0db5ab2f11ac924b8609c2791be5d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.871428336128179, + "test_acc": 0.8716577540106952, + "test_auc": 0.9208555244422482, + "val_auc": 0.926060606060606, + "k": 1, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839 + ] + }, + { + "test_f1": 0.8768566048036167, + "test_acc": 0.8770053475935828, + "test_auc": 0.9347215358992661, + "val_auc": 0.9385858585858586, + "k": 2, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1274 + ] + }, + { + "test_f1": 0.881173056835927, + "test_acc": 0.881283422459893, + "test_auc": 0.9434835923058621, + "val_auc": 0.9470707070707071, + "k": 5, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1274, + 1682, + 1305, + 930 + ] + }, + { + "test_f1": 0.8983775882883429, + "test_acc": 0.8983957219251337, + "test_auc": 0.9529594245868335, + "val_auc": 0.9583838383838383, + "k": 10, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1274, + 1682, + 1305, + 930, + 3417, + 343, + 340, + 888, + 874 + ] + }, + { + "test_f1": 0.9080022381610273, + "test_acc": 0.9080213903743316, + "test_auc": 0.9565969362543237, + "val_auc": 0.9466666666666667, + "k": 16, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1274, + 1682, + 1305, + 930, + 3417, + 343, + 340, + 888, + 874, + 1895, + 1441, + 238, + 456, + 1744, + 2884 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e1151c08e557f4b7b3f6769e4d44cef098c89bd7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9457402444655052, + "test_acc": 0.9457471264367816, + "test_auc": 0.9846189424481845, + "val_auc": 0.9943181818181818, + "k": 1, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554 + ] + }, + { + "test_f1": 0.9576938876419444, + "test_acc": 0.9577011494252874, + "test_auc": 0.9920978712322095, + "val_auc": 0.9991883116883117, + "k": 2, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072 + ] + }, + { + "test_f1": 0.9613765180533485, + "test_acc": 0.9613793103448276, + "test_auc": 0.989743002191677, + "val_auc": 0.9991883116883117, + "k": 5, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 1101, + 4323, + 1065 + ] + }, + { + "test_f1": 0.9765517142245282, + "test_acc": 0.976551724137931, + "test_auc": 0.99499600898317, + "val_auc": 1.0, + "k": 10, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 1101, + 4323, + 1065, + 760, + 192, + 1682, + 1876, + 1622 + ] + }, + { + "test_f1": 0.9788503243482447, + "test_acc": 0.9788505747126437, + "test_auc": 0.9951017032848096, + "val_auc": 1.0, + "k": 16, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 1101, + 4323, + 1065, + 760, + 192, + 1682, + 1876, + 1622, + 763, + 1684, + 770, + 1872, + 393, + 652 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6a252cea1cdc8ec97af0864c75cc081db76b4769 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9675107501194458, + "test_acc": 0.9675107501194458, + "test_auc": 0.9880921726648662, + "val_auc": 0.9995709995709996, + "k": 1, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286 + ] + }, + { + "test_f1": 0.9574699655435616, + "test_acc": 0.9574773053033923, + "test_auc": 0.9880921726648662, + "val_auc": 0.9995709995709996, + "k": 2, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905 + ] + }, + { + "test_f1": 0.9827988665017662, + "test_acc": 0.9827998088867654, + "test_auc": 0.9972561137073783, + "val_auc": 1.0, + "k": 5, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581 + ] + }, + { + "test_f1": 0.9880549974637796, + "test_acc": 0.9880554228380315, + "test_auc": 0.9987974381872271, + "val_auc": 1.0, + "k": 10, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581, + 2874, + 948, + 304, + 1079, + 1682 + ] + }, + { + "test_f1": 0.9894887433037904, + "test_acc": 0.9894887720974678, + "test_auc": 0.9993726955464124, + "val_auc": 1.0, + "k": 16, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581, + 2874, + 948, + 304, + 1079, + 1682, + 1684, + 760, + 1156, + 239, + 88, + 3877 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9b8382c9be13aab70b2b45e604f89ef789a267f1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.992643566195959, + "test_acc": 0.9926436781609196, + "test_auc": 0.9983372172466043, + "val_auc": 1.0, + "k": 1, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559 + ] + }, + { + "test_f1": 0.9926436781609196, + "test_acc": 0.9926436781609196, + "test_auc": 0.9988542737702257, + "val_auc": 1.0, + "k": 2, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300 + ] + }, + { + "test_f1": 0.9921839047415093, + "test_acc": 0.992183908045977, + "test_auc": 0.9997319592510417, + "val_auc": 1.0, + "k": 5, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 2236, + 334 + ] + }, + { + "test_f1": 0.9954022969067715, + "test_acc": 0.9954022988505747, + "test_auc": 0.999863865739488, + "val_auc": 1.0, + "k": 10, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 2236, + 334, + 3825, + 982, + 1268, + 1158, + 913 + ] + }, + { + "test_f1": 0.993103419118721, + "test_acc": 0.993103448275862, + "test_auc": 0.9997167392716056, + "val_auc": 1.0, + "k": 16, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 2236, + 334, + 3825, + 982, + 1268, + 1158, + 913, + 1872, + 1592, + 1934, + 1769, + 896, + 1305 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f05eb81e20a87a23bff31b6223adca215660e5fa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9457456125644097, + "test_acc": 0.9457471264367816, + "test_auc": 0.9796982385410467, + "val_auc": 0.9598214285714286, + "k": 1, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095 + ] + }, + { + "test_f1": 0.9650533934875399, + "test_acc": 0.9650574712643678, + "test_auc": 0.9892166445695113, + "val_auc": 0.9760551948051948, + "k": 2, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023 + ] + }, + { + "test_f1": 0.9747108154063739, + "test_acc": 0.9747126436781609, + "test_auc": 0.9931882136479246, + "val_auc": 0.976461038961039, + "k": 5, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 1682 + ] + }, + { + "test_f1": 0.9806877691190784, + "test_acc": 0.9806896551724138, + "test_auc": 0.9960368864657179, + "val_auc": 0.9878246753246753, + "k": 10, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 1682, + 494, + 1301, + 211, + 1079, + 1701 + ] + }, + { + "test_f1": 0.9816084878327882, + "test_acc": 0.9816091954022989, + "test_auc": 0.9957561624005627, + "val_auc": 0.9866071428571428, + "k": 16, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 1682, + 494, + 1301, + 211, + 1079, + 1701, + 1193, + 586, + 2011, + 235, + 292, + 888 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..530fe4d6b5c1a3381bdf6e7fecb1a230b0e1a9d6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/69_blood-pressure_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9627522272400656, + "test_acc": 0.9627586206896551, + "test_auc": 0.991551220304129, + "val_auc": 0.9841720779220778, + "k": 1, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32 + ] + }, + { + "test_f1": 0.9820685485627189, + "test_acc": 0.9820689655172414, + "test_auc": 0.9969923629525408, + "val_auc": 0.9931006493506493, + "k": 2, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683 + ] + }, + { + "test_f1": 0.9972413781440628, + "test_acc": 0.9972413793103448, + "test_auc": 0.9997514070025435, + "val_auc": 0.9931006493506493, + "k": 5, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 1047, + 816 + ] + }, + { + "test_f1": 0.9972413793103448, + "test_acc": 0.9972413793103448, + "test_auc": 0.9987071473023431, + "val_auc": 0.9931006493506493, + "k": 10, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 1047, + 816, + 87, + 1458, + 1558, + 5094, + 495 + ] + }, + { + "test_f1": 0.9977011494252873, + "test_acc": 0.9977011494252873, + "test_auc": 0.9999095256777965, + "val_auc": 0.9931006493506493, + "k": 16, + "dataset": "69_blood-pressure", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32, + 2683, + 12798, + 1047, + 816, + 87, + 1458, + 1558, + 5094, + 495, + 1023, + 232, + 868, + 1627, + 1051, + 1682 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c6674171a0b9c331ab6806d973dff1a16d162771 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8239150189632679, + "test_acc": 0.8241509433962264, + "test_auc": 0.8697205882501855, + "val_auc": 0.9235694277711084, + "k": 1, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.8937573230539363, + "test_acc": 0.8938364779874214, + "test_auc": 0.9472324637305465, + "val_auc": 0.9915966386554622, + "k": 2, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028 + ] + }, + { + "test_f1": 0.9121815827932492, + "test_acc": 0.9122012578616352, + "test_auc": 0.963685484826422, + "val_auc": 0.9967987194877952, + "k": 5, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 778, + 3417, + 1839 + ] + }, + { + "test_f1": 0.9184691004351123, + "test_acc": 0.9184905660377358, + "test_auc": 0.9667276431614347, + "val_auc": 0.9971988795518207, + "k": 10, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 778, + 3417, + 1839, + 1305, + 4663, + 1602, + 988, + 1895 + ] + }, + { + "test_f1": 0.9187213976249557, + "test_acc": 0.9187421383647799, + "test_auc": 0.9666709365402278, + "val_auc": 0.9967987194877951, + "k": 16, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 778, + 3417, + 1839, + 1305, + 4663, + 1602, + 988, + 1895, + 1279, + 8038, + 1478, + 3301, + 5539, + 30649 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c6ca2f76dcfb4e70c48b8012c8cc3572ea438141 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999797066940852, + "val_auc": 1.0, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9999830889117376, + "val_auc": 1.0, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686 + ] + }, + { + "test_f1": 0.999080459770115, + "test_acc": 0.999080459770115, + "test_auc": 0.999993235564695, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 4787, + 547 + ] + }, + { + "test_f1": 0.999080459770115, + "test_acc": 0.999080459770115, + "test_auc": 0.9999720967043672, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 4787, + 547, + 1162, + 1840, + 4276, + 24163, + 3818 + ] + }, + { + "test_f1": 0.999080459770115, + "test_acc": 0.999080459770115, + "test_auc": 0.9999627956058228, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1379, + 1686, + 682, + 4787, + 547, + 1162, + 1840, + 4276, + 24163, + 3818, + 1362, + 104, + 119, + 6777, + 2709, + 1245 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..401afb695363eafca022f6ea95fafbdb90a9d396 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/71_social-media_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9816066681247114, + "test_acc": 0.9816091954022989, + "test_auc": 0.993981343687429, + "val_auc": 1.0, + "k": 1, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211 + ] + }, + { + "test_f1": 0.9862065116689178, + "test_acc": 0.9862068965517241, + "test_auc": 0.9952868797012826, + "val_auc": 1.0, + "k": 2, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586 + ] + }, + { + "test_f1": 0.9926436315094049, + "test_acc": 0.9926436781609196, + "test_auc": 0.996742078846258, + "val_auc": 1.0, + "k": 5, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 10, + 559, + 21724 + ] + }, + { + "test_f1": 0.9944827236320538, + "test_acc": 0.9944827586206897, + "test_auc": 0.9962626494940202, + "val_auc": 1.0, + "k": 10, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 10, + 559, + 21724, + 620, + 3223, + 495, + 878, + 1874 + ] + }, + { + "test_f1": 0.994942483833445, + "test_acc": 0.9949425287356322, + "test_auc": 0.9974219045944046, + "val_auc": 1.0, + "k": 16, + "dataset": "71_social-media", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2211, + 586, + 10, + 559, + 21724, + 620, + 3223, + 495, + 878, + 1874, + 1872, + 1682, + 760, + 1603, + 1023, + 239 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0c0574432b6bd0dead420b9c3408fc37fe5b03b3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/72_gene-expression_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8932479892405967, + "test_acc": 0.8933333333333333, + "test_auc": 0.9380047959846312, + "val_auc": 0.9241071428571428, + "k": 1, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979 + ] + }, + { + "test_f1": 0.9291913309494634, + "test_acc": 0.9291954022988506, + "test_auc": 0.9802191000595271, + "val_auc": 0.9853896103896104, + "k": 2, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 27839 + ] + }, + { + "test_f1": 0.9434476064073148, + "test_acc": 0.9434482758620689, + "test_auc": 0.9859105268142215, + "val_auc": 0.9963474025974026, + "k": 5, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 27839, + 495, + 1682, + 1708 + ] + }, + { + "test_f1": 0.9475847443418873, + "test_acc": 0.9475862068965517, + "test_auc": 0.9916408490719194, + "val_auc": 0.9987824675324676, + "k": 10, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 27839, + 495, + 1682, + 1708, + 1027, + 1631, + 117, + 1840, + 2024 + ] + }, + { + "test_f1": 0.964135611158452, + "test_acc": 0.9641379310344828, + "test_auc": 0.9949359746198387, + "val_auc": 0.9983766233766234, + "k": 16, + "dataset": "72_gene-expression", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6979, + 27839, + 495, + 1682, + 1708, + 1027, + 1631, + 117, + 1840, + 2024, + 1034, + 2616, + 1841, + 27304, + 532, + 1583 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bd77cb38516511f27bc6bee19933c3db57674bc0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9019630233526316, + "test_acc": 0.9020689655172414, + "test_auc": 0.947487688727745, + "val_auc": 0.9472402597402597, + "k": 1, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961 + ] + }, + { + "test_f1": 0.9163164260363882, + "test_acc": 0.9163218390804597, + "test_auc": 0.9680481898371125, + "val_auc": 0.9683441558441559, + "k": 2, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182 + ] + }, + { + "test_f1": 0.9328731375233197, + "test_acc": 0.9328735632183908, + "test_auc": 0.9738820925104172, + "val_auc": 0.9768668831168832, + "k": 5, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182, + 248, + 732, + 1264 + ] + }, + { + "test_f1": 0.9599992389851021, + "test_acc": 0.96, + "test_auc": 0.9858179386059852, + "val_auc": 0.9943181818181818, + "k": 10, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182, + 248, + 732, + 1264, + 8734, + 1441, + 1482, + 1536, + 859 + ] + }, + { + "test_f1": 0.9655171539078867, + "test_acc": 0.9655172413793104, + "test_auc": 0.9909212822663565, + "val_auc": 0.9987824675324676, + "k": 16, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 7182, + 248, + 732, + 1264, + 8734, + 1441, + 1482, + 1536, + 859, + 1682, + 653, + 306, + 1960, + 1856, + 23259 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d93fddffdd606c895d4fc34101cdc274edb017cc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/74_magnetic-field_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9530771434907823, + "test_acc": 0.953103448275862, + "test_auc": 0.9813462240922126, + "val_auc": 0.9890422077922078, + "k": 1, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786 + ] + }, + { + "test_f1": 0.9544541641773988, + "test_acc": 0.9544827586206897, + "test_auc": 0.9813462240922126, + "val_auc": 0.9890422077922078, + "k": 2, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 3355 + ] + }, + { + "test_f1": 0.9558585231985594, + "test_acc": 0.9558620689655173, + "test_auc": 0.9894601642404892, + "val_auc": 0.9910714285714286, + "k": 5, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 3355, + 1558, + 3930, + 271 + ] + }, + { + "test_f1": 0.964124872203006, + "test_acc": 0.9641379310344828, + "test_auc": 0.9916366212998539, + "val_auc": 0.9935064935064934, + "k": 10, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 3355, + 1558, + 3930, + 271, + 292, + 1910, + 653, + 760, + 276 + ] + }, + { + "test_f1": 0.9664239634755235, + "test_acc": 0.9664367816091954, + "test_auc": 0.9915588302938472, + "val_auc": 0.9959415584415584, + "k": 16, + "dataset": "74_magnetic-field", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 26786, + 3355, + 1558, + 3930, + 271, + 292, + 1910, + 653, + 760, + 276, + 67, + 327, + 1051, + 532, + 1593, + 1596 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a11f1af7cdf3bb0260ab60de73a5ea9985629962 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/75_cell-lines_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.967815030599757, + "test_acc": 0.967816091954023, + "test_auc": 0.9878455780886412, + "val_auc": 0.9815340909090908, + "k": 1, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831 + ] + }, + { + "test_f1": 0.9655170955936044, + "test_acc": 0.9655172413793104, + "test_auc": 0.9904012663022891, + "val_auc": 0.9890422077922078, + "k": 2, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856 + ] + }, + { + "test_f1": 0.9742523837165158, + "test_acc": 0.9742528735632184, + "test_auc": 0.9925236078792141, + "val_auc": 0.9914772727272727, + "k": 5, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 7739, + 1168, + 1167 + ] + }, + { + "test_f1": 0.9783895620516864, + "test_acc": 0.9783908045977011, + "test_auc": 0.9941935778451216, + "val_auc": 0.9926948051948052, + "k": 10, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 7739, + 1168, + 1167, + 1037, + 822, + 517, + 1449, + 1536 + ] + }, + { + "test_f1": 0.9724100021870672, + "test_acc": 0.9724137931034482, + "test_auc": 0.9946704705341199, + "val_auc": 0.9967532467532467, + "k": 16, + "dataset": "75_cell-lines", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11831, + 1856, + 7739, + 1168, + 1167, + 1037, + 822, + 517, + 1449, + 1536, + 586, + 1252, + 581, + 18934, + 1366, + 1101 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2efd6535233f59f42111d54af35e7b54bc7ed796 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/76_trial-court_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.963678145563485, + "test_acc": 0.9636781609195403, + "test_auc": 0.9857718558904702, + "val_auc": 0.9995941558441559, + "k": 1, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840 + ] + }, + { + "test_f1": 0.9645939140396502, + "test_acc": 0.9645977011494253, + "test_auc": 0.9939991003301045, + "val_auc": 1.0, + "k": 2, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495 + ] + }, + { + "test_f1": 0.9765509508696238, + "test_acc": 0.976551724137931, + "test_auc": 0.9974565723253422, + "val_auc": 0.9995941558441559, + "k": 5, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495, + 1765, + 1763, + 789 + ] + }, + { + "test_f1": 0.9788488756993098, + "test_acc": 0.9788505747126437, + "test_auc": 0.9973534146869419, + "val_auc": 0.9995941558441559, + "k": 10, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495, + 1765, + 1763, + 789, + 1682, + 901, + 1797, + 434, + 1051 + ] + }, + { + "test_f1": 0.9793103448275862, + "test_acc": 0.9793103448275862, + "test_auc": 0.9975470466475458, + "val_auc": 1.0, + "k": 16, + "dataset": "76_trial-court", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23840, + 5495, + 1765, + 1763, + 789, + 1682, + 901, + 1797, + 434, + 1051, + 1815, + 20223, + 677, + 1310, + 7461, + 2120 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fdcb9a2fbf46b22e5ebe8a2865f678b9b4662170 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9636710190010682, + "test_acc": 0.9636781609195403, + "test_auc": 0.977867613236647, + "val_auc": 0.9756493506493507, + "k": 1, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879 + ] + }, + { + "test_f1": 0.9696550954390544, + "test_acc": 0.9696551724137931, + "test_auc": 0.9915199347908438, + "val_auc": 0.9837662337662337, + "k": 2, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293 + ] + }, + { + "test_f1": 0.9903443377213826, + "test_acc": 0.9903448275862069, + "test_auc": 0.9966050990313328, + "val_auc": 1.0, + "k": 5, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 4455, + 277, + 1051 + ] + }, + { + "test_f1": 0.9857467287226145, + "test_acc": 0.9857471264367816, + "test_auc": 0.9941352345906165, + "val_auc": 1.0, + "k": 10, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 4455, + 277, + 1051, + 571, + 7367, + 1501, + 30210, + 1454 + ] + }, + { + "test_f1": 0.9885056110583972, + "test_acc": 0.9885057471264368, + "test_auc": 0.9981550002705774, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1879, + 1293, + 4455, + 277, + 1051, + 571, + 7367, + 1501, + 30210, + 1454, + 1484, + 1283, + 30, + 6859, + 1682, + 613 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2ac473fd6bd19f8779423f10c2f7dd3c9329d9f2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9825278492159298, + "test_acc": 0.9825287356321839, + "test_auc": 0.9913609705611776, + "val_auc": 0.9886363636363636, + "k": 1, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439 + ] + }, + { + "test_f1": 0.9802290073941264, + "test_acc": 0.9802298850574712, + "test_auc": 0.9909559499972942, + "val_auc": 0.9995941558441559, + "k": 2, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161 + ] + }, + { + "test_f1": 0.9816080057222831, + "test_acc": 0.9816091954022989, + "test_auc": 0.99001907570756, + "val_auc": 0.9991883116883117, + "k": 5, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161, + 26232, + 1121, + 1795 + ] + }, + { + "test_f1": 0.9806884060083972, + "test_acc": 0.9806896551724138, + "test_auc": 0.9904672195465123, + "val_auc": 0.9991883116883117, + "k": 10, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161, + 26232, + 1121, + 1795, + 5535, + 5907, + 802, + 493, + 5797 + ] + }, + { + "test_f1": 0.9848269126513065, + "test_acc": 0.9848275862068966, + "test_auc": 0.9980188660100655, + "val_auc": 0.9987824675324676, + "k": 16, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9439, + 6161, + 26232, + 1121, + 1795, + 5535, + 5907, + 802, + 493, + 5797, + 1441, + 2020, + 2530, + 1790, + 345, + 495 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e33bea183058fb02de13638098e82e357075fee7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9576973936480825, + "test_acc": 0.9577011494252874, + "test_auc": 0.9893493966123708, + "val_auc": 0.9922889610389609, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272 + ] + }, + { + "test_f1": 0.9572379443202836, + "test_acc": 0.9572413793103448, + "test_auc": 0.9889777754478056, + "val_auc": 0.9931006493506493, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866 + ] + }, + { + "test_f1": 0.9622955030341607, + "test_acc": 0.9622988505747126, + "test_auc": 0.99342835110125, + "val_auc": 0.9971590909090909, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 165, + 5703 + ] + }, + { + "test_f1": 0.9825285805151106, + "test_acc": 0.9825287356321839, + "test_auc": 0.9976603509389036, + "val_auc": 1.0, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 165, + 5703, + 1072, + 918, + 1682, + 508, + 495 + ] + }, + { + "test_f1": 0.9862065758160522, + "test_acc": 0.9862068965517241, + "test_auc": 0.997511533362195, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 3198, + 165, + 5703, + 1072, + 918, + 1682, + 508, + 495, + 1726, + 1079, + 1645, + 1358, + 1281, + 1715 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e3427905ed1af99cb626877571dd6a354af7c69d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/7_hist_fig_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5517352586322463, + "test_acc": 0.5733333333333334, + "test_auc": 0.5863327929327348, + "val_auc": 0.6122448979591837, + "k": 1, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475 + ] + }, + { + "test_f1": 0.5979584938226857, + "test_acc": 0.600754716981132, + "test_auc": 0.6145466153741777, + "val_auc": 0.6276510604241696, + "k": 2, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475, + 1556 + ] + }, + { + "test_f1": 0.6161828416626474, + "test_acc": 0.6163522012578616, + "test_auc": 0.6531961016223157, + "val_auc": 0.6882753101240496, + "k": 5, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475, + 1556, + 776, + 1279, + 1132 + ] + }, + { + "test_f1": 0.6149732310183584, + "test_acc": 0.6150943396226415, + "test_auc": 0.6608693428816482, + "val_auc": 0.6850740296118447, + "k": 10, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475, + 1556, + 776, + 1279, + 1132, + 17275, + 1815, + 238, + 7893, + 581 + ] + }, + { + "test_f1": 0.6230738040616313, + "test_acc": 0.6231446540880503, + "test_auc": 0.6741074276560216, + "val_auc": 0.6942777110844338, + "k": 16, + "dataset": "7_hist_fig_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6475, + 1556, + 776, + 1279, + 1132, + 17275, + 1815, + 238, + 7893, + 581, + 1895, + 7740, + 6758, + 2126, + 1866, + 925 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4534bbba7e8a55aed63c6d0351c6c82bb837e4fa --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9595148120161089, + "test_acc": 0.9595402298850575, + "test_auc": 0.9762872720385303, + "val_auc": 0.9768668831168832, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963 + ] + }, + { + "test_f1": 0.958592749681999, + "test_acc": 0.9586206896551724, + "test_auc": 0.9790082661399427, + "val_auc": 0.9829545454545455, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217 + ] + }, + { + "test_f1": 0.9747110078560357, + "test_acc": 0.9747126436781609, + "test_auc": 0.9933640889658532, + "val_auc": 0.9922889610389611, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217, + 3772, + 352, + 1942 + ] + }, + { + "test_f1": 0.9765512780257494, + "test_acc": 0.976551724137931, + "test_auc": 0.9941166323935278, + "val_auc": 0.9967532467532467, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217, + 3772, + 352, + 1942, + 1682, + 1261, + 1898, + 111, + 760 + ] + }, + { + "test_f1": 0.9783899731987808, + "test_acc": 0.9783908045977011, + "test_auc": 0.9944083486660533, + "val_auc": 0.9939123376623377, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10963, + 217, + 3772, + 352, + 1942, + 1682, + 1261, + 1898, + 111, + 760, + 1726, + 1441, + 581, + 1221, + 640, + 393 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e48e4820a56dc784f534550160eb7fb5b8285172 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/81_public-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9852871323919179, + "test_acc": 0.9852873563218391, + "test_auc": 0.9962237539910168, + "val_auc": 1.0, + "k": 1, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833 + ] + }, + { + "test_f1": 0.9857468552705536, + "test_acc": 0.9857471264367816, + "test_auc": 0.9982357507170301, + "val_auc": 1.0, + "k": 2, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675 + ] + }, + { + "test_f1": 0.9894252873563218, + "test_acc": 0.9894252873563218, + "test_auc": 0.9990775001352887, + "val_auc": 1.0, + "k": 5, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 829, + 1682 + ] + }, + { + "test_f1": 0.9908045977011494, + "test_acc": 0.9908045977011494, + "test_auc": 0.999381899724011, + "val_auc": 1.0, + "k": 10, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 829, + 1682, + 596, + 1588, + 913, + 361, + 859 + ] + }, + { + "test_f1": 0.9908045938135429, + "test_acc": 0.9908045977011494, + "test_auc": 0.9996178094052709, + "val_auc": 1.0, + "k": 16, + "dataset": "81_public-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12833, + 1675, + 165, + 829, + 1682, + 596, + 1588, + 913, + 361, + 859, + 491, + 1181, + 1281, + 1079, + 1515, + 1631 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..61599f159a1c9509d546834dfbb4a840b19e1172 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8532773198450958, + "test_acc": 0.8533333333333334, + "test_auc": 0.9004313173061312, + "val_auc": 0.9200487012987013, + "k": 1, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020 + ] + }, + { + "test_f1": 0.9131026767746447, + "test_acc": 0.9131034482758621, + "test_auc": 0.9643395881811786, + "val_auc": 0.9853896103896104, + "k": 2, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760 + ] + }, + { + "test_f1": 0.9438740905158824, + "test_acc": 0.9439080459770115, + "test_auc": 0.9848620393419558, + "val_auc": 0.9845779220779222, + "k": 5, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1682, + 1482, + 913 + ] + }, + { + "test_f1": 0.9512411152963881, + "test_acc": 0.9512643678160919, + "test_auc": 0.9899108447426809, + "val_auc": 0.9918831168831168, + "k": 10, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1682, + 1482, + 913, + 49, + 18571, + 2023, + 170, + 1960 + ] + }, + { + "test_f1": 0.9572329717354346, + "test_acc": 0.9572413793103448, + "test_auc": 0.9938655027328319, + "val_auc": 0.9987824675324676, + "k": 16, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1682, + 1482, + 913, + 49, + 18571, + 2023, + 170, + 1960, + 2817, + 1264, + 364, + 1120, + 3842, + 586 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..70529579d9e6c5332f18b536379e6cd1fb961e25 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/83_third-party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9558610426113668, + "test_acc": 0.9558620689655173, + "test_auc": 0.9675247916553926, + "val_auc": 0.9551542207792209, + "k": 1, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123 + ] + }, + { + "test_f1": 0.9590752691845337, + "test_acc": 0.9590804597701149, + "test_auc": 0.9766394454515938, + "val_auc": 0.9630681818181818, + "k": 2, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491 + ] + }, + { + "test_f1": 0.9687273027756429, + "test_acc": 0.968735632183908, + "test_auc": 0.982936711943287, + "val_auc": 0.9719967532467533, + "k": 5, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 1763, + 1281, + 1682 + ] + }, + { + "test_f1": 0.9678109478589979, + "test_acc": 0.967816091954023, + "test_auc": 0.9914480626657287, + "val_auc": 0.9849837662337663, + "k": 10, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 1763, + 1281, + 1682, + 679, + 1111, + 167, + 866, + 1486 + ] + }, + { + "test_f1": 0.9691908304043998, + "test_acc": 0.9691954022988506, + "test_auc": 0.9939931814492127, + "val_auc": 0.9902597402597402, + "k": 16, + "dataset": "83_third-party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17123, + 1491, + 1763, + 1281, + 1682, + 679, + 1111, + 167, + 866, + 1486, + 204, + 7316, + 523, + 969, + 687, + 1310 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..59e4a6e39707697602b575700052ba4c8b37829d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/84_clinical-trials_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8481697566946521, + "test_acc": 0.8482758620689655, + "test_auc": 0.9064275664267547, + "val_auc": 0.906047077922078, + "k": 1, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 352 + ] + }, + { + "test_f1": 0.8677168022718386, + "test_acc": 0.8680459770114942, + "test_auc": 0.9210831382921154, + "val_auc": 0.9257305194805194, + "k": 2, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 352, + 7967 + ] + }, + { + "test_f1": 0.9282471649877408, + "test_acc": 0.9282758620689655, + "test_auc": 0.9735992545592294, + "val_auc": 0.9866071428571429, + "k": 5, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 352, + 7967, + 495, + 1682, + 1242 + ] + }, + { + "test_f1": 0.9489618274565002, + "test_acc": 0.9489655172413793, + "test_auc": 0.9869437943070513, + "val_auc": 0.9955357142857143, + "k": 10, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 352, + 7967, + 495, + 1682, + 1242, + 1762, + 178, + 19873, + 859, + 108 + ] + }, + { + "test_f1": 0.9503356939767229, + "test_acc": 0.9503448275862069, + "test_auc": 0.9885275177228204, + "val_auc": 0.9983766233766234, + "k": 16, + "dataset": "84_clinical-trials", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 352, + 7967, + 495, + 1682, + 1242, + 1762, + 178, + 19873, + 859, + 108, + 960, + 1168, + 12346, + 1037, + 1797, + 987 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a52b90b103e51e92aa8c3a183ab55e95f8d38e0e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/85_mental-health_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8963723198194128, + "test_acc": 0.896551724137931, + "test_auc": 0.9437503382217652, + "val_auc": 0.9642857142857142, + "k": 1, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335 + ] + }, + { + "test_f1": 0.9203156525791584, + "test_acc": 0.9204597701149425, + "test_auc": 0.971087112397857, + "val_auc": 0.9760551948051949, + "k": 2, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1682 + ] + }, + { + "test_f1": 0.9609137400652832, + "test_acc": 0.960919540229885, + "test_auc": 0.9914827303966666, + "val_auc": 0.9971590909090908, + "k": 5, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1682, + 1443, + 235, + 1079 + ] + }, + { + "test_f1": 0.9793103448275862, + "test_acc": 0.9793103448275862, + "test_auc": 0.9967124844417988, + "val_auc": 0.9983766233766234, + "k": 10, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1682, + 1443, + 235, + 1079, + 1281, + 1515, + 21, + 760, + 165 + ] + }, + { + "test_f1": 0.980689532712188, + "test_acc": 0.9806896551724138, + "test_auc": 0.997066771740895, + "val_auc": 1.0, + "k": 16, + "dataset": "85_mental-health", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4335, + 1682, + 1443, + 235, + 1079, + 1281, + 1515, + 21, + 760, + 165, + 482, + 878, + 1928, + 559, + 335, + 571 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c5133dc68f00028f4d924318b5d637dbaf603744 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/87_glue_cola_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6893505002315113, + "test_acc": 0.6950943396226416, + "test_auc": 0.7091409807612661, + "val_auc": 0.7154861944777912, + "k": 1, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313 + ] + }, + { + "test_f1": 0.6893505002315113, + "test_acc": 0.6950943396226416, + "test_auc": 0.7091409807612661, + "val_auc": 0.7154861944777912, + "k": 2, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 19000 + ] + }, + { + "test_f1": 0.6972034595413633, + "test_acc": 0.7001257861635221, + "test_auc": 0.7661483495841683, + "val_auc": 0.733093237294918, + "k": 5, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 19000, + 433, + 6718, + 25757 + ] + }, + { + "test_f1": 0.6992487691242028, + "test_acc": 0.700377358490566, + "test_auc": 0.7724350633240814, + "val_auc": 0.7587034813925569, + "k": 10, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 19000, + 433, + 6718, + 25757, + 1105, + 1928, + 2424, + 836, + 4333 + ] + }, + { + "test_f1": 0.7124180219206508, + "test_acc": 0.7137106918238993, + "test_auc": 0.7782258472829934, + "val_auc": 0.7671068427370948, + "k": 16, + "dataset": "87_glue_cola", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1313, + 19000, + 433, + 6718, + 25757, + 1105, + 1928, + 2424, + 836, + 4333, + 121, + 24630, + 1067, + 26569, + 1019, + 1532 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..15a2d2e5cac06c638f35b277a0c7775c5bb641e1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6991836418288707, + "test_acc": 0.7006603081438004, + "test_auc": 0.759787013233084, + "val_auc": 0.7920673076923077, + "k": 1, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394 + ] + }, + { + "test_f1": 0.6947536731725328, + "test_acc": 0.6955245781364637, + "test_auc": 0.7584434654919237, + "val_auc": 0.8036858974358975, + "k": 2, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 5133 + ] + }, + { + "test_f1": 0.6963119209717638, + "test_acc": 0.6969919295671313, + "test_auc": 0.7589419130914086, + "val_auc": 0.8052884615384616, + "k": 5, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 5133, + 2041, + 584, + 1928 + ] + }, + { + "test_f1": 0.7003825934291014, + "test_acc": 0.7006603081438004, + "test_auc": 0.7629800922397199, + "val_auc": 0.8263221153846154, + "k": 10, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 5133, + 2041, + 584, + 1928, + 28397, + 1551, + 1257, + 1737, + 28053 + ] + }, + { + "test_f1": 0.6946064632030183, + "test_acc": 0.6947909024211298, + "test_auc": 0.7665348956382068, + "val_auc": 0.8353365384615385, + "k": 16, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 5133, + 2041, + 584, + 1928, + 28397, + 1551, + 1257, + 1737, + 28053, + 7649, + 3632, + 31, + 1697, + 2367, + 9221 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..33604198bdf45473386079242409af778293ce60 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6469452188301973, + "test_acc": 0.6470440251572327, + "test_auc": 0.6799351721805417, + "val_auc": 0.6826730692276911, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520 + ] + }, + { + "test_f1": 0.695311086319976, + "test_acc": 0.6953459119496855, + "test_auc": 0.7658165905346523, + "val_auc": 0.7338935574229692, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1011 + ] + }, + { + "test_f1": 0.7373581581210553, + "test_acc": 0.7373584905660377, + "test_auc": 0.8213227528229263, + "val_auc": 0.7763105242096839, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1011, + 1927, + 1848, + 24630 + ] + }, + { + "test_f1": 0.7625051716843877, + "test_acc": 0.7625157232704403, + "test_auc": 0.8416854929273679, + "val_auc": 0.803921568627451, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1011, + 1927, + 1848, + 24630, + 1036, + 4674, + 1928, + 3089, + 7614 + ] + }, + { + "test_f1": 0.7640228274217502, + "test_acc": 0.7640251572327044, + "test_auc": 0.8462389839793669, + "val_auc": 0.8095238095238094, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1011, + 1927, + 1848, + 24630, + 1036, + 4674, + 1928, + 3089, + 7614, + 1520, + 5653, + 635, + 6417, + 1470, + 1500 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1004a807094474bbe87019baa64fbdd7ddadf3f2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7649122713753604, + "test_acc": 0.7650314465408805, + "test_auc": 0.7892764235134005, + "val_auc": 0.7378951580632253, + "k": 1, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16477 + ] + }, + { + "test_f1": 0.7773453854896029, + "test_acc": 0.7773584905660378, + "test_auc": 0.8516754781330155, + "val_auc": 0.8257302921168468, + "k": 2, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16477, + 1503 + ] + }, + { + "test_f1": 0.7838993710691824, + "test_acc": 0.7838993710691824, + "test_auc": 0.8560537862302147, + "val_auc": 0.8255302120848339, + "k": 5, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16477, + 1503, + 1643, + 2472, + 1928 + ] + }, + { + "test_f1": 0.7858967462130348, + "test_acc": 0.7859119496855346, + "test_auc": 0.8626332732175641, + "val_auc": 0.8543417366946778, + "k": 10, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16477, + 1503, + 1643, + 2472, + 1928, + 1848, + 7303, + 5891, + 7032, + 7615 + ] + }, + { + "test_f1": 0.7850601049574645, + "test_acc": 0.7851572327044025, + "test_auc": 0.858508119679324, + "val_auc": 0.8771508603441376, + "k": 16, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16477, + 1503, + 1643, + 2472, + 1928, + 1848, + 7303, + 5891, + 7032, + 7615, + 17225, + 18220, + 7708, + 6327, + 2471, + 12335 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0f8b188bf0a0d9a936099ae7803722999ed04f7d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/92_glue_sst2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8115668173726301, + "test_acc": 0.8115723270440252, + "test_auc": 0.8596293412209544, + "val_auc": 0.8781512605042017, + "k": 1, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.855232066258414, + "test_acc": 0.8558490566037736, + "test_auc": 0.9204139026408071, + "val_auc": 0.9507803121248499, + "k": 2, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8665753806794965, + "test_acc": 0.8669182389937107, + "test_auc": 0.9315342229522074, + "val_auc": 0.9531812725090036, + "k": 5, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 1132, + 7284 + ] + }, + { + "test_f1": 0.8793914713004995, + "test_acc": 0.8794968553459119, + "test_auc": 0.9437627273454516, + "val_auc": 0.960984393757503, + "k": 10, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 1132, + 7284, + 100, + 794, + 6745, + 1874, + 505 + ] + }, + { + "test_f1": 0.8727562079938187, + "test_acc": 0.8729559748427673, + "test_auc": 0.9439460112461381, + "val_auc": 0.9637855142056824, + "k": 16, + "dataset": "92_glue_sst2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 1438, + 1132, + 7284, + 100, + 794, + 6745, + 1874, + 505, + 1335, + 5660, + 1928, + 1960, + 1417, + 4675 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c998f630e801a5fe5701fae98cc7e620afb784c8 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8840018444686781, + "test_acc": 0.8850314465408805, + "test_auc": 0.8924974102288619, + "val_auc": 0.879951980792317, + "k": 1, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480 + ] + }, + { + "test_f1": 0.8847131006604747, + "test_acc": 0.8857861635220126, + "test_auc": 0.9469099448224323, + "val_auc": 0.9437775110044018, + "k": 2, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433 + ] + }, + { + "test_f1": 0.9203407084658264, + "test_acc": 0.920503144654088, + "test_auc": 0.9655147796694613, + "val_auc": 0.9645858343337336, + "k": 5, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 1335 + ] + }, + { + "test_f1": 0.9297939655111941, + "test_acc": 0.929811320754717, + "test_auc": 0.9803495102471901, + "val_auc": 0.9899959983993598, + "k": 10, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 1335, + 7666, + 3481, + 6553, + 1927, + 1928 + ] + }, + { + "test_f1": 0.9350552676976203, + "test_acc": 0.9350943396226415, + "test_auc": 0.9832293965099099, + "val_auc": 0.9963985594237695, + "k": 16, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 1335, + 7666, + 3481, + 6553, + 1927, + 1928, + 1064, + 4110, + 988, + 993, + 15841, + 1500 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d78d2b8ba8c3d629d7ca3a71ea81e459a52e2208 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7572974721150446, + "test_acc": 0.7676767676767676, + "test_auc": 0.7653061224489797, + "val_auc": 0.7411858974358975, + "k": 1, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629 + ] + }, + { + "test_f1": 0.7373737373737373, + "test_acc": 0.7373737373737373, + "test_auc": 0.8193877551020409, + "val_auc": 0.8183092948717948, + "k": 2, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625 + ] + }, + { + "test_f1": 0.7373737373737373, + "test_acc": 0.7373737373737373, + "test_auc": 0.823265306122449, + "val_auc": 0.8485576923076923, + "k": 5, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 794, + 530 + ] + }, + { + "test_f1": 0.767629340253613, + "test_acc": 0.7676767676767676, + "test_auc": 0.8689795918367347, + "val_auc": 0.8998397435897435, + "k": 10, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 794, + 530, + 567, + 2046, + 1345, + 7669, + 777 + ] + }, + { + "test_f1": 0.8080416289051587, + "test_acc": 0.8080808080808081, + "test_auc": 0.8893877551020408, + "val_auc": 0.8898237179487181, + "k": 16, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1629, + 1625, + 1874, + 794, + 530, + 567, + 2046, + 1345, + 7669, + 777, + 765, + 1182, + 1758, + 1335, + 680, + 920 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2596e2941fd1a61df280e0197e68220a792119d9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.923209519594522, + "test_acc": 0.9232409381663113, + "test_auc": 0.9407255864702673, + "val_auc": 0.9643999999999999, + "k": 1, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655 + ] + }, + { + "test_f1": 0.910323016187499, + "test_acc": 0.9104477611940298, + "test_auc": 0.9648208765230042, + "val_auc": 0.9764, + "k": 2, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188 + ] + }, + { + "test_f1": 0.9466849160876467, + "test_acc": 0.9466950959488273, + "test_auc": 0.9857701400254592, + "val_auc": 0.9952, + "k": 5, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188, + 1417, + 12485, + 1927 + ] + }, + { + "test_f1": 0.9637523356595585, + "test_acc": 0.9637526652452025, + "test_auc": 0.9899436261138389, + "val_auc": 0.9952000000000001, + "k": 10, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188, + 1417, + 12485, + 1927, + 727, + 1625, + 330, + 4080, + 5406 + ] + }, + { + "test_f1": 0.9658817591285327, + "test_acc": 0.9658848614072495, + "test_auc": 0.9923258774322604, + "val_auc": 0.994, + "k": 16, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 28188, + 1417, + 12485, + 1927, + 727, + 1625, + 330, + 4080, + 5406, + 1472, + 984, + 1118, + 567, + 2187, + 6096 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/scr/saebench_scr_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/scr/saebench_scr_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..86960ac43a5d6a8cc47a1d3e0fb55471e1038f7b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/scr/saebench_scr_custom_sae_eval_results.json @@ -0,0 +1,326 @@ +{ + "eval_type_id": "scr", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": true, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "fd430e96-ad83-4e95-83c3-201921eb8e00", + "datetime_epoch_millis": 1776879710777, + "eval_result_metrics": { + "scr_metrics": { + "scr_dir1_threshold_2": 0.25142477242713573, + "scr_metric_threshold_2": 0.17821499949985128, + "scr_dir2_threshold_2": 0.1709856871482001, + "scr_dir1_threshold_5": 0.10963573574984177, + "scr_metric_threshold_5": 0.2659162012413723, + "scr_dir2_threshold_5": 0.2521337401628084, + "scr_dir1_threshold_10": 0.15397685795196597, + "scr_metric_threshold_10": 0.3114682599546464, + "scr_dir2_threshold_10": 0.2894830915889919, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results", + "scr_dir1_threshold_2": 0.5128207283533913, + "scr_metric_threshold_2": 0.06476677137111979, + "scr_dir2_threshold_2": 0.06476677137111979, + "scr_dir1_threshold_5": 0.5512825313325533, + "scr_metric_threshold_5": 0.21761653710635845, + "scr_dir2_threshold_5": 0.21761653710635845, + "scr_dir1_threshold_10": 0.5512825313325533, + "scr_metric_threshold_10": 0.222797971465746, + "scr_dir2_threshold_10": 0.222797971465746, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results", + "scr_dir1_threshold_2": 0.22727291199782534, + "scr_metric_threshold_2": 0.29969413887699026, + "scr_dir2_threshold_2": 0.29969413887699026, + "scr_dir1_threshold_5": -0.09848491348071632, + "scr_metric_threshold_5": 0.3333333940923963, + "scr_dir2_threshold_5": 0.3333333940923963, + "scr_dir1_threshold_10": -0.09848491348071632, + "scr_metric_threshold_10": 0.4189602909143935, + "scr_dir2_threshold_10": 0.4189602909143935, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results", + "scr_dir1_threshold_2": 0.3749997671693945, + "scr_metric_threshold_2": 0.05276388042950681, + "scr_dir2_threshold_2": 0.05276388042950681, + "scr_dir1_threshold_5": -0.18750034924590825, + "scr_metric_threshold_5": 0.13065336115482035, + "scr_dir2_threshold_5": 0.13065336115482035, + "scr_dir1_threshold_10": 0.03125052386886235, + "scr_metric_threshold_10": 0.1482412214577287, + "scr_dir2_threshold_10": 0.1482412214577287, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results", + "scr_dir1_threshold_2": 0.17449641072003616, + "scr_metric_threshold_2": 0.012738878893616343, + "scr_dir2_threshold_2": 0.012738878893616343, + "scr_dir1_threshold_5": -0.4496646738445599, + "scr_metric_threshold_5": 0.006369534358659739, + "scr_dir2_threshold_5": 0.006369534358659739, + "scr_dir1_threshold_10": -0.5973157906316693, + "scr_metric_threshold_10": -0.08598729016413297, + "scr_dir2_threshold_10": -0.08598729016413297, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results", + "scr_dir1_threshold_2": 0.03684225881529179, + "scr_metric_threshold_2": 0.48015863348328314, + "scr_dir2_threshold_2": 0.48015863348328314, + "scr_dir1_threshold_5": 0.057894933322792726, + "scr_metric_threshold_5": 0.5833333136228053, + "scr_dir2_threshold_5": 0.5833333136228053, + "scr_dir1_threshold_10": 0.08421069803000372, + "scr_metric_threshold_10": 0.6190476077844602, + "scr_dir2_threshold_10": 0.6190476077844602, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results", + "scr_dir1_threshold_2": 0.2896173902280937, + "scr_metric_threshold_2": 0.11068695256721958, + "scr_dir2_threshold_2": 0.11068695256721958, + "scr_dir1_threshold_5": 0.3989072479911928, + "scr_metric_threshold_5": 0.1603053122520976, + "scr_dir2_threshold_5": 0.1603053122520976, + "scr_dir1_threshold_10": 0.502732368584764, + "scr_metric_threshold_10": 0.2519083153249215, + "scr_dir2_threshold_10": 0.2519083153249215, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results", + "scr_dir1_threshold_2": 0.2651161514146259, + "scr_metric_threshold_2": 0.274678179658647, + "scr_dir2_threshold_2": 0.274678179658647, + "scr_dir1_threshold_5": 0.3767442672816017, + "scr_metric_threshold_5": 0.46781131470206244, + "scr_dir2_threshold_5": 0.46781131470206244, + "scr_dir1_threshold_10": 0.4465116317753144, + "scr_metric_threshold_10": 0.6051501487174384, + "scr_dir2_threshold_10": 0.6051501487174384, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results", + "scr_dir1_threshold_2": 0.13023256071842734, + "scr_metric_threshold_2": 0.13023256071842734, + "scr_dir2_threshold_2": 0.07239806190521783, + "scr_dir1_threshold_5": 0.227906842641778, + "scr_metric_threshold_5": 0.227906842641778, + "scr_dir2_threshold_5": 0.11764715401326675, + "scr_dir1_threshold_10": 0.3116278141366155, + "scr_metric_threshold_10": 0.3116278141366155, + "scr_dir2_threshold_10": 0.13574646721137937, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_scr", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": null +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/tpp/saebench_tpp_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/tpp/saebench_tpp_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..8ef3bc646f545a2aa11f10781b670c38db453b3f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-170898-tokens-700000000/tpp/saebench_tpp_custom_sae_eval_results.json @@ -0,0 +1,297 @@ +{ + "eval_type_id": "tpp", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": false, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "2cf9cd7a-9f3b-4811-9852-aa3fc4118a3a", + "datetime_epoch_millis": 1776879883789, + "eval_result_metrics": { + "tpp_metrics": { + "tpp_threshold_2_total_metric": 0.04157500416040421, + "tpp_threshold_2_intended_diff_only": 0.052000010013580324, + "tpp_threshold_2_unintended_diff_only": 0.010425005853176118, + "tpp_threshold_5_total_metric": 0.08317500352859497, + "tpp_threshold_5_intended_diff_only": 0.10290000438690186, + "tpp_threshold_5_unintended_diff_only": 0.019725000858306887, + "tpp_threshold_10_total_metric": 0.13842501044273375, + "tpp_threshold_10_intended_diff_only": 0.18930001258850096, + "tpp_threshold_10_unintended_diff_only": 0.05087500214576721, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results", + "tpp_threshold_2_total_metric": 0.07130000293254853, + "tpp_threshold_2_intended_diff_only": 0.0910000205039978, + "tpp_threshold_2_unintended_diff_only": 0.01970001757144928, + "tpp_threshold_5_total_metric": 0.15030002295970918, + "tpp_threshold_5_intended_diff_only": 0.18640003204345704, + "tpp_threshold_5_unintended_diff_only": 0.036100009083747865, + "tpp_threshold_10_total_metric": 0.23570001721382142, + "tpp_threshold_10_intended_diff_only": 0.33000003099441527, + "tpp_threshold_10_unintended_diff_only": 0.09430001378059387, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results", + "tpp_threshold_2_total_metric": 0.011850005388259888, + "tpp_threshold_2_intended_diff_only": 0.012999999523162841, + "tpp_threshold_2_unintended_diff_only": 0.0011499941349029541, + "tpp_threshold_5_total_metric": 0.016049984097480773, + "tpp_threshold_5_intended_diff_only": 0.01939997673034668, + "tpp_threshold_5_unintended_diff_only": 0.003349992632865906, + "tpp_threshold_10_total_metric": 0.04115000367164612, + "tpp_threshold_10_intended_diff_only": 0.04859999418258667, + "tpp_threshold_10_unintended_diff_only": 0.007449990510940552, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_tpp", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1": { + "0": { + "tpp_threshold_2_total_metric": 0.1577499806880951, + "tpp_threshold_2_intended_diff_only": 0.2120000123977661, + "tpp_threshold_2_unintended_diff_only": 0.05425003170967102, + "tpp_threshold_5_total_metric": 0.18049997091293335, + "tpp_threshold_5_intended_diff_only": 0.2459999918937683, + "tpp_threshold_5_unintended_diff_only": 0.06550002098083496, + "tpp_threshold_10_total_metric": 0.21775001287460327, + "tpp_threshold_10_intended_diff_only": 0.33100003004074097, + "tpp_threshold_10_unintended_diff_only": 0.1132500171661377 + }, + "1": { + "tpp_threshold_2_total_metric": 0.09425000846385956, + "tpp_threshold_2_intended_diff_only": 0.12800002098083496, + "tpp_threshold_2_unintended_diff_only": 0.0337500125169754, + "tpp_threshold_5_total_metric": 0.11950001120567322, + "tpp_threshold_5_intended_diff_only": 0.16100001335144043, + "tpp_threshold_5_unintended_diff_only": 0.04150000214576721, + "tpp_threshold_10_total_metric": 0.22750000655651093, + "tpp_threshold_10_intended_diff_only": 0.2940000295639038, + "tpp_threshold_10_unintended_diff_only": 0.06650002300739288 + }, + "2": { + "tpp_threshold_2_total_metric": 0.021250009536743164, + "tpp_threshold_2_intended_diff_only": 0.022000014781951904, + "tpp_threshold_2_unintended_diff_only": 0.0007500052452087402, + "tpp_threshold_5_total_metric": 0.200250044465065, + "tpp_threshold_5_intended_diff_only": 0.23600006103515625, + "tpp_threshold_5_unintended_diff_only": 0.03575001657009125, + "tpp_threshold_10_total_metric": 0.20650005340576172, + "tpp_threshold_10_intended_diff_only": 0.29600006341934204, + "tpp_threshold_10_unintended_diff_only": 0.08950001001358032 + }, + "6": { + "tpp_threshold_2_total_metric": 0.07300001382827759, + "tpp_threshold_2_intended_diff_only": 0.08100003004074097, + "tpp_threshold_2_unintended_diff_only": 0.008000016212463379, + "tpp_threshold_5_total_metric": 0.11225003004074097, + "tpp_threshold_5_intended_diff_only": 0.12700003385543823, + "tpp_threshold_5_unintended_diff_only": 0.014750003814697266, + "tpp_threshold_10_total_metric": 0.3385000079870224, + "tpp_threshold_10_intended_diff_only": 0.4020000100135803, + "tpp_threshold_10_unintended_diff_only": 0.06350000202655792 + }, + "9": { + "tpp_threshold_2_total_metric": 0.010250002145767212, + "tpp_threshold_2_intended_diff_only": 0.012000024318695068, + "tpp_threshold_2_unintended_diff_only": 0.0017500221729278564, + "tpp_threshold_5_total_metric": 0.1390000581741333, + "tpp_threshold_5_intended_diff_only": 0.16200006008148193, + "tpp_threshold_5_unintended_diff_only": 0.023000001907348633, + "tpp_threshold_10_total_metric": 0.18825000524520874, + "tpp_threshold_10_intended_diff_only": 0.3270000219345093, + "tpp_threshold_10_unintended_diff_only": 0.13875001668930054 + } + }, + "canrager/amazon_reviews_mcauley_1and5": { + "1": { + "tpp_threshold_2_total_metric": 0.009499996900558472, + "tpp_threshold_2_intended_diff_only": 0.004999995231628418, + "tpp_threshold_2_unintended_diff_only": -0.004500001668930054, + "tpp_threshold_5_total_metric": 0.010999977588653564, + "tpp_threshold_5_intended_diff_only": 0.010999977588653564, + "tpp_threshold_5_unintended_diff_only": 0.0, + "tpp_threshold_10_total_metric": 0.01299998164176941, + "tpp_threshold_10_intended_diff_only": 0.019999980926513672, + "tpp_threshold_10_unintended_diff_only": 0.006999999284744263 + }, + "2": { + "tpp_threshold_2_total_metric": 0.005250036716461182, + "tpp_threshold_2_intended_diff_only": 0.012000024318695068, + "tpp_threshold_2_unintended_diff_only": 0.006749987602233887, + "tpp_threshold_5_total_metric": 0.006500020623207092, + "tpp_threshold_5_intended_diff_only": 0.013999998569488525, + "tpp_threshold_5_unintended_diff_only": 0.007499977946281433, + "tpp_threshold_10_total_metric": 0.05800005793571472, + "tpp_threshold_10_intended_diff_only": 0.07100003957748413, + "tpp_threshold_10_unintended_diff_only": 0.01299998164176941 + }, + "3": { + "tpp_threshold_2_total_metric": 0.021250009536743164, + "tpp_threshold_2_intended_diff_only": 0.018000006675720215, + "tpp_threshold_2_unintended_diff_only": -0.0032500028610229492, + "tpp_threshold_5_total_metric": 0.02175000309944153, + "tpp_threshold_5_intended_diff_only": 0.023000001907348633, + "tpp_threshold_5_unintended_diff_only": 0.0012499988079071045, + "tpp_threshold_10_total_metric": 0.018249988555908203, + "tpp_threshold_10_intended_diff_only": 0.02499997615814209, + "tpp_threshold_10_unintended_diff_only": 0.006749987602233887 + }, + "5": { + "tpp_threshold_2_total_metric": -0.0007500350475311279, + "tpp_threshold_2_intended_diff_only": 0.007999956607818604, + "tpp_threshold_2_unintended_diff_only": 0.008749991655349731, + "tpp_threshold_5_total_metric": 0.006499946117401123, + "tpp_threshold_5_intended_diff_only": 0.008999943733215332, + "tpp_threshold_5_unintended_diff_only": 0.002499997615814209, + "tpp_threshold_10_total_metric": 0.018749982118606567, + "tpp_threshold_10_intended_diff_only": 0.02399998903274536, + "tpp_threshold_10_unintended_diff_only": 0.005250006914138794 + }, + "6": { + "tpp_threshold_2_total_metric": 0.02400001883506775, + "tpp_threshold_2_intended_diff_only": 0.022000014781951904, + "tpp_threshold_2_unintended_diff_only": -0.0020000040531158447, + "tpp_threshold_5_total_metric": 0.03449997305870056, + "tpp_threshold_5_intended_diff_only": 0.039999961853027344, + "tpp_threshold_5_unintended_diff_only": 0.005499988794326782, + "tpp_threshold_10_total_metric": 0.09775000810623169, + "tpp_threshold_10_intended_diff_only": 0.1029999852180481, + "tpp_threshold_10_unintended_diff_only": 0.005249977111816406 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..405e90523222dede92980c24bcce696bd045f6f3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9723087771204134, + "test_acc": 0.9723150357995227, + "test_auc": 0.9782028988677208, + "val_auc": 0.9854139854139854, + "k": 1, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431 + ] + }, + { + "test_f1": 0.9742169419417541, + "test_acc": 0.9742243436754177, + "test_auc": 0.9923778042681015, + "val_auc": 0.9927069927069927, + "k": 2, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928 + ] + }, + { + "test_f1": 0.9789957948374844, + "test_acc": 0.9789976133651551, + "test_auc": 0.997273653550311, + "val_auc": 0.9974259974259974, + "k": 5, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 126 + ] + }, + { + "test_f1": 0.987111713748006, + "test_acc": 0.9871121718377088, + "test_auc": 0.9991733925355615, + "val_auc": 0.9982839982839984, + "k": 10, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 126, + 461, + 1411, + 786, + 5951, + 1472 + ] + }, + { + "test_f1": 0.993317419389236, + "test_acc": 0.9933174224343675, + "test_auc": 0.999477788228089, + "val_auc": 0.9982839982839983, + "k": 16, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1500, + 1758, + 126, + 461, + 1411, + 786, + 5951, + 1472, + 542, + 1319, + 1848, + 1634, + 5020, + 7026 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..94fe60528c28a49b5736a2cad156c5b5a443ecc6 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9261162142244893, + "test_acc": 0.9261179520414776, + "test_auc": 0.9569212314267859, + "val_auc": 0.9462587846217446, + "k": 1, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838 + ] + }, + { + "test_f1": 0.956573717032053, + "test_acc": 0.9565780946208684, + "test_auc": 0.9659163457725987, + "val_auc": 0.95824720959074, + "k": 2, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441 + ] + }, + { + "test_f1": 0.9572094678356017, + "test_acc": 0.9572261827608555, + "test_auc": 0.9677190648038011, + "val_auc": 0.968582058701943, + "k": 5, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 1411 + ] + }, + { + "test_f1": 0.9598007051082187, + "test_acc": 0.9598185353208036, + "test_auc": 0.9841972272064407, + "val_auc": 0.9818106655642828, + "k": 10, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 1411, + 943, + 1697, + 1180, + 601, + 908 + ] + }, + { + "test_f1": 0.9636963598092818, + "test_acc": 0.9637070641607258, + "test_auc": 0.9844542784755684, + "val_auc": 0.9851178172798678, + "k": 16, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 1500, + 1411, + 943, + 1697, + 1180, + 601, + 908, + 1417, + 698, + 654, + 30737, + 920, + 1703 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c7f8158f3db4ada59c27cfd27c4a5f8187f2d56e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.572985218296196, + "test_acc": 0.5828930817610063, + "test_auc": 0.5930256931624979, + "val_auc": 0.5826330532212886, + "k": 1, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777 + ] + }, + { + "test_f1": 0.562963230407103, + "test_acc": 0.5811320754716981, + "test_auc": 0.6001004770444508, + "val_auc": 0.5882352941176471, + "k": 2, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921 + ] + }, + { + "test_f1": 0.5785566268948489, + "test_acc": 0.590440251572327, + "test_auc": 0.6211534430538946, + "val_auc": 0.6290516206482593, + "k": 5, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921, + 7284, + 1438, + 1380 + ] + }, + { + "test_f1": 0.5895637942405438, + "test_acc": 0.5924528301886792, + "test_auc": 0.6206521970271554, + "val_auc": 0.6754701880752302, + "k": 10, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921, + 7284, + 1438, + 1380, + 2046, + 969, + 1335, + 505, + 559 + ] + }, + { + "test_f1": 0.5988562433804158, + "test_acc": 0.6027672955974843, + "test_auc": 0.6292559078679425, + "val_auc": 0.6626650660264106, + "k": 16, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 777, + 2921, + 7284, + 1438, + 1380, + 2046, + 969, + 1335, + 505, + 559, + 1319, + 1178, + 6411, + 521, + 26301, + 394 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..748cfe0bb3c2b5d77f813cf080bdec92ef518106 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6514376684847857, + "test_acc": 0.6631446540880503, + "test_auc": 0.6720218138220364, + "val_auc": 0.7262905162064826, + "k": 1, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567 + ] + }, + { + "test_f1": 0.6678647985250032, + "test_acc": 0.670188679245283, + "test_auc": 0.7075229433976783, + "val_auc": 0.855142056822729, + "k": 2, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 433 + ] + }, + { + "test_f1": 0.6963158256656153, + "test_acc": 0.6976100628930818, + "test_auc": 0.7457833310886962, + "val_auc": 0.8027210884353742, + "k": 5, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 433, + 1164, + 4470, + 1345 + ] + }, + { + "test_f1": 0.7205609588091761, + "test_acc": 0.7220125786163522, + "test_auc": 0.7737555175036126, + "val_auc": 0.8095238095238095, + "k": 10, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 433, + 1164, + 4470, + 1345, + 1629, + 559, + 4505, + 1731, + 19328 + ] + }, + { + "test_f1": 0.7190053330821625, + "test_acc": 0.72, + "test_auc": 0.7757572106013029, + "val_auc": 0.8217286914765906, + "k": 16, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 433, + 1164, + 4470, + 1345, + 1629, + 559, + 4505, + 1731, + 19328, + 794, + 6411, + 4196, + 1335, + 993, + 523 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9253ee05ac9e4c9809bf4fbd0e0a6b8735556580 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/110_aimade_humangpt3_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7765972414422038, + "test_acc": 0.7766037735849056, + "test_auc": 0.8158089199515158, + "val_auc": 0.854341736694678, + "k": 1, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411 + ] + }, + { + "test_f1": 0.7499852827791499, + "test_acc": 0.7511949685534591, + "test_auc": 0.8221541123945485, + "val_auc": 0.86234493797519, + "k": 2, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786 + ] + }, + { + "test_f1": 0.7658535096732796, + "test_acc": 0.7660377358490567, + "test_auc": 0.8351862559352086, + "val_auc": 0.8927571028411365, + "k": 5, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417 + ] + }, + { + "test_f1": 0.7920346804061211, + "test_acc": 0.7922012578616352, + "test_auc": 0.8876638036573745, + "val_auc": 0.9297719087635055, + "k": 10, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417, + 1758, + 1335, + 181, + 5054, + 1360 + ] + }, + { + "test_f1": 0.7969776371487346, + "test_acc": 0.7969811320754717, + "test_auc": 0.8944056892942963, + "val_auc": 0.9425770308123249, + "k": 16, + "dataset": "110_aimade_humangpt3", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1411, + 786, + 800, + 6718, + 1417, + 1758, + 1335, + 181, + 5054, + 1360, + 1183, + 191, + 1108, + 19471, + 1703, + 12220 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b956982f4cee5b70f99029adfb8d654d75b89aee --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/113_movie_sent_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8400601970919783, + "test_acc": 0.8402515723270441, + "test_auc": 0.8914339079266743, + "val_auc": 0.9443777511004402, + "k": 1, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.8744474824673062, + "test_acc": 0.8744654088050314, + "test_auc": 0.9384862268730652, + "val_auc": 0.9579831932773109, + "k": 2, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186 + ] + }, + { + "test_f1": 0.8797403903698112, + "test_acc": 0.879748427672956, + "test_auc": 0.9431209805384903, + "val_auc": 0.9611844737895159, + "k": 5, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 7867, + 821 + ] + }, + { + "test_f1": 0.8832660074834806, + "test_acc": 0.8832704402515723, + "test_auc": 0.9458051783271343, + "val_auc": 0.9631852741096438, + "k": 10, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 7867, + 821, + 505, + 1697, + 4675, + 7284, + 6745 + ] + }, + { + "test_f1": 0.8890523907462876, + "test_acc": 0.8890566037735849, + "test_auc": 0.9454555718812118, + "val_auc": 0.9615846338535414, + "k": 16, + "dataset": "113_movie_sent", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 186, + 6319, + 7867, + 821, + 505, + 1697, + 4675, + 7284, + 6745, + 1791, + 5256, + 100, + 7143, + 1438, + 2632 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b23eef78ac761ae770d259ce4354bc59cc2875f9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/114_nyc_borough_Manhattan_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5180242187214908, + "test_acc": 0.5241025641025641, + "test_auc": 0.5360773555054363, + "val_auc": 0.5569155446756426, + "k": 1, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264 + ] + }, + { + "test_f1": 0.5473552694312842, + "test_acc": 0.5487179487179488, + "test_auc": 0.5705263069310264, + "val_auc": 0.5879232966136271, + "k": 2, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945 + ] + }, + { + "test_f1": 0.5501666629428935, + "test_acc": 0.556923076923077, + "test_auc": 0.5726890968458612, + "val_auc": 0.6050591595267238, + "k": 5, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 470 + ] + }, + { + "test_f1": 0.5587373178937548, + "test_acc": 0.5620512820512821, + "test_auc": 0.604781701282526, + "val_auc": 0.6077111383108935, + "k": 10, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 470, + 913, + 1924, + 7073, + 4531, + 375 + ] + }, + { + "test_f1": 0.5719927037712071, + "test_acc": 0.5743589743589743, + "test_auc": 0.6299146665769011, + "val_auc": 0.6436148510811913, + "k": 16, + "dataset": "114_nyc_borough_Manhattan", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1264, + 1945, + 238, + 6892, + 470, + 913, + 1924, + 7073, + 4531, + 375, + 1851, + 1523, + 386, + 1132, + 5888, + 26171 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..603cdca3e1d454ec1aed5d14cbad834102e87d4c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/115_nyc_borough_Brooklyn_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3339032896927634, + "test_acc": 0.5005128205128205, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239 + ] + }, + { + "test_f1": 0.3339032896927634, + "test_acc": 0.5005128205128205, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 2, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 456 + ] + }, + { + "test_f1": 0.4997729416337038, + "test_acc": 0.5005128205128205, + "test_auc": 0.5182385464705288, + "val_auc": 0.5263157894736842, + "k": 5, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 456, + 1409, + 827, + 1264 + ] + }, + { + "test_f1": 0.4998744663967559, + "test_acc": 0.5025641025641026, + "test_auc": 0.5186950886996331, + "val_auc": 0.5724194206446349, + "k": 10, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 456, + 1409, + 827, + 1264, + 238, + 6341, + 326, + 8141, + 957 + ] + }, + { + "test_f1": 0.4970043636131784, + "test_acc": 0.49948717948717947, + "test_auc": 0.5131134917696166, + "val_auc": 0.5524275805793555, + "k": 16, + "dataset": "115_nyc_borough_Brooklyn", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 239, + 456, + 1409, + 827, + 1264, + 238, + 6341, + 326, + 8141, + 957, + 11202, + 15805, + 538, + 1478, + 3301, + 16809 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..49018e209985977564e648e61959740e57f08bf2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/116_nyc_borough_Bronx_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.523982387982388, + "test_acc": 0.5241025641025641, + "test_auc": 0.5407079981149224, + "val_auc": 0.6036311709506323, + "k": 1, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815 + ] + }, + { + "test_f1": 0.5450035162487642, + "test_acc": 0.5456410256410257, + "test_auc": 0.5516103107011815, + "val_auc": 0.6307629538963688, + "k": 2, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156 + ] + }, + { + "test_f1": 0.5221025572876735, + "test_acc": 0.5271794871794871, + "test_auc": 0.5454774968862558, + "val_auc": 0.6242350061199511, + "k": 5, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 12594, + 16809, + 1770 + ] + }, + { + "test_f1": 0.5326668254140436, + "test_acc": 0.5374358974358975, + "test_auc": 0.5461991281516141, + "val_auc": 0.62749898000816, + "k": 10, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 12594, + 16809, + 1770, + 698, + 191, + 1409, + 400, + 360 + ] + }, + { + "test_f1": 0.5482395625653732, + "test_acc": 0.5517948717948717, + "test_auc": 0.5697226242973037, + "val_auc": 0.6352509179926561, + "k": 16, + "dataset": "116_nyc_borough_Bronx", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1815, + 1156, + 12594, + 16809, + 1770, + 698, + 191, + 1409, + 400, + 360, + 1611, + 4110, + 1265, + 4347, + 848, + 5633 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..08daf306b050faa81c2d19d30bacf80c048b2189 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/117_us_state_FL_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8577738003496822, + "test_acc": 0.8585858585858586, + "test_auc": 0.8940816326530612, + "val_auc": 0.8687900641025642, + "k": 1, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997 + ] + }, + { + "test_f1": 0.8671979042043984, + "test_acc": 0.8686868686868687, + "test_auc": 0.9081632653061225, + "val_auc": 0.874198717948718, + "k": 2, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369 + ] + }, + { + "test_f1": 0.8467668125435367, + "test_acc": 0.8484848484848485, + "test_auc": 0.8891836734693878, + "val_auc": 0.9154647435897436, + "k": 5, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 2335, + 1156, + 1499 + ] + }, + { + "test_f1": 0.8473644735326978, + "test_acc": 0.8484848484848485, + "test_auc": 0.9016326530612245, + "val_auc": 0.9350961538461537, + "k": 10, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 2335, + 1156, + 1499, + 1279, + 314, + 3526, + 2412, + 1684 + ] + }, + { + "test_f1": 0.8080416289051587, + "test_acc": 0.8080808080808081, + "test_auc": 0.9093877551020407, + "val_auc": 0.9334935897435899, + "k": 16, + "dataset": "117_us_state_FL", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7997, + 6369, + 2335, + 1156, + 1499, + 1279, + 314, + 3526, + 2412, + 1684, + 1473, + 5836, + 1734, + 337, + 4201, + 581 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8e8af102500d19e094c64affa11d794d3be292c0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/118_us_state_CA_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8880672805906451, + "test_acc": 0.8888888888888888, + "test_auc": 0.9089795918367346, + "val_auc": 0.9270833333333334, + "k": 1, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8984098573926301, + "test_acc": 0.898989898989899, + "test_auc": 0.9522448979591835, + "val_auc": 0.9577323717948718, + "k": 2, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789 + ] + }, + { + "test_f1": 0.8880672805906451, + "test_acc": 0.8888888888888888, + "test_auc": 0.9342857142857143, + "val_auc": 0.9747596153846154, + "k": 5, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789, + 2335, + 4889, + 1116 + ] + }, + { + "test_f1": 0.8880672805906451, + "test_acc": 0.8888888888888888, + "test_auc": 0.9685714285714285, + "val_auc": 0.9815705128205128, + "k": 10, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789, + 2335, + 4889, + 1116, + 4201, + 1924, + 5388, + 2832, + 581 + ] + }, + { + "test_f1": 0.89867992527567, + "test_acc": 0.898989898989899, + "test_auc": 0.9640816326530611, + "val_auc": 0.984375, + "k": 16, + "dataset": "118_us_state_CA", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 789, + 2335, + 4889, + 1116, + 4201, + 1924, + 5388, + 2832, + 581, + 2900, + 461, + 925, + 7481, + 737, + 3526 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e77c258e9b357e26cb11f6817ba7a026e3ce6f85 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/119_us_state_TX_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8677158770616714, + "test_acc": 0.8686868686868687, + "test_auc": 0.8873469387755102, + "val_auc": 0.8279246794871794, + "k": 1, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388 + ] + }, + { + "test_f1": 0.8681220810253067, + "test_acc": 0.8686868686868687, + "test_auc": 0.9191836734693878, + "val_auc": 0.8810096153846153, + "k": 2, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335 + ] + }, + { + "test_f1": 0.8681220810253067, + "test_acc": 0.8686868686868687, + "test_auc": 0.952857142857143, + "val_auc": 0.9401041666666666, + "k": 5, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 1769, + 337, + 5888 + ] + }, + { + "test_f1": 0.8382848088730442, + "test_acc": 0.8383838383838383, + "test_auc": 0.9567346938775511, + "val_auc": 0.9567307692307693, + "k": 10, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 1769, + 337, + 5888, + 7481, + 1596, + 2000, + 4201, + 2832 + ] + }, + { + "test_f1": 0.8584992077639138, + "test_acc": 0.8585858585858586, + "test_auc": 0.9612244897959183, + "val_auc": 0.9599358974358975, + "k": 16, + "dataset": "119_us_state_TX", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 2335, + 1769, + 337, + 5888, + 7481, + 1596, + 2000, + 4201, + 2832, + 2900, + 5463, + 23520, + 7384, + 1473, + 1499 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..daa5b71b9c29edb759e91ddeeb7a977210ea51d5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/120_us_timezone_Chicago_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6440080589493589, + "test_acc": 0.6601647160814911, + "test_auc": 0.673656695441475, + "val_auc": 0.6868000000000001, + "k": 1, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388 + ] + }, + { + "test_f1": 0.6559494020632926, + "test_acc": 0.6749024707412223, + "test_auc": 0.7789839932299285, + "val_auc": 0.8392000000000001, + "k": 2, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 6747 + ] + }, + { + "test_f1": 0.7004262206792526, + "test_acc": 0.7026441265713047, + "test_auc": 0.8032008279208335, + "val_auc": 0.8772, + "k": 5, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 6747, + 789, + 12946, + 337 + ] + }, + { + "test_f1": 0.7312523840065449, + "test_acc": 0.7325530992631123, + "test_auc": 0.8285848385870032, + "val_auc": 0.906, + "k": 10, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 6747, + 789, + 12946, + 337, + 7384, + 5352, + 7118, + 4889, + 2000 + ] + }, + { + "test_f1": 0.7766462552008653, + "test_acc": 0.7767663632423061, + "test_auc": 0.8784645134912917, + "val_auc": 0.9211999999999999, + "k": 16, + "dataset": "120_us_timezone_Chicago", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5388, + 6747, + 789, + 12946, + 337, + 7384, + 5352, + 7118, + 4889, + 2000, + 2659, + 7481, + 239, + 4201, + 1602, + 1116 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..87f0836933efb30fb5258478c64d2cfd5c74b111 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/121_us_timezone_New_York_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6251822754396058, + "test_acc": 0.6545296922410057, + "test_auc": 0.6638593316207737, + "val_auc": 0.6812, + "k": 1, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.7174648979609495, + "test_acc": 0.7208495882097963, + "test_auc": 0.7578196280969998, + "val_auc": 0.7508, + "k": 2, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384 + ] + }, + { + "test_f1": 0.7648440628163885, + "test_acc": 0.765496315561335, + "test_auc": 0.8644632869419087, + "val_auc": 0.9028, + "k": 5, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 4201, + 3526, + 5388 + ] + }, + { + "test_f1": 0.7968718622691713, + "test_acc": 0.7975726051148678, + "test_auc": 0.8984004503360234, + "val_auc": 0.9172, + "k": 10, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 4201, + 3526, + 5388, + 7481, + 1924, + 3842, + 1091, + 6244 + ] + }, + { + "test_f1": 0.8253470370977912, + "test_acc": 0.8257477243172952, + "test_auc": 0.9185006035043839, + "val_auc": 0.9292, + "k": 16, + "dataset": "121_us_timezone_New_York", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 7384, + 4201, + 3526, + 5388, + 7481, + 1924, + 3842, + 1091, + 6244, + 2900, + 2050, + 2659, + 3825, + 2832, + 1101 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b22893aa3a4ebe4a698edc63ac86b773b43013a7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/122_us_timezone_Los_Angeles_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8161133627731844, + "test_acc": 0.8188123103597746, + "test_auc": 0.834885559635703, + "val_auc": 0.8728, + "k": 1, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747 + ] + }, + { + "test_f1": 0.8100022246247579, + "test_acc": 0.8140442132639792, + "test_auc": 0.834885559635703, + "val_auc": 0.8728, + "k": 2, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889 + ] + }, + { + "test_f1": 0.8239563332569495, + "test_acc": 0.8266146510619853, + "test_auc": 0.9090835301173489, + "val_auc": 0.8884000000000001, + "k": 5, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 5388, + 14154 + ] + }, + { + "test_f1": 0.8369347088050375, + "test_acc": 0.8383181621153013, + "test_auc": 0.9134703982227059, + "val_auc": 0.9008, + "k": 10, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 5388, + 14154, + 1924, + 4531, + 737, + 12946, + 1091 + ] + }, + { + "test_f1": 0.8129314948143386, + "test_acc": 0.8170784568703945, + "test_auc": 0.9134313169923686, + "val_auc": 0.956, + "k": 16, + "dataset": "122_us_timezone_Los_Angeles", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6747, + 4889, + 789, + 5388, + 14154, + 1924, + 4531, + 737, + 12946, + 1091, + 4201, + 1158, + 1851, + 3526, + 1116, + 470 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..29a04d89f129287021847c123d91cdfe3886eaad --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/123_world_country_United_Kingdom_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9415217295759696, + "test_acc": 0.9415384615384615, + "test_auc": 0.9788559935368769, + "val_auc": 0.9885760913912689, + "k": 1, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958 + ] + }, + { + "test_f1": 0.9405045593085622, + "test_acc": 0.9405128205128205, + "test_auc": 0.9781701282526005, + "val_auc": 0.9906160750713994, + "k": 2, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818 + ] + }, + { + "test_f1": 0.9764084696111573, + "test_acc": 0.9764102564102564, + "test_auc": 0.9924302352980778, + "val_auc": 0.9983680130558956, + "k": 5, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 5888, + 508, + 1499 + ] + }, + { + "test_f1": 0.977435612600348, + "test_acc": 0.9774358974358974, + "test_auc": 0.9940439122092437, + "val_auc": 1.0, + "k": 10, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 5888, + 508, + 1499, + 4586, + 2780, + 730, + 1279, + 1158 + ] + }, + { + "test_f1": 0.978461402518014, + "test_acc": 0.9784615384615385, + "test_auc": 0.9946940115124382, + "val_auc": 0.9995920032639738, + "k": 16, + "dataset": "123_world_country_United_Kingdom", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3958, + 3818, + 5888, + 508, + 1499, + 4586, + 2780, + 730, + 1279, + 1158, + 3475, + 386, + 5463, + 1596, + 779, + 851 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7fbfb983ff12a8447dc98b9f6bda5bae87fd78a7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/124_world_country_United_States_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.926839190384715, + "test_acc": 0.9271794871794872, + "test_auc": 0.9286468677416098, + "val_auc": 0.9418604651162791, + "k": 1, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065 + ] + }, + { + "test_f1": 0.9278594089471172, + "test_acc": 0.9282051282051282, + "test_auc": 0.9762008954118558, + "val_auc": 0.9826601387188902, + "k": 2, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508 + ] + }, + { + "test_f1": 0.9558542493069856, + "test_acc": 0.9558974358974359, + "test_auc": 0.9883781600296226, + "val_auc": 0.993064055487556, + "k": 5, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 5406, + 12813, + 3291 + ] + }, + { + "test_f1": 0.9548455756589671, + "test_acc": 0.9548717948717949, + "test_auc": 0.9919463426128522, + "val_auc": 0.9983680130558955, + "k": 10, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 5406, + 12813, + 3291, + 1499, + 1279, + 3475, + 992, + 581 + ] + }, + { + "test_f1": 0.9733302472512828, + "test_acc": 0.9733333333333334, + "test_auc": 0.997231292288013, + "val_auc": 1.0, + "k": 16, + "dataset": "124_world_country_United_States", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1065, + 508, + 5406, + 12813, + 3291, + 1499, + 1279, + 3475, + 992, + 581, + 5888, + 3958, + 778, + 2708, + 6828, + 1482 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..772c767df89e1e79d715e7c8c5c96b21d0a1a58d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/125_world_country_Italy_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9671635291367049, + "test_acc": 0.9671794871794872, + "test_auc": 0.9809977446393106, + "val_auc": 1.0, + "k": 1, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037 + ] + }, + { + "test_f1": 0.9671635291367049, + "test_acc": 0.9671794871794872, + "test_auc": 0.977711482142256, + "val_auc": 1.0, + "k": 2, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508 + ] + }, + { + "test_f1": 0.9753766894429566, + "test_acc": 0.9753846153846154, + "test_auc": 0.9885485744100717, + "val_auc": 1.0, + "k": 5, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 5472, + 1253 + ] + }, + { + "test_f1": 0.9805096221322112, + "test_acc": 0.9805128205128205, + "test_auc": 0.9977109772107584, + "val_auc": 1.0, + "k": 10, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 5472, + 1253, + 778, + 1279, + 1079, + 3958, + 7734 + ] + }, + { + "test_f1": 0.9815358977164503, + "test_acc": 0.9815384615384616, + "test_auc": 0.9982327397583062, + "val_auc": 1.0, + "k": 16, + "dataset": "125_world_country_Italy", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4037, + 508, + 1028, + 5472, + 1253, + 778, + 1279, + 1079, + 3958, + 7734, + 26034, + 582, + 581, + 3475, + 343, + 493 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dce52f35e77a66c632fd8b4d459f9fabd37b873a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/126_art_type_book_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5809232744791196, + "test_acc": 0.6107498916341569, + "test_auc": 0.6174360909149668, + "val_auc": 0.6576, + "k": 1, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351 + ] + }, + { + "test_f1": 0.6426018087282287, + "test_acc": 0.6493281317728652, + "test_auc": 0.6743898442913596, + "val_auc": 0.6816000000000001, + "k": 2, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1894 + ] + }, + { + "test_f1": 0.677444174583433, + "test_acc": 0.6801040312093628, + "test_auc": 0.7291340050294537, + "val_auc": 0.8064, + "k": 5, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1894, + 1441, + 559, + 3078 + ] + }, + { + "test_f1": 0.6888080411142931, + "test_acc": 0.6918075422626788, + "test_auc": 0.7615451215351107, + "val_auc": 0.7768, + "k": 10, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1894, + 1441, + 559, + 3078, + 3274, + 654, + 343, + 7595, + 619 + ] + }, + { + "test_f1": 0.7096143265868258, + "test_acc": 0.7104464672735153, + "test_auc": 0.7877566020974596, + "val_auc": 0.8288, + "k": 16, + "dataset": "126_art_type_book", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1351, + 1894, + 1441, + 559, + 3078, + 3274, + 654, + 343, + 7595, + 619, + 4510, + 1960, + 649, + 1602, + 52, + 2046 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..793cb705ba47d1ab10d574f3b956c1d997f031ca --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/127_art_type_song_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6334724334342453, + "test_acc": 0.64629388816645, + "test_auc": 0.6581264157551471, + "val_auc": 0.6892, + "k": 1, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274 + ] + }, + { + "test_f1": 0.6334724334342453, + "test_acc": 0.64629388816645, + "test_auc": 0.6569498452533591, + "val_auc": 0.7054, + "k": 2, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121 + ] + }, + { + "test_f1": 0.6985218445200152, + "test_acc": 0.6991764195925444, + "test_auc": 0.7501758655365176, + "val_auc": 0.7762, + "k": 5, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 698, + 567, + 649 + ] + }, + { + "test_f1": 0.7528120903770202, + "test_acc": 0.752925877763329, + "test_auc": 0.8135690031730952, + "val_auc": 0.836, + "k": 10, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 698, + 567, + 649, + 52, + 3078, + 4510, + 1895, + 1602 + ] + }, + { + "test_f1": 0.7567872911903648, + "test_acc": 0.7568270481144344, + "test_auc": 0.8448302296322908, + "val_auc": 0.8660000000000001, + "k": 16, + "dataset": "127_art_type_song", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3274, + 121, + 698, + 567, + 649, + 52, + 3078, + 4510, + 1895, + 1602, + 1524, + 1499, + 1158, + 2046, + 19018, + 1894 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..75713086cd3fd07fe42bd63e1828c5a8f6d135f5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/128_art_type_movie_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5409996600881487, + "test_acc": 0.5708712613784135, + "test_auc": 0.577756241347641, + "val_auc": 0.5367999999999999, + "k": 1, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.5468759321401215, + "test_acc": 0.5825747724317295, + "test_auc": 0.64622918736594, + "val_auc": 0.6659999999999999, + "k": 2, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508 + ] + }, + { + "test_f1": 0.6364910127413209, + "test_acc": 0.6415257910706545, + "test_auc": 0.7018098367456759, + "val_auc": 0.7388, + "k": 5, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121 + ] + }, + { + "test_f1": 0.6537779379848431, + "test_acc": 0.6545296922410057, + "test_auc": 0.7286409802775068, + "val_auc": 0.792, + "k": 10, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 7893, + 51, + 1441, + 3274, + 52 + ] + }, + { + "test_f1": 0.6626985711270891, + "test_acc": 0.6627654963155614, + "test_auc": 0.7389561704001768, + "val_auc": 0.8284, + "k": 16, + "dataset": "128_art_type_movie", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 508, + 778, + 581, + 121, + 7893, + 51, + 1441, + 3274, + 52, + 567, + 925, + 957, + 4510, + 4654, + 84 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8fd0b599d20513a04a5aabd48fd7b983c5095600 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/129_arith_mc_A_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.632056836001399, + "test_acc": 0.6460554371002132, + "test_auc": 0.6551463902527732, + "val_auc": 0.6207999999999999, + "k": 1, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504 + ] + }, + { + "test_f1": 0.6193262494838208, + "test_acc": 0.6226012793176973, + "test_auc": 0.6996544826332061, + "val_auc": 0.7004, + "k": 2, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 1048 + ] + }, + { + "test_f1": 0.6978187050040029, + "test_acc": 0.6993603411513859, + "test_auc": 0.7835515548281505, + "val_auc": 0.7639999999999999, + "k": 5, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 1048, + 4672, + 927, + 696 + ] + }, + { + "test_f1": 0.7171921846854867, + "test_acc": 0.7185501066098081, + "test_auc": 0.8006728496090197, + "val_auc": 0.7744, + "k": 10, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 1048, + 4672, + 927, + 696, + 1902, + 6611, + 1818, + 246, + 1974 + ] + }, + { + "test_f1": 0.7285297487699793, + "test_acc": 0.7292110874200426, + "test_auc": 0.8046190216402982, + "val_auc": 0.7936, + "k": 16, + "dataset": "129_arith_mc_A", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1504, + 1048, + 4672, + 927, + 696, + 1902, + 6611, + 1818, + 246, + 1974, + 437, + 1499, + 7129, + 741, + 1927, + 29085 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..fb58759775b1871e7f63ee2a49d07dd192965582 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/130_temp_cat_Frequency_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9393815708101421, + "test_acc": 0.9393939393939394, + "test_auc": 0.9608163265306123, + "val_auc": 0.9512882447665058, + "k": 1, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299 + ] + }, + { + "test_f1": 0.9494949494949495, + "test_acc": 0.9494949494949495, + "test_auc": 0.9702040816326531, + "val_auc": 0.9530998389694042, + "k": 2, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114 + ] + }, + { + "test_f1": 0.9494949494949495, + "test_acc": 0.9494949494949495, + "test_auc": 0.9971428571428571, + "val_auc": 0.969404186795491, + "k": 5, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 432, + 2535, + 23633 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 0.9995918367346939, + "val_auc": 0.9923510466988728, + "k": 10, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 432, + 2535, + 23633, + 6877, + 1232, + 15752, + 27804, + 19940 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 0.9955102040816326, + "val_auc": 0.998389694041868, + "k": 16, + "dataset": "130_temp_cat_Frequency", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1299, + 114, + 432, + 2535, + 23633, + 6877, + 1232, + 15752, + 27804, + 19940, + 3959, + 1610, + 1244, + 1420, + 373, + 2887 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..104c0b9f3dfbeae7fd35cd41d2aaebe9118fc822 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/131_temp_cat_Typical Time_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7402786376070819, + "test_acc": 0.7474747474747475, + "test_auc": 0.7589795918367347, + "val_auc": 0.7238325281803544, + "k": 1, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791 + ] + }, + { + "test_f1": 0.8383838383838383, + "test_acc": 0.8383838383838383, + "test_auc": 0.8755102040816327, + "val_auc": 0.8599033816425122, + "k": 2, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 1657 + ] + }, + { + "test_f1": 0.8888888888888888, + "test_acc": 0.8888888888888888, + "test_auc": 0.9510204081632654, + "val_auc": 0.932769726247987, + "k": 5, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 1657, + 114, + 23557, + 278 + ] + }, + { + "test_f1": 0.8989692846835704, + "test_acc": 0.898989898989899, + "test_auc": 0.9640816326530612, + "val_auc": 0.9694041867954911, + "k": 10, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 1657, + 114, + 23557, + 278, + 4580, + 1396, + 253, + 1299, + 521 + ] + }, + { + "test_f1": 0.9392699499082477, + "test_acc": 0.9393939393939394, + "test_auc": 0.9926530612244898, + "val_auc": 0.9891304347826088, + "k": 16, + "dataset": "131_temp_cat_Typical Time", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23791, + 1657, + 114, + 23557, + 278, + 4580, + 1396, + 253, + 1299, + 521, + 25266, + 5387, + 1420, + 1658, + 1610, + 16 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..211b75286c9e70e9e1fa1d0df3a2a0f1e3bf9be2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/132_temp_cat_Event Ordering_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9494330117733285, + "test_acc": 0.9494949494949495, + "test_auc": 0.9538775510204082, + "val_auc": 0.9428341384863125, + "k": 1, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9595464448405625, + "test_acc": 0.9595959595959596, + "test_auc": 0.996734693877551, + "val_auc": 0.9855072463768116, + "k": 2, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244 + ] + }, + { + "test_f1": 0.979793856936714, + "test_acc": 0.9797979797979798, + "test_auc": 0.9983673469387755, + "val_auc": 0.9887278582930757, + "k": 5, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244, + 1610, + 177, + 1501 + ] + }, + { + "test_f1": 1.0, + "test_acc": 1.0, + "test_auc": 1.0, + "val_auc": 0.9987922705314011, + "k": 10, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244, + 1610, + 177, + 1501, + 15752, + 3644, + 7577, + 1550, + 1960 + ] + }, + { + "test_f1": 0.98989898989899, + "test_acc": 0.98989898989899, + "test_auc": 0.9987755102040816, + "val_auc": 0.999597423510467, + "k": 16, + "dataset": "132_temp_cat_Event Ordering", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 1244, + 1610, + 177, + 1501, + 15752, + 3644, + 7577, + 1550, + 1960, + 114, + 1360, + 25152, + 27804, + 943, + 635 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..804ffe25b357802cc2e6365a151c444a60b93c80 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/133_context_type_Causality_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9021835748792271, + "test_acc": 0.9022222222222223, + "test_auc": 0.9361567635903918, + "val_auc": 0.9184253246753247, + "k": 1, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752 + ] + }, + { + "test_f1": 0.9244444444444444, + "test_acc": 0.9244444444444444, + "test_auc": 0.9468236409608091, + "val_auc": 0.9182224025974025, + "k": 2, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309 + ] + }, + { + "test_f1": 0.9422153739444982, + "test_acc": 0.9422222222222222, + "test_auc": 0.9635745891276866, + "val_auc": 0.9811282467532467, + "k": 5, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580 + ] + }, + { + "test_f1": 0.9422153739444982, + "test_acc": 0.9422222222222222, + "test_auc": 0.9629424778761062, + "val_auc": 0.9857954545454546, + "k": 10, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 943, + 3301, + 27804, + 505, + 10260 + ] + }, + { + "test_f1": 0.951109179545599, + "test_acc": 0.9511111111111111, + "test_auc": 0.9687104930467763, + "val_auc": 0.9902597402597402, + "k": 16, + "dataset": "133_context_type_Causality", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15752, + 1309, + 959, + 1369, + 22580, + 943, + 3301, + 27804, + 505, + 10260, + 969, + 5563, + 1182, + 1915, + 1073, + 1541 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4a8e8248b516c382255f8378aebfb108b1ccdcad --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/134_context_type_Belief_states_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8130009496676164, + "test_acc": 0.8133333333333334, + "test_auc": 0.8350189633375474, + "val_auc": 0.8374594155844156, + "k": 1, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406 + ] + }, + { + "test_f1": 0.825730641201268, + "test_acc": 0.8266666666666667, + "test_auc": 0.8704962073324906, + "val_auc": 0.8478084415584417, + "k": 2, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854 + ] + }, + { + "test_f1": 0.8072907471931862, + "test_acc": 0.8088888888888889, + "test_auc": 0.8865755372945638, + "val_auc": 0.8701298701298701, + "k": 5, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1232, + 1182 + ] + }, + { + "test_f1": 0.8439637355308598, + "test_acc": 0.8444444444444444, + "test_auc": 0.9184576485461441, + "val_auc": 0.8798701298701299, + "k": 10, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1232, + 1182, + 1804, + 6662, + 15752, + 1570, + 311 + ] + }, + { + "test_f1": 0.844105778217604, + "test_acc": 0.8444444444444444, + "test_auc": 0.9210651074589127, + "val_auc": 0.8859577922077921, + "k": 16, + "dataset": "134_context_type_Belief_states", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 406, + 854, + 1915, + 1232, + 1182, + 1804, + 6662, + 15752, + 1570, + 311, + 800, + 3481, + 920, + 1411, + 748, + 629 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4c8834a92b52fca5b3afce4a86dbfbeb9fd8d97b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/135_context_type_Event_duration_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9288860795055486, + "test_acc": 0.9288888888888889, + "test_auc": 0.9567398862199747, + "val_auc": 0.9399350649350648, + "k": 1, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396 + ] + }, + { + "test_f1": 0.9243817167104839, + "test_acc": 0.9244444444444444, + "test_auc": 0.9679203539823009, + "val_auc": 0.9602272727272727, + "k": 2, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16 + ] + }, + { + "test_f1": 0.9510705225773719, + "test_acc": 0.9511111111111111, + "test_auc": 0.9842762326169405, + "val_auc": 0.9667207792207793, + "k": 5, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 21467, + 114 + ] + }, + { + "test_f1": 0.9377408871321915, + "test_acc": 0.9377777777777778, + "test_auc": 0.9919405815423514, + "val_auc": 0.9825487012987012, + "k": 10, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 21467, + 114, + 15752, + 1869, + 1960, + 7033, + 1835 + ] + }, + { + "test_f1": 0.9555292050944225, + "test_acc": 0.9555555555555556, + "test_auc": 0.9920986093552464, + "val_auc": 0.9849837662337663, + "k": 16, + "dataset": "135_context_type_Event_duration", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1396, + 16, + 1658, + 21467, + 114, + 15752, + 1869, + 1960, + 7033, + 1835, + 969, + 79, + 177, + 20486, + 4114, + 3481 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a748afd80425cc75e6eb0bb72c0e55ede154ef22 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/136_glue_mnli_entailment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6575262237184158, + "test_acc": 0.6605981794538361, + "test_auc": 0.6850905857825491, + "val_auc": 0.7696000000000001, + "k": 1, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649 + ] + }, + { + "test_f1": 0.692747996640028, + "test_acc": 0.6965756393584742, + "test_auc": 0.7322217980071579, + "val_auc": 0.8244, + "k": 2, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 28053 + ] + }, + { + "test_f1": 0.7152953992532521, + "test_acc": 0.717381881231036, + "test_auc": 0.7818583425650215, + "val_auc": 0.8528000000000001, + "k": 5, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 28053, + 5394, + 943, + 1928 + ] + }, + { + "test_f1": 0.7246854372134179, + "test_acc": 0.7264846120502817, + "test_auc": 0.7970286239949737, + "val_auc": 0.8716, + "k": 10, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 28053, + 5394, + 943, + 1928, + 7614, + 7246, + 1560, + 1737, + 811 + ] + }, + { + "test_f1": 0.7426522175739533, + "test_acc": 0.7433896835717382, + "test_auc": 0.8182301914529349, + "val_auc": 0.8752, + "k": 16, + "dataset": "136_glue_mnli_entailment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7649, + 28053, + 5394, + 943, + 1928, + 7614, + 7246, + 1560, + 1737, + 811, + 253, + 584, + 17225, + 1551, + 1848, + 13 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..63ada7d61f6bdca6b0f3d22e160384842a444d36 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/137_glue_mnli_neutral_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.610744625733531, + "test_acc": 0.6107498916341569, + "test_auc": 0.6437888651562271, + "val_auc": 0.6502, + "k": 1, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943 + ] + }, + { + "test_f1": 0.6299389563215626, + "test_acc": 0.6302557433896836, + "test_auc": 0.6733936486988205, + "val_auc": 0.7008, + "k": 2, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253 + ] + }, + { + "test_f1": 0.6439877024835846, + "test_acc": 0.6445600346770698, + "test_auc": 0.7091867947528938, + "val_auc": 0.7575999999999999, + "k": 5, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 28053, + 1011, + 7649 + ] + }, + { + "test_f1": 0.6430423929575055, + "test_acc": 0.6436931079323797, + "test_auc": 0.708705043432775, + "val_auc": 0.7476, + "k": 10, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 28053, + 1011, + 7649, + 1551, + 5394, + 1928, + 7614, + 32305 + ] + }, + { + "test_f1": 0.6785898583985397, + "test_acc": 0.6788036410923277, + "test_auc": 0.7443719270503741, + "val_auc": 0.7467999999999999, + "k": 16, + "dataset": "137_glue_mnli_neutral", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 943, + 253, + 28053, + 1011, + 7649, + 1551, + 5394, + 1928, + 7614, + 32305, + 27731, + 100, + 1417, + 6718, + 1500, + 4674 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..98a79adbed050edb522b4b1906ebdc5cef917ff0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/138_glue_mnli_contradiction_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6125476251082548, + "test_acc": 0.6250541829215431, + "test_auc": 0.6423601455625516, + "val_auc": 0.6956000000000001, + "k": 1, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253 + ] + }, + { + "test_f1": 0.615762417301407, + "test_acc": 0.6289553532726485, + "test_auc": 0.7008948098622988, + "val_auc": 0.7468, + "k": 2, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394 + ] + }, + { + "test_f1": 0.7005096055579836, + "test_acc": 0.7048114434330299, + "test_auc": 0.7883845322502823, + "val_auc": 0.8075999999999999, + "k": 5, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 1541, + 7649, + 844 + ] + }, + { + "test_f1": 0.7246024681329558, + "test_acc": 0.7273515387949718, + "test_auc": 0.8062502912303222, + "val_auc": 0.7796, + "k": 10, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 1541, + 7649, + 844, + 100, + 1099, + 28053, + 1960, + 7246 + ] + }, + { + "test_f1": 0.7349655742573894, + "test_acc": 0.7373211963589077, + "test_auc": 0.823670749653154, + "val_auc": 0.776, + "k": 16, + "dataset": "138_glue_mnli_contradiction", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 253, + 5394, + 1541, + 7649, + 844, + 100, + 1099, + 28053, + 1960, + 7246, + 794, + 7143, + 24568, + 2429, + 1737, + 4675 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ee9a919cc289845bd3e3649acbde60a8ee1b0b8b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/139_news_class_Politics_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7998008568808032, + "test_acc": 0.8033898305084746, + "test_auc": 0.8159682003875669, + "val_auc": 0.798913043478261, + "k": 1, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7998008568808032, + "test_acc": 0.8033898305084746, + "test_auc": 0.840571716436296, + "val_auc": 0.8870772946859904, + "k": 2, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5020 + ] + }, + { + "test_f1": 0.7998008568808032, + "test_acc": 0.8033898305084746, + "test_auc": 0.8566774405871604, + "val_auc": 0.8884863123993559, + "k": 5, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5020, + 1847, + 6838, + 32278 + ] + }, + { + "test_f1": 0.8275201509359603, + "test_acc": 0.8291525423728814, + "test_auc": 0.9226649457810725, + "val_auc": 0.9039855072463768, + "k": 10, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5020, + 1847, + 6838, + 32278, + 266, + 508, + 4413, + 4674, + 1528 + ] + }, + { + "test_f1": 0.8689736795112867, + "test_acc": 0.8691525423728813, + "test_auc": 0.9500143407132852, + "val_auc": 0.9408212560386473, + "k": 16, + "dataset": "139_news_class_Politics", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5020, + 1847, + 6838, + 32278, + 266, + 508, + 4413, + 4674, + 1528, + 776, + 27673, + 1697, + 1758, + 1628, + 1433 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4a04fb4a5a880aa13f9350c55d9e861e71c12fcd --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/140_news_class_Technology_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8127583562806044, + "test_acc": 0.8155932203389831, + "test_auc": 0.8286468617739094, + "val_auc": 0.770330112721417, + "k": 1, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528 + ] + }, + { + "test_f1": 0.8199388486695353, + "test_acc": 0.8223728813559322, + "test_auc": 0.8347618890028791, + "val_auc": 0.7874396135265699, + "k": 2, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791 + ] + }, + { + "test_f1": 0.8209600654803185, + "test_acc": 0.8216949152542373, + "test_auc": 0.898840976198093, + "val_auc": 0.8530595813204508, + "k": 5, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1628, + 1815, + 2023 + ] + }, + { + "test_f1": 0.8283715971850414, + "test_acc": 0.8291525423728814, + "test_auc": 0.9173358999533008, + "val_auc": 0.855475040257649, + "k": 10, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1628, + 1815, + 2023, + 1103, + 203, + 16223, + 1180, + 1866 + ] + }, + { + "test_f1": 0.8317922998680725, + "test_acc": 0.8325423728813559, + "test_auc": 0.9177900225406597, + "val_auc": 0.855877616747182, + "k": 16, + "dataset": "140_news_class_Technology", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1528, + 4791, + 1628, + 1815, + 2023, + 1103, + 203, + 16223, + 1180, + 1866, + 461, + 416, + 4558, + 31149, + 4185, + 20997 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3b068ad3b1f7f0274109e4f6c65973f4e8d54e2a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/141_news_class_Entertainment_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6451309045439563, + "test_acc": 0.6494915254237288, + "test_auc": 0.6882218618658371, + "val_auc": 0.6996779388083736, + "k": 1, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020 + ] + }, + { + "test_f1": 0.7200821307506052, + "test_acc": 0.7213559322033898, + "test_auc": 0.7946612098414064, + "val_auc": 0.8413848631239935, + "k": 2, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266 + ] + }, + { + "test_f1": 0.7871466817123987, + "test_acc": 0.7884745762711864, + "test_auc": 0.874272392656084, + "val_auc": 0.9182769726247988, + "k": 5, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 431, + 1628 + ] + }, + { + "test_f1": 0.7888106357909721, + "test_acc": 0.7898305084745763, + "test_auc": 0.8767351343798377, + "val_auc": 0.9251207729468599, + "k": 10, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 431, + 1628, + 373, + 32278, + 27673, + 16223, + 6918 + ] + }, + { + "test_f1": 0.8035531001684714, + "test_acc": 0.8040677966101695, + "test_auc": 0.8832794637308653, + "val_auc": 0.9325684380032206, + "k": 16, + "dataset": "141_news_class_Entertainment", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5020, + 266, + 1866, + 431, + 1628, + 373, + 32278, + 27673, + 16223, + 6918, + 1758, + 1847, + 508, + 2333, + 1103, + 1634 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a9b6e0b76a10dd65bc7ecbc4f08ab636a9853883 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/142_cancer_cat_Thyroid_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5960105969872169, + "test_acc": 0.6005830903790087, + "test_auc": 0.5849653202774378, + "val_auc": 0.5186074429771909, + "k": 1, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381 + ] + }, + { + "test_f1": 0.5671235590300698, + "test_acc": 0.577259475218659, + "test_auc": 0.6064021487828097, + "val_auc": 0.548219287715086, + "k": 2, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219 + ] + }, + { + "test_f1": 0.5684507765370299, + "test_acc": 0.5714285714285714, + "test_auc": 0.6350809193526452, + "val_auc": 0.5980392156862745, + "k": 5, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219, + 1122, + 799, + 1799 + ] + }, + { + "test_f1": 0.6433874689227137, + "test_acc": 0.6443148688046647, + "test_auc": 0.6735006119951041, + "val_auc": 0.6630652260904362, + "k": 10, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219, + 1122, + 799, + 1799, + 275, + 993, + 304, + 5897, + 673 + ] + }, + { + "test_f1": 0.6792184970878199, + "test_acc": 0.6793002915451894, + "test_auc": 0.7218822249422003, + "val_auc": 0.7470988395358142, + "k": 16, + "dataset": "142_cancer_cat_Thyroid_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4381, + 219, + 1122, + 799, + 1799, + 275, + 993, + 304, + 5897, + 673, + 1697, + 801, + 1508, + 1039, + 1411, + 1152 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..67ba30fe6b5aaccce8fb75cff9f188940404429c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/143_cancer_cat_Lung_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.724166953770739, + "test_acc": 0.7405247813411079, + "test_auc": 0.7431660546715626, + "val_auc": 0.7222889155662265, + "k": 1, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860 + ] + }, + { + "test_f1": 0.7810153915502788, + "test_acc": 0.7871720116618076, + "test_auc": 0.798313613491092, + "val_auc": 0.8125250100040017, + "k": 2, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275 + ] + }, + { + "test_f1": 0.8450780425408725, + "test_acc": 0.8454810495626822, + "test_auc": 0.8839079287365702, + "val_auc": 0.8787515006002401, + "k": 5, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 635, + 191, + 534 + ] + }, + { + "test_f1": 0.8920071014905018, + "test_acc": 0.892128279883382, + "test_auc": 0.9230926152590779, + "val_auc": 0.9115646258503401, + "k": 10, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 635, + 191, + 534, + 1928, + 8006, + 1799, + 506, + 433 + ] + }, + { + "test_f1": 0.9241853619052776, + "test_acc": 0.924198250728863, + "test_auc": 0.9570583435332518, + "val_auc": 0.9371748699479792, + "k": 16, + "dataset": "143_cancer_cat_Lung_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1860, + 275, + 635, + 191, + 534, + 1928, + 8006, + 1799, + 506, + 433, + 1122, + 801, + 799, + 1239, + 821, + 26639 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..32c0d1c78cb255f613b79c9e888c78eb7bc1d153 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/144_cancer_cat_Colon_Cancer_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.684000409773275, + "test_acc": 0.6997084548104956, + "test_auc": 0.703828369373045, + "val_auc": 0.683873549419768, + "k": 1, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 635 + ] + }, + { + "test_f1": 0.7347696035742682, + "test_acc": 0.7376093294460642, + "test_auc": 0.7992316061471508, + "val_auc": 0.7763105242096838, + "k": 2, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 635, + 1860 + ] + }, + { + "test_f1": 0.7839223300585374, + "test_acc": 0.7842565597667639, + "test_auc": 0.8635930912552701, + "val_auc": 0.8577430972388955, + "k": 5, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 635, + 1860, + 191, + 1239, + 5897 + ] + }, + { + "test_f1": 0.8279444625166507, + "test_acc": 0.8279883381924198, + "test_auc": 0.8876478988168095, + "val_auc": 0.883953581432573, + "k": 10, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 635, + 1860, + 191, + 1239, + 5897, + 534, + 275, + 433, + 8006, + 1195 + ] + }, + { + "test_f1": 0.792809211725787, + "test_acc": 0.793002915451895, + "test_auc": 0.8862709098327213, + "val_auc": 0.8939575830332133, + "k": 16, + "dataset": "144_cancer_cat_Colon_Cancer", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 635, + 1860, + 191, + 1239, + 5897, + 534, + 275, + 433, + 8006, + 1195, + 1464, + 3233, + 1596, + 167, + 1799, + 181 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..8701d89c42112273f068a3fb70ad5e265e65b7a1 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/145_disease_class_digestive system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.564322177225403, + "test_acc": 0.5656565656565656, + "test_auc": 0.5712244897959183, + "val_auc": 0.5062626262626263, + "k": 1, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779 + ] + }, + { + "test_f1": 0.3389600705036947, + "test_acc": 0.5050505050505051, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 2, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779, + 320 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 5, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779, + 320, + 1073, + 987, + 859 + ] + }, + { + "test_f1": 0.3277368277368277, + "test_acc": 0.494949494949495, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 10, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779, + 320, + 1073, + 987, + 859, + 1030, + 6534, + 4712, + 680, + 1703 + ] + }, + { + "test_f1": 0.5631585444669557, + "test_acc": 0.5656565656565656, + "test_auc": 0.5359183673469388, + "val_auc": 0.5074747474747474, + "k": 16, + "dataset": "145_disease_class_digestive system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2779, + 320, + 1073, + 987, + 859, + 1030, + 6534, + 4712, + 680, + 1703, + 31245, + 6804, + 473, + 7666, + 798, + 5653 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c331c401b6e4167b938c3a93c06e7f25359c8c56 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/146_disease_class_cardiovascular diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.564322177225403, + "test_acc": 0.5656565656565656, + "test_auc": 0.5742857142857143, + "val_auc": 0.5915151515151514, + "k": 1, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520 + ] + }, + { + "test_f1": 0.5526054331058645, + "test_acc": 0.5757575757575758, + "test_auc": 0.6020408163265306, + "val_auc": 0.662020202020202, + "k": 2, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534 + ] + }, + { + "test_f1": 0.5356125356125355, + "test_acc": 0.5555555555555556, + "test_auc": 0.6034693877551021, + "val_auc": 0.6008080808080809, + "k": 5, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534, + 5883, + 7666, + 2046 + ] + }, + { + "test_f1": 0.5483573274270948, + "test_acc": 0.5555555555555556, + "test_auc": 0.5459183673469388, + "val_auc": 0.582020202020202, + "k": 10, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534, + 5883, + 7666, + 2046, + 1791, + 320, + 1116, + 13612, + 1818 + ] + }, + { + "test_f1": 0.5237940076649754, + "test_acc": 0.5252525252525253, + "test_auc": 0.5416326530612244, + "val_auc": 0.601010101010101, + "k": 16, + "dataset": "146_disease_class_cardiovascular diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 6534, + 5883, + 7666, + 2046, + 1791, + 320, + 1116, + 13612, + 1818, + 1647, + 5129, + 2489, + 4712, + 347, + 2548 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0388d57570b561a748a316dac866448b689bf8a7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5853506965412945, + "test_acc": 0.5858585858585859, + "test_auc": 0.5877551020408164, + "val_auc": 0.5785858585858586, + "k": 1, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666 + ] + }, + { + "test_f1": 0.6055774918319629, + "test_acc": 0.6060606060606061, + "test_auc": 0.5897959183673469, + "val_auc": 0.5981818181818181, + "k": 2, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245 + ] + }, + { + "test_f1": 0.5957120221826104, + "test_acc": 0.5959595959595959, + "test_auc": 0.590204081632653, + "val_auc": 0.5773737373737373, + "k": 5, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 7026, + 4985 + ] + }, + { + "test_f1": 0.5744556861578137, + "test_acc": 0.5757575757575758, + "test_auc": 0.5828571428571427, + "val_auc": 0.7187878787878789, + "k": 10, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 7026, + 4985, + 2779, + 1116, + 111, + 27174, + 6499 + ] + }, + { + "test_f1": 0.5653905596745241, + "test_acc": 0.5656565656565656, + "test_auc": 0.5971428571428571, + "val_auc": 0.6711111111111112, + "k": 16, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7666, + 31245, + 1030, + 7026, + 4985, + 2779, + 1116, + 111, + 27174, + 6499, + 373, + 680, + 2595, + 347, + 1717, + 1011 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ce0d6115fa6b6c3aa16da01c32222b2746b9dee7 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/148_twt_emotion_worry_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5745104132986361, + "test_acc": 0.5749152542372882, + "test_auc": 0.58818159755546, + "val_auc": 0.5221417069243157, + "k": 1, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319 + ] + }, + { + "test_f1": 0.5864406779661017, + "test_acc": 0.5864406779661017, + "test_auc": 0.6091705184351709, + "val_auc": 0.587157809983897, + "k": 2, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 217 + ] + }, + { + "test_f1": 0.5850243274539423, + "test_acc": 0.5884745762711865, + "test_auc": 0.6173529984960636, + "val_auc": 0.5853462157809984, + "k": 5, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 217, + 539, + 1960, + 1541 + ] + }, + { + "test_f1": 0.5678648478044258, + "test_acc": 0.5694915254237288, + "test_auc": 0.6076196989921053, + "val_auc": 0.6099033816425121, + "k": 10, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 217, + 539, + 1960, + 1541, + 186, + 311, + 20, + 836, + 1791 + ] + }, + { + "test_f1": 0.5675452531020186, + "test_acc": 0.568135593220339, + "test_auc": 0.6002590521156229, + "val_auc": 0.6409017713365539, + "k": 16, + "dataset": "148_twt_emotion_worry", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 217, + 539, + 1960, + 1541, + 186, + 311, + 20, + 836, + 1791, + 490, + 6088, + 373, + 1812, + 443, + 6982 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..320f46f8d0e431e0859eba3bc8280906bc337293 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/149_twt_emotion_happiness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6243912417588506, + "test_acc": 0.624406779661017, + "test_auc": 0.6483739837398375, + "val_auc": 0.6674718196457325, + "k": 1, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319 + ] + }, + { + "test_f1": 0.6498808890444132, + "test_acc": 0.6508474576271186, + "test_auc": 0.7025340775795819, + "val_auc": 0.6799516908212561, + "k": 2, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541 + ] + }, + { + "test_f1": 0.6741073784156596, + "test_acc": 0.6752542372881356, + "test_auc": 0.7265906241151963, + "val_auc": 0.711755233494364, + "k": 5, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1791, + 1874 + ] + }, + { + "test_f1": 0.6737439662579083, + "test_acc": 0.6745762711864407, + "test_auc": 0.7339393203972746, + "val_auc": 0.7437600644122383, + "k": 10, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1791, + 1874, + 186, + 1960, + 1335, + 1255, + 908 + ] + }, + { + "test_f1": 0.6870593435794048, + "test_acc": 0.688135593220339, + "test_auc": 0.7354542880571275, + "val_auc": 0.7431561996779388, + "k": 16, + "dataset": "149_twt_emotion_happiness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1319, + 1541, + 20, + 1791, + 1874, + 186, + 1960, + 1335, + 1255, + 908, + 311, + 765, + 253, + 217, + 821, + 1815 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..954d6ac165c233b03f369dd06c3afb787ead6d2d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/150_twt_emotion_sadness_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.596207214566625, + "test_acc": 0.6047457627118644, + "test_auc": 0.612177471842561, + "val_auc": 0.6298309178743962, + "k": 1, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541 + ] + }, + { + "test_f1": 0.5934630155394679, + "test_acc": 0.6033898305084746, + "test_auc": 0.612177471842561, + "val_auc": 0.6298309178743962, + "k": 2, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255 + ] + }, + { + "test_f1": 0.6083593998562475, + "test_acc": 0.6128813559322034, + "test_auc": 0.6403799553599335, + "val_auc": 0.6880032206119162, + "k": 5, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 1791, + 20, + 12485 + ] + }, + { + "test_f1": 0.6198429182647663, + "test_acc": 0.6230508474576271, + "test_auc": 0.6587820321893857, + "val_auc": 0.711352657004831, + "k": 10, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 1791, + 20, + 12485, + 186, + 311, + 521, + 4511, + 5144 + ] + }, + { + "test_f1": 0.6198231489172695, + "test_acc": 0.6216949152542373, + "test_auc": 0.6581100410732736, + "val_auc": 0.712157809983897, + "k": 16, + "dataset": "150_twt_emotion_sadness", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1541, + 1255, + 1791, + 20, + 12485, + 186, + 311, + 521, + 4511, + 5144, + 1942, + 6212, + 1319, + 567, + 1500, + 1596 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..bf907a9328ae1c238cd8ba4b1d7634190ada8c20 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/151_it_tick_HR Support_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5799417236966361, + "test_acc": 0.5806451612903226, + "test_auc": 0.5970026107088321, + "val_auc": 0.5514205682272909, + "k": 1, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067 + ] + }, + { + "test_f1": 0.609599824740427, + "test_acc": 0.6104218362282878, + "test_auc": 0.643145657849367, + "val_auc": 0.5822328931572629, + "k": 2, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268 + ] + }, + { + "test_f1": 0.5856079404466501, + "test_acc": 0.5856079404466501, + "test_auc": 0.6440569430077336, + "val_auc": 0.6172468987595038, + "k": 5, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 5731, + 1441, + 4802 + ] + }, + { + "test_f1": 0.601948146973369, + "test_acc": 0.6029776674937966, + "test_auc": 0.6611496970592581, + "val_auc": 0.6590636254501799, + "k": 10, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 5731, + 1441, + 4802, + 24630, + 1674, + 8006, + 10, + 1619 + ] + }, + { + "test_f1": 0.5973719199195288, + "test_acc": 0.6004962779156328, + "test_auc": 0.661629968967046, + "val_auc": 0.6486594637855142, + "k": 16, + "dataset": "151_it_tick_HR Support", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1067, + 1268, + 5731, + 1441, + 4802, + 24630, + 1674, + 8006, + 10, + 1619, + 898, + 1101, + 1551, + 2, + 710, + 8013 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3c16520e3e1064b129b3a0bf8a1c68bd5ae5bf87 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/152_it_tick_Hardware_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5796594504908162, + "test_acc": 0.607940446650124, + "test_auc": 0.6165582976208068, + "val_auc": 0.6158463385354142, + "k": 1, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575 + ] + }, + { + "test_f1": 0.6282205585007892, + "test_acc": 0.6327543424317618, + "test_auc": 0.6711245751440815, + "val_auc": 0.6624649859943977, + "k": 2, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 1625 + ] + }, + { + "test_f1": 0.6142872412441657, + "test_acc": 0.6153846153846154, + "test_auc": 0.6676764691394512, + "val_auc": 0.6856742697078833, + "k": 5, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 1625, + 583, + 28711, + 1381 + ] + }, + { + "test_f1": 0.5875559775319984, + "test_acc": 0.5880893300248139, + "test_auc": 0.644906654844589, + "val_auc": 0.6842737094837935, + "k": 10, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 1625, + 583, + 28711, + 1381, + 433, + 10, + 822, + 1044, + 1619 + ] + }, + { + "test_f1": 0.5896103832849282, + "test_acc": 0.5905707196029777, + "test_auc": 0.6348948327668587, + "val_auc": 0.7250900360144057, + "k": 16, + "dataset": "152_it_tick_Hardware", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 575, + 1625, + 583, + 28711, + 1381, + 433, + 10, + 822, + 1044, + 1619, + 4802, + 765, + 1551, + 1799, + 434, + 1653 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..857dec470c0257d9bc1e5055a7f20f3afed90526 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/153_it_tick_Administrative rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5549927200988423, + "test_acc": 0.5558312655086849, + "test_auc": 0.5804393872223044, + "val_auc": 0.5610244097639056, + "k": 1, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619 + ] + }, + { + "test_f1": 0.5955334987593052, + "test_acc": 0.5955334987593052, + "test_auc": 0.6241564454952959, + "val_auc": 0.6626650660264105, + "k": 2, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042 + ] + }, + { + "test_f1": 0.6972331858289159, + "test_acc": 0.6972704714640199, + "test_auc": 0.7447908970001478, + "val_auc": 0.7354941976790716, + "k": 5, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 32279, + 575, + 2 + ] + }, + { + "test_f1": 0.6817118060047275, + "test_acc": 0.6823821339950372, + "test_auc": 0.7401605832225013, + "val_auc": 0.7507002801120447, + "k": 10, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 32279, + 575, + 2, + 1067, + 125, + 10, + 1625, + 7319 + ] + }, + { + "test_f1": 0.7021011772506556, + "test_acc": 0.7022332506203474, + "test_auc": 0.7612186591793508, + "val_auc": 0.7122849139655862, + "k": 16, + "dataset": "153_it_tick_Administrative rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1619, + 1042, + 32279, + 575, + 2, + 1067, + 125, + 10, + 1625, + 7319, + 1662, + 1872, + 25533, + 772, + 30127, + 549 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..107d8633243917b6b50ae4c3185f12f2d9eac8ed --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/154_athlete_sport_football_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.831585539474205, + "test_acc": 0.8340026773761714, + "test_auc": 0.8491347794296855, + "val_auc": 0.8470707070707072, + "k": 1, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731 + ] + }, + { + "test_f1": 0.8376398152025932, + "test_acc": 0.8406961178045516, + "test_auc": 0.9049762727416093, + "val_auc": 0.8965656565656567, + "k": 2, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648 + ] + }, + { + "test_f1": 0.8622957908329483, + "test_acc": 0.8634538152610441, + "test_auc": 0.9513949620793968, + "val_auc": 0.9648484848484848, + "k": 5, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648, + 23226, + 1158, + 737 + ] + }, + { + "test_f1": 0.868643307078443, + "test_acc": 0.8688085676037484, + "test_auc": 0.9533841808719589, + "val_auc": 0.9684848484848485, + "k": 10, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648, + 23226, + 1158, + 737, + 6567, + 805, + 1274, + 23697, + 9346 + ] + }, + { + "test_f1": 0.8699710814784704, + "test_acc": 0.8701472556894244, + "test_auc": 0.9539003025046235, + "val_auc": 0.9696969696969697, + "k": 16, + "dataset": "154_athlete_sport_football", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4731, + 4648, + 23226, + 1158, + 737, + 6567, + 805, + 1274, + 23697, + 9346, + 13701, + 5872, + 1499, + 3733, + 1447, + 1708 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5f6607fd44150bcce6b69d8db05e1491b5bbd84f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/155_athlete_sport_basketball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8681255466750192, + "test_acc": 0.8701472556894244, + "test_auc": 0.8713674355923213, + "val_auc": 0.8755555555555555, + "k": 1, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226 + ] + }, + { + "test_f1": 0.8681255466750192, + "test_acc": 0.8701472556894244, + "test_auc": 0.9219150979914266, + "val_auc": 0.9260606060606061, + "k": 2, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648 + ] + }, + { + "test_f1": 0.874567002631305, + "test_acc": 0.8755020080321285, + "test_auc": 0.9643087554300296, + "val_auc": 0.9634343434343434, + "k": 5, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648, + 4731, + 32, + 4505 + ] + }, + { + "test_f1": 0.8706655347803586, + "test_acc": 0.8714859437751004, + "test_auc": 0.9658212785479778, + "val_auc": 0.9595959595959596, + "k": 10, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648, + 4731, + 32, + 4505, + 1065, + 6567, + 1447, + 805, + 1682 + ] + }, + { + "test_f1": 0.8733600028057678, + "test_acc": 0.8741633199464525, + "test_auc": 0.9654700291035254, + "val_auc": 0.9593939393939395, + "k": 16, + "dataset": "155_athlete_sport_basketball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23226, + 4648, + 4731, + 32, + 4505, + 1065, + 6567, + 1447, + 805, + 1682, + 1643, + 7828, + 4387, + 176, + 1380, + 1274 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3b9ae6da507e8393ce4e514b25bcc4ba60081430 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/156_athlete_sport_baseball_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8690462698002057, + "test_acc": 0.8701472556894244, + "test_auc": 0.8845034479792404, + "val_auc": 0.8868686868686868, + "k": 1, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648 + ] + }, + { + "test_f1": 0.8575230850233125, + "test_acc": 0.85809906291834, + "test_auc": 0.9240691889722011, + "val_auc": 0.915959595959596, + "k": 2, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 23226 + ] + }, + { + "test_f1": 0.8968189430758178, + "test_acc": 0.8969210174029452, + "test_auc": 0.9760218491491162, + "val_auc": 0.9878787878787878, + "k": 5, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 23226, + 4505, + 32, + 4731 + ] + }, + { + "test_f1": 0.8942265809721436, + "test_acc": 0.8942436412315931, + "test_auc": 0.9741975025447663, + "val_auc": 0.9846464646464647, + "k": 10, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 23226, + 4505, + 32, + 4731, + 2832, + 1682, + 6828, + 1305, + 1065 + ] + }, + { + "test_f1": 0.9036020193822134, + "test_acc": 0.9036144578313253, + "test_auc": 0.9757566199767745, + "val_auc": 0.9818181818181819, + "k": 16, + "dataset": "156_athlete_sport_baseball", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4648, + 23226, + 4505, + 32, + 4731, + 2832, + 1682, + 6828, + 1305, + 1065, + 1101, + 1158, + 152, + 5872, + 2302, + 2960 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..46f2edeb696851ca328fca343f79f6dacdd6b7cc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7630701317041225, + "test_acc": 0.7723076923076924, + "test_auc": 0.7778322449254385, + "val_auc": 0.7674418604651163, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518 + ] + }, + { + "test_f1": 0.7768442754005693, + "test_acc": 0.7835897435897435, + "test_auc": 0.8017975561315515, + "val_auc": 0.777641778865769, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541 + ] + }, + { + "test_f1": 0.7940878005721161, + "test_acc": 0.7979487179487179, + "test_auc": 0.8885784495236813, + "val_auc": 0.8588331293349654, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 1791, + 186, + 1335 + ] + }, + { + "test_f1": 0.7942804693771546, + "test_acc": 0.796923076923077, + "test_auc": 0.8859464940923013, + "val_auc": 0.875968992248062, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 1791, + 186, + 1335, + 821, + 2581, + 253, + 7813, + 2203 + ] + }, + { + "test_f1": 0.8202972941037457, + "test_acc": 0.8225641025641026, + "test_auc": 0.9050981250210388, + "val_auc": 0.8829049367605059, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 11518, + 1541, + 1791, + 186, + 1335, + 821, + 2581, + 253, + 7813, + 2203, + 549, + 1030, + 7867, + 1438, + 1361, + 1417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7c844ae38c5bd8eeef1a2b179d5b3ddbdffbb717 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/158_code_C_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8444771626021648, + "test_acc": 0.8473949579831933, + "test_auc": 0.8507047638674968, + "val_auc": 0.8538739462063427, + "k": 1, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852 + ] + }, + { + "test_f1": 0.901331027381649, + "test_acc": 0.9018487394957984, + "test_auc": 0.9123031325248931, + "val_auc": 0.9409875551987154, + "k": 2, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782 + ] + }, + { + "test_f1": 0.9112478596627563, + "test_acc": 0.9115966386554621, + "test_auc": 0.9616210111287068, + "val_auc": 0.9863508631071859, + "k": 5, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 173, + 60 + ] + }, + { + "test_f1": 0.918443969541621, + "test_acc": 0.9186554621848739, + "test_auc": 0.9733162317142836, + "val_auc": 0.9855479727017261, + "k": 10, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 173, + 60, + 1275, + 1521, + 8815, + 259, + 1688 + ] + }, + { + "test_f1": 0.9161509233728821, + "test_acc": 0.9163025210084034, + "test_auc": 0.9772011555343442, + "val_auc": 0.9855479727017262, + "k": 16, + "dataset": "158_code_C", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 852, + 1782, + 1318, + 173, + 60, + 1275, + 1521, + 8815, + 259, + 1688, + 1283, + 11077, + 228, + 42, + 325, + 6919 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4d5e9400999ad85a33fccf3a574850d765554370 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.757311288129863, + "test_acc": 0.7630252100840336, + "test_auc": 0.7815480580804246, + "val_auc": 0.7972701726214372, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021 + ] + }, + { + "test_f1": 0.7734416937997142, + "test_acc": 0.7741176470588236, + "test_auc": 0.8811979810689055, + "val_auc": 0.8900040144520274, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914 + ] + }, + { + "test_f1": 0.7781872367595614, + "test_acc": 0.7794957983193277, + "test_auc": 0.8911999877070814, + "val_auc": 0.9177037334403854, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1180, + 1493, + 6919 + ] + }, + { + "test_f1": 0.8071557274604708, + "test_acc": 0.8077310924369748, + "test_auc": 0.9038616938195544, + "val_auc": 0.9305499799277399, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1180, + 1493, + 6919, + 4790, + 1275, + 1067, + 1794, + 1091 + ] + }, + { + "test_f1": 0.835242813527289, + "test_acc": 0.8352941176470589, + "test_auc": 0.9296831500242242, + "val_auc": 0.960256924929747, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1021, + 1914, + 1180, + 1493, + 6919, + 4790, + 1275, + 1067, + 1794, + 1091, + 2962, + 98, + 1500, + 727, + 347, + 964 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..477f99d7cb7091e7ae1c9f191ec763011a05726c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7473932601763248, + "test_acc": 0.7573109243697479, + "test_auc": 0.7639429717045939, + "val_auc": 0.7476916900843035, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 764 + ] + }, + { + "test_f1": 0.7916028105607112, + "test_acc": 0.7959663865546218, + "test_auc": 0.8128495798714305, + "val_auc": 0.8018867924528301, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 764, + 4235 + ] + }, + { + "test_f1": 0.8035085877773488, + "test_acc": 0.8067226890756303, + "test_auc": 0.8566010712916966, + "val_auc": 0.83701324769169, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 764, + 4235, + 3494, + 5189, + 692 + ] + }, + { + "test_f1": 0.8293294481879439, + "test_acc": 0.8305882352941176, + "test_auc": 0.921906523201076, + "val_auc": 0.9040545965475713, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 764, + 4235, + 3494, + 5189, + 692, + 17338, + 42, + 8189, + 2046, + 60 + ] + }, + { + "test_f1": 0.824651610227343, + "test_acc": 0.825546218487395, + "test_auc": 0.9259261267906083, + "val_auc": 0.9203131272581293, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 764, + 4235, + 3494, + 5189, + 692, + 17338, + 42, + 8189, + 2046, + 60, + 576, + 964, + 30916, + 32655, + 22147, + 1980 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6936890fa18d2ff03b4aab7838a7d5b92883e420 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/161_agnews_0_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9472261036747394, + "test_acc": 0.9472268907563025, + "test_auc": 0.9869509765639122, + "val_auc": 0.9867523083099157, + "k": 1, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9865541569753477, + "test_acc": 0.9865546218487395, + "test_auc": 0.999407951348967, + "val_auc": 0.9991971095945402, + "k": 2, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847 + ] + }, + { + "test_f1": 0.9889068109997484, + "test_acc": 0.9889075630252101, + "test_auc": 0.9995701997960822, + "val_auc": 0.9991971095945403, + "k": 5, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1411, + 786 + ] + }, + { + "test_f1": 0.9969747653056478, + "test_acc": 0.9969747899159664, + "test_auc": 0.9999502859911347, + "val_auc": 1.0, + "k": 10, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1411, + 786, + 461, + 1697, + 431, + 279, + 5020 + ] + }, + { + "test_f1": 0.9973109237620901, + "test_acc": 0.9973109243697479, + "test_auc": 0.9999887013616215, + "val_auc": 1.0, + "k": 16, + "dataset": "161_agnews_0", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1847, + 1703, + 1411, + 786, + 461, + 1697, + 431, + 279, + 5020, + 1634, + 1360, + 27673, + 1758, + 1472, + 126 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..89bc4cc987c5d81867c350fd7bf78a3403d0c15d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/162_agnews_1_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9455458739435183, + "test_acc": 0.945546218487395, + "test_auc": 0.9880182459451446, + "val_auc": 0.9891609795262946, + "k": 1, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9610015295704211, + "test_acc": 0.9610084033613445, + "test_auc": 0.9952132188645681, + "val_auc": 0.9967884383781614, + "k": 2, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9653618706502495, + "test_acc": 0.9653781512605042, + "test_auc": 0.9967392129639673, + "val_auc": 0.9971898835808912, + "k": 5, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 431 + ] + }, + { + "test_f1": 0.9875617182331379, + "test_acc": 0.987563025210084, + "test_auc": 0.9987788431640526, + "val_auc": 0.9991971095945404, + "k": 10, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 431, + 1758, + 1847, + 279, + 1628, + 126 + ] + }, + { + "test_f1": 0.9986554618810449, + "test_acc": 0.998655462184874, + "test_auc": 0.9999882494160863, + "val_auc": 1.0, + "k": 16, + "dataset": "162_agnews_1", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 1411, + 431, + 1758, + 1847, + 279, + 1628, + 126, + 1551, + 1472, + 1634, + 1360, + 1914, + 461 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5404ef112ecb8e3bb15b73e977507bc316e7e79b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.929411429732669, + "test_acc": 0.9294117647058824, + "test_auc": 0.980863270205581, + "val_auc": 0.9807306302689682, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680 + ] + }, + { + "test_f1": 0.9616769051922055, + "test_acc": 0.9616806722689075, + "test_auc": 0.9954086853085161, + "val_auc": 0.9971898835808912, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703 + ] + }, + { + "test_f1": 0.9973109073552354, + "test_acc": 0.9973109243697479, + "test_auc": 0.9997473624458569, + "val_auc": 1.0, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 461, + 1847 + ] + }, + { + "test_f1": 0.9966386212811212, + "test_acc": 0.9966386554621849, + "test_auc": 0.9997387754806892, + "val_auc": 1.0, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 461, + 1847, + 1411, + 279, + 431, + 1758, + 1360 + ] + }, + { + "test_f1": 0.9983193239332243, + "test_acc": 0.9983193277310924, + "test_auc": 0.9999905091437621, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 680, + 1703, + 786, + 461, + 1847, + 1411, + 279, + 431, + 1758, + 1360, + 1697, + 27673, + 1634, + 1472, + 1914, + 5020 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0779296155ef9162e6e140d46192061fd68e697b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/21_headline_istrump_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7678842093540024, + "test_acc": 0.7690566037735849, + "test_auc": 0.828130585222457, + "val_auc": 0.8403361344537815, + "k": 1, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559 + ] + }, + { + "test_f1": 0.8100682374461682, + "test_acc": 0.810566037735849, + "test_auc": 0.8738714369761599, + "val_auc": 0.9087635054021608, + "k": 2, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585 + ] + }, + { + "test_f1": 0.8372383903969872, + "test_acc": 0.8374842767295597, + "test_auc": 0.9050095743054201, + "val_auc": 0.9567827130852341, + "k": 5, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 20310, + 776, + 32700 + ] + }, + { + "test_f1": 0.8602243885881374, + "test_acc": 0.8603773584905661, + "test_auc": 0.9260951972529692, + "val_auc": 0.954781912765106, + "k": 10, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 20310, + 776, + 32700, + 508, + 431, + 1866, + 400, + 1327 + ] + }, + { + "test_f1": 0.8637635700560867, + "test_acc": 0.8638993710691824, + "test_auc": 0.9289426797321422, + "val_auc": 0.9603841536614646, + "k": 16, + "dataset": "21_headline_istrump", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 559, + 1585, + 20310, + 776, + 32700, + 508, + 431, + 1866, + 400, + 1327, + 3832, + 5888, + 1499, + 7976, + 28521, + 1608 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7441e2bfcaff87d5387feccb10a6a9758a3c4abc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6401262768268545, + "test_acc": 0.6450079239302694, + "test_auc": 0.6613019891500904, + "val_auc": 0.7066826730692277, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28521 + ] + }, + { + "test_f1": 0.6704546751616374, + "test_acc": 0.6719492868462758, + "test_auc": 0.7232167972674303, + "val_auc": 0.7354941976790716, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28521, + 431 + ] + }, + { + "test_f1": 0.7748721720915382, + "test_acc": 0.7749603803486529, + "test_auc": 0.8527225236085996, + "val_auc": 0.8743497398959584, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28521, + 431, + 913, + 776, + 508 + ] + }, + { + "test_f1": 0.7874934907272458, + "test_acc": 0.7876386687797148, + "test_auc": 0.8757132810930279, + "val_auc": 0.8811524609843937, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28521, + 431, + 913, + 776, + 508, + 1327, + 20310, + 1042, + 5186, + 562 + ] + }, + { + "test_f1": 0.8145381089825003, + "test_acc": 0.8145800316957211, + "test_auc": 0.9093932087602974, + "val_auc": 0.921968787515006, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 28521, + 431, + 913, + 776, + 508, + 1327, + 20310, + 1042, + 5186, + 562, + 2023, + 272, + 3507, + 1866, + 1288, + 559 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..22e30607dcdf5989fd0cbcae64cabc17d3d9353a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/23_headline_ischina_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6994224941708784, + "test_acc": 0.6998799519807923, + "test_auc": 0.7077482244973251, + "val_auc": 0.7342995169082126, + "k": 1, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.7022628906284063, + "test_acc": 0.7022809123649459, + "test_auc": 0.7722744881018262, + "val_auc": 0.8192431561996779, + "k": 2, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1528 + ] + }, + { + "test_f1": 0.7610975559183988, + "test_acc": 0.7611044417767107, + "test_auc": 0.832606991330013, + "val_auc": 0.7918679549114332, + "k": 5, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1528, + 1447, + 6501, + 32700 + ] + }, + { + "test_f1": 0.7622117082126969, + "test_acc": 0.7623049219687875, + "test_auc": 0.8434214167127836, + "val_auc": 0.8305152979066022, + "k": 10, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1528, + 1447, + 6501, + 32700, + 5186, + 1417, + 347, + 776, + 559 + ] + }, + { + "test_f1": 0.792313335051649, + "test_acc": 0.7923169267707083, + "test_auc": 0.8598736395498985, + "val_auc": 0.8200483091787439, + "k": 16, + "dataset": "23_headline_ischina", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1528, + 1447, + 6501, + 32700, + 5186, + 1417, + 347, + 776, + 559, + 1866, + 10642, + 5020, + 2837, + 6718, + 3507 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4941b7fa5ecc993700344cb0800bb3ca2d8414b0 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7272106181718806, + "test_acc": 0.7272727272727273, + "test_auc": 0.7495446542407869, + "val_auc": 0.7162865146058425, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866 + ] + }, + { + "test_f1": 0.7245179063360881, + "test_acc": 0.7245179063360881, + "test_auc": 0.7665594074433855, + "val_auc": 0.7675070028011205, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186 + ] + }, + { + "test_f1": 0.7190039995163049, + "test_acc": 0.71900826446281, + "test_auc": 0.780174245643859, + "val_auc": 0.7735094037615047, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 3507, + 1681, + 1528 + ] + }, + { + "test_f1": 0.7299660716066088, + "test_acc": 0.7300275482093664, + "test_auc": 0.8001032117054216, + "val_auc": 0.7719087635054022, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 3507, + 1681, + 1528, + 1447, + 12612, + 6501, + 3301, + 508 + ] + }, + { + "test_f1": 0.7437977641763308, + "test_acc": 0.743801652892562, + "test_auc": 0.8043531054580777, + "val_auc": 0.7735094037615046, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1866, + 5186, + 3507, + 1681, + 1528, + 1447, + 12612, + 6501, + 3301, + 508, + 1377, + 601, + 2247, + 1417, + 4336, + 2135 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..4a6a2d6db5b6f995ac7189c2b9fbe1410073668e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/26_headline_isfrontpage_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6199998691998356, + "test_acc": 0.6274213836477988, + "test_auc": 0.6347615132161869, + "val_auc": 0.6558623449379751, + "k": 1, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.6428083605339663, + "test_acc": 0.6430188679245283, + "test_auc": 0.6917792613760064, + "val_auc": 0.7282913165266106, + "k": 2, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585 + ] + }, + { + "test_f1": 0.6706760669698467, + "test_acc": 0.6706918238993711, + "test_auc": 0.726179674929294, + "val_auc": 0.7464985994397759, + "k": 5, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 4605, + 2837, + 27673 + ] + }, + { + "test_f1": 0.6769799872402072, + "test_acc": 0.6769811320754717, + "test_auc": 0.7376332732175641, + "val_auc": 0.7961184473789517, + "k": 10, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 4605, + 2837, + 27673, + 776, + 126, + 32700, + 4674, + 3832 + ] + }, + { + "test_f1": 0.6759705362128402, + "test_acc": 0.6759748427672956, + "test_auc": 0.737890857981305, + "val_auc": 0.8191276510604242, + "k": 16, + "dataset": "26_headline_isfrontpage", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1585, + 4605, + 2837, + 27673, + 776, + 126, + 32700, + 4674, + 3832, + 1804, + 20310, + 2135, + 10642, + 1499, + 1866 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c26e05533bb8d6d9b6109d13dd99cb2a14157e03 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8011847479998445, + "test_acc": 0.8020125786163522, + "test_auc": 0.8435560519635175, + "val_auc": 0.8327330932372948, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293 + ] + }, + { + "test_f1": 0.8095184952926592, + "test_acc": 0.809559748427673, + "test_auc": 0.8919819875468209, + "val_auc": 0.8923569427771109, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062 + ] + }, + { + "test_f1": 0.8271589599361426, + "test_acc": 0.8271698113207547, + "test_auc": 0.9106501110335895, + "val_auc": 0.9375750300120048, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1166 + ] + }, + { + "test_f1": 0.8406702952538312, + "test_acc": 0.840754716981132, + "test_auc": 0.9207256624801652, + "val_auc": 0.9519807923169268, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1166, + 1305, + 1441, + 1116, + 1313, + 409 + ] + }, + { + "test_f1": 0.8432278751489318, + "test_acc": 0.8432704402515723, + "test_auc": 0.9234954265097379, + "val_auc": 0.9571828731492596, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 293, + 1062, + 1429, + 1221, + 1166, + 1305, + 1441, + 1116, + 1313, + 409, + 1928, + 524, + 765, + 14, + 14876, + 396 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a683e33a8e85bf7893e6f7631f63a426853e4924 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/41_truthqa_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6407536062708477, + "test_acc": 0.6464646464646465, + "test_auc": 0.6599999999999999, + "val_auc": 0.6065837013247692, + "k": 1, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377 + ] + }, + { + "test_f1": 0.6387350864185846, + "test_acc": 0.6464646464646465, + "test_auc": 0.6904081632653061, + "val_auc": 0.6501405058209554, + "k": 2, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027 + ] + }, + { + "test_f1": 0.6453785163462583, + "test_acc": 0.6464646464646465, + "test_auc": 0.7246938775510204, + "val_auc": 0.6824568446407064, + "k": 5, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339 + ] + }, + { + "test_f1": 0.6211540694299315, + "test_acc": 0.6262626262626263, + "test_auc": 0.7167346938775511, + "val_auc": 0.73243677238057, + "k": 10, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339, + 100, + 1551, + 681, + 524, + 505 + ] + }, + { + "test_f1": 0.6168653533766157, + "test_acc": 0.6262626262626263, + "test_auc": 0.7391836734693877, + "val_auc": 0.7430750702529104, + "k": 16, + "dataset": "41_truthqa_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1377, + 2027, + 238, + 473, + 1339, + 100, + 1551, + 681, + 524, + 505, + 272, + 1202, + 1790, + 334, + 160, + 2043 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9bdd23769cfd681b62f77a7b89414c851f9f86c3 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6727326253949535, + "test_acc": 0.6755629691409508, + "test_auc": 0.698831385642738, + "val_auc": 0.8103864734299517, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6751018609576207, + "test_acc": 0.6805671392827356, + "test_auc": 0.7240345019476906, + "val_auc": 0.8564814814814814, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114 + ] + }, + { + "test_f1": 0.7250013463012397, + "test_acc": 0.7289407839866555, + "test_auc": 0.7819643850862549, + "val_auc": 0.8675523349436394, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 30704 + ] + }, + { + "test_f1": 0.7643226617923177, + "test_acc": 0.7664720600500416, + "test_auc": 0.8521424596549805, + "val_auc": 0.8909017713365539, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 30704, + 801, + 3276, + 238, + 6086, + 4481 + ] + }, + { + "test_f1": 0.790706455461578, + "test_acc": 0.79232693911593, + "test_auc": 0.8701585976627713, + "val_auc": 0.9299516908212561, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1114, + 486, + 1062, + 30704, + 801, + 3276, + 238, + 6086, + 4481, + 363, + 1335, + 6719, + 4534, + 7400, + 2744 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c36ad8e4f1d30e21149922af96615302e2b48a41 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/44_phys_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.3334731051125547, + "test_acc": 0.500125786163522, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 1, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.5727463678183866, + "test_acc": 0.5781132075471698, + "test_auc": 0.592366731845527, + "val_auc": 0.5260104041616647, + "k": 2, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 588 + ] + }, + { + "test_f1": 0.6121127663704543, + "test_acc": 0.6150943396226415, + "test_auc": 0.6705028864682813, + "val_auc": 0.5682272909163666, + "k": 5, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 588, + 1960, + 3727, + 486 + ] + }, + { + "test_f1": 0.6274314510866236, + "test_acc": 0.629182389937107, + "test_auc": 0.6913950993327858, + "val_auc": 0.5986394557823129, + "k": 10, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 588, + 1960, + 3727, + 486, + 293, + 1928, + 30704, + 1438, + 8609 + ] + }, + { + "test_f1": 0.6294108382112915, + "test_acc": 0.6304402515723271, + "test_auc": 0.6932946445659361, + "val_auc": 0.6024409763905563, + "k": 16, + "dataset": "44_phys_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 588, + 1960, + 3727, + 486, + 293, + 1928, + 30704, + 1438, + 8609, + 7143, + 765, + 7614, + 1739, + 3589, + 238 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e19bb17869dafa6694af64d26f119d02f6a56a43 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5101301712615386, + "test_acc": 0.5362058647516457, + "test_auc": 0.5416253617167579, + "val_auc": 0.5455640305098355, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498 + ] + }, + { + "test_f1": 0.3330009174868407, + "test_acc": 0.49970077797725915, + "test_auc": 0.5, + "val_auc": 0.5, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727 + ] + }, + { + "test_f1": 0.5489758839327065, + "test_acc": 0.5493716337522442, + "test_auc": 0.5750136091453457, + "val_auc": 0.5030108390204737, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 293, + 150, + 860 + ] + }, + { + "test_f1": 0.5360613102534078, + "test_acc": 0.5362058647516457, + "test_auc": 0.5671546858436238, + "val_auc": 0.5523885989562425, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 293, + 150, + 860, + 1336, + 378, + 26677, + 672, + 100 + ] + }, + { + "test_f1": 0.5643121383250304, + "test_acc": 0.5643327348892878, + "test_auc": 0.5752514110534911, + "val_auc": 0.5644319550381373, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7498, + 3727, + 293, + 150, + 860, + 1336, + 378, + 26677, + 672, + 100, + 844, + 1130, + 1641, + 2797, + 22608, + 874 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f9008609d6ab797ab5c87c147a4e2435f03f64cc --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/48_cm_correct_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5514990176396354, + "test_acc": 0.5690566037735849, + "test_auc": 0.5835467004341096, + "val_auc": 0.6486594637855142, + "k": 1, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.5676499243869692, + "test_acc": 0.5738364779874214, + "test_auc": 0.5998256777707008, + "val_auc": 0.6994797919167666, + "k": 2, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1345 + ] + }, + { + "test_f1": 0.5985793916802752, + "test_acc": 0.6158490566037735, + "test_auc": 0.6256259246470266, + "val_auc": 0.7553021208483394, + "k": 5, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1345, + 6659, + 7143, + 13203 + ] + }, + { + "test_f1": 0.599420537288073, + "test_acc": 0.610817610062893, + "test_auc": 0.6620060068513749, + "val_auc": 0.7679071628651459, + "k": 10, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1345, + 6659, + 7143, + 13203, + 1377, + 6400, + 1327, + 777, + 121 + ] + }, + { + "test_f1": 0.6263931044656049, + "test_acc": 0.6264150943396226, + "test_auc": 0.7074483893800649, + "val_auc": 0.8307322929171669, + "k": 16, + "dataset": "48_cm_correct", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 1345, + 6659, + 7143, + 13203, + 1377, + 6400, + 1327, + 777, + 121, + 1224, + 598, + 1596, + 1433, + 448, + 19189 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6ddd874791e31ada151d13690bfc3f8be12144b9 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/49_cm_isshort_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9984905641271326, + "test_acc": 0.9984905660377359, + "test_auc": 0.9997306435492673, + "val_auc": 1.0, + "k": 1, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625 + ] + }, + { + "test_f1": 0.9984905641271326, + "test_acc": 0.9984905660377359, + "test_auc": 0.9994790079176619, + "val_auc": 1.0, + "k": 2, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625, + 680 + ] + }, + { + "test_f1": 0.9992452829233379, + "test_acc": 0.999245283018868, + "test_auc": 0.999981013408078, + "val_auc": 1.0, + "k": 5, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625, + 680, + 1500, + 1703, + 1525 + ] + }, + { + "test_f1": 0.9984905620254628, + "test_acc": 0.9984905660377359, + "test_auc": 0.9999916458995544, + "val_auc": 1.0, + "k": 10, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625, + 680, + 1500, + 1703, + 1525, + 1551, + 347, + 1232, + 6718, + 1255 + ] + }, + { + "test_f1": 0.9994968551548519, + "test_acc": 0.9994968553459119, + "test_auc": 0.9999967089907336, + "val_auc": 1.0, + "k": 16, + "dataset": "49_cm_isshort", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1625, + 680, + 1500, + 1703, + 1525, + 1551, + 347, + 1232, + 6718, + 1255, + 1042, + 1335, + 549, + 1758, + 11655, + 1472 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d940b831cfd220a441908a04e510113506c3eb96 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/50_deon_isvalid_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6462808013522341, + "test_acc": 0.6462893081761006, + "test_auc": 0.692271900147741, + "val_auc": 0.7609043617446979, + "k": 1, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6591557428704151, + "test_acc": 0.659874213836478, + "test_auc": 0.7197265120668652, + "val_auc": 0.8191276510604242, + "k": 2, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143 + ] + }, + { + "test_f1": 0.6882958502224759, + "test_acc": 0.6883018867924529, + "test_auc": 0.7562639298296067, + "val_auc": 0.8507402961184474, + "k": 5, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1928, + 3780, + 5827 + ] + }, + { + "test_f1": 0.6847421244690326, + "test_acc": 0.6847798742138365, + "test_auc": 0.7598278397106342, + "val_auc": 0.8415366146458584, + "k": 10, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1928, + 3780, + 5827, + 6400, + 56, + 152, + 7614, + 1255 + ] + }, + { + "test_f1": 0.7069174599446536, + "test_acc": 0.7069182389937106, + "test_auc": 0.7844247163909477, + "val_auc": 0.8447378951580633, + "k": 16, + "dataset": "50_deon_isvalid", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1928, + 3780, + 5827, + 6400, + 56, + 152, + 7614, + 1255, + 4460, + 1691, + 1327, + 1736, + 1396, + 406 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..106fef2762fe0eddcfde7acba5bb6e69e6bc4954 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/51_just_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6671530734095308, + "test_acc": 0.6689308176100629, + "test_auc": 0.7148520210341059, + "val_auc": 0.7226890756302521, + "k": 1, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794 + ] + }, + { + "test_f1": 0.6654701385414821, + "test_acc": 0.6681761006289308, + "test_auc": 0.7148520210341059, + "val_auc": 0.7226890756302521, + "k": 2, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143 + ] + }, + { + "test_f1": 0.6888270932266418, + "test_acc": 0.690062893081761, + "test_auc": 0.7619349716821311, + "val_auc": 0.7370948379351742, + "k": 5, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 1928, + 1011 + ] + }, + { + "test_f1": 0.7114734114860636, + "test_acc": 0.7119496855345911, + "test_auc": 0.7832014735620567, + "val_auc": 0.7523009203681472, + "k": 10, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 1928, + 1011, + 19940, + 16883, + 598, + 30704, + 7614 + ] + }, + { + "test_f1": 0.7324109553699342, + "test_acc": 0.7325786163522012, + "test_auc": 0.8015577106321878, + "val_auc": 0.7863145258103241, + "k": 16, + "dataset": "51_just_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 794, + 7143, + 1360, + 1928, + 1011, + 19940, + 16883, + 598, + 30704, + 7614, + 1960, + 7068, + 253, + 6400, + 5827, + 836 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..125baca287432498b6f4e13bfd1be4ef200e8f41 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/52_virtue_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7719244277438061, + "test_acc": 0.7721419185282523, + "test_auc": 0.8023139069254204, + "val_auc": 0.7223999999999999, + "k": 1, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062 + ] + }, + { + "test_f1": 0.7930365424951341, + "test_acc": 0.7942181340341655, + "test_auc": 0.8700645889245658, + "val_auc": 0.7912, + "k": 2, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 811 + ] + }, + { + "test_f1": 0.8212675332391279, + "test_acc": 0.8212877792378449, + "test_auc": 0.8942601559439327, + "val_auc": 0.8, + "k": 5, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 811, + 7319, + 1662, + 1030 + ] + }, + { + "test_f1": 0.8356922601312297, + "test_acc": 0.835742444152431, + "test_auc": 0.9065079599260231, + "val_auc": 0.8424, + "k": 10, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 811, + 7319, + 1662, + 1030, + 1248, + 5448, + 3892, + 1313, + 628 + ] + }, + { + "test_f1": 0.8420476432411987, + "test_acc": 0.8420499342969776, + "test_auc": 0.9162095600891391, + "val_auc": 0.8787999999999999, + "k": 16, + "dataset": "52_virtue_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1062, + 811, + 7319, + 1662, + 1030, + 1248, + 5448, + 3892, + 1313, + 628, + 4334, + 2046, + 296, + 1942, + 6778, + 1090 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..85d1132eb685c92bafcd83ed3b0f3072cb574b0d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/54_cs_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5171079402615442, + "test_acc": 0.5212578616352201, + "test_auc": 0.533683986151433, + "val_auc": 0.6044417767106843, + "k": 1, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011 + ] + }, + { + "test_f1": 0.5204662900069686, + "test_acc": 0.5212578616352201, + "test_auc": 0.5386331577790852, + "val_auc": 0.6506602641056423, + "k": 2, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411 + ] + }, + { + "test_f1": 0.5204662900069686, + "test_acc": 0.5212578616352201, + "test_auc": 0.5386331577790852, + "val_auc": 0.6506602641056423, + "k": 5, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928 + ] + }, + { + "test_f1": 0.5663355304045627, + "test_acc": 0.5665408805031447, + "test_auc": 0.5978867670036323, + "val_auc": 0.6522609043617447, + "k": 10, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928, + 1703, + 786, + 1202, + 15467, + 6718 + ] + }, + { + "test_f1": 0.5204662900069686, + "test_acc": 0.5212578616352201, + "test_auc": 0.5386331577790852, + "val_auc": 0.6506602641056423, + "k": 16, + "dataset": "54_cs_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1011, + 1411, + 1377, + 800, + 1928, + 1703, + 786, + 1202, + 15467, + 6718, + 1417, + 334, + 1596, + 448, + 1643, + 1791 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f41f4fd0e6ec73a28a500936ed028479c42a7649 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/56_wikidatasex_or_gender_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9932074156222817, + "test_acc": 0.9932075471698113, + "test_auc": 0.9971420116066303, + "val_auc": 1.0, + "k": 1, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.9939622045395757, + "test_acc": 0.9939622641509434, + "test_auc": 0.9986849886434864, + "val_auc": 1.0, + "k": 2, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 6475 + ] + }, + { + "test_f1": 0.9937106392819116, + "test_acc": 0.9937106918238994, + "test_auc": 0.9994696411989805, + "val_auc": 1.0, + "k": 5, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 6475, + 1744, + 487, + 152 + ] + }, + { + "test_f1": 0.9916975919619981, + "test_acc": 0.9916981132075472, + "test_auc": 0.9993155966498538, + "val_auc": 1.0, + "k": 10, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 6475, + 1744, + 487, + 152, + 6935, + 292, + 26988, + 821, + 334 + ] + }, + { + "test_f1": 0.992955822373378, + "test_acc": 0.9929559748427673, + "test_auc": 0.999442553661172, + "val_auc": 1.0, + "k": 16, + "dataset": "56_wikidatasex_or_gender", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 6475, + 1744, + 487, + 152, + 6935, + 292, + 26988, + 821, + 334, + 559, + 347, + 3417, + 1960, + 7368, + 1221 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..6b64a90088b7373da7e921468d1bb0792c90af78 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8576761514561052, + "test_acc": 0.8578616352201258, + "test_auc": 0.9013752874570017, + "val_auc": 0.935374149659864, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487 + ] + }, + { + "test_f1": 0.8938049431979801, + "test_acc": 0.8938364779874214, + "test_auc": 0.9405010840078214, + "val_auc": 0.9687875150060024, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744 + ] + }, + { + "test_f1": 0.9139336344665348, + "test_acc": 0.9139622641509434, + "test_auc": 0.9475666277483724, + "val_auc": 0.9791916766706683, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 6475, + 26988, + 771 + ] + }, + { + "test_f1": 0.9363154427947931, + "test_acc": 0.9363522012578617, + "test_auc": 0.9797104215630978, + "val_auc": 0.9811924769907963, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 6475, + 26988, + 771, + 7368, + 1524, + 1028, + 559, + 878 + ] + }, + { + "test_f1": 0.9531959769159073, + "test_acc": 0.9532075471698114, + "test_auc": 0.9882055290980912, + "val_auc": 0.9843937575030012, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 487, + 1744, + 6475, + 26988, + 771, + 7368, + 1524, + 1028, + 559, + 878, + 24947, + 6065, + 152, + 1288, + 6935, + 3417 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..94f0c41a3ca93c427c0503cc705bb0a0a4baa38f --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/58_wikidatapolitical_party_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8234157348996112, + "test_acc": 0.8273417721518987, + "test_auc": 0.829014024422759, + "val_auc": 0.7991196478591437, + "k": 1, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104 + ] + }, + { + "test_f1": 0.824286962461306, + "test_acc": 0.8283544303797469, + "test_auc": 0.8968729105907156, + "val_auc": 0.8979591836734694, + "k": 2, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555 + ] + }, + { + "test_f1": 0.8448195674288332, + "test_acc": 0.8465822784810126, + "test_auc": 0.9224867610925841, + "val_auc": 0.9027611044417766, + "k": 5, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 337, + 292, + 1708 + ] + }, + { + "test_f1": 0.8554678829427647, + "test_acc": 0.8556962025316456, + "test_auc": 0.9347950481769072, + "val_auc": 0.9327731092436975, + "k": 10, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 337, + 292, + 1708, + 918, + 1815, + 67, + 677, + 4797 + ] + }, + { + "test_f1": 0.8572085273612798, + "test_acc": 0.8572151898734177, + "test_auc": 0.9384621537477079, + "val_auc": 0.9455782312925171, + "k": 16, + "dataset": "58_wikidatapolitical_party", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3104, + 4555, + 337, + 292, + 1708, + 918, + 1815, + 67, + 677, + 4797, + 1851, + 1380, + 1438, + 968, + 3502, + 1457 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..7a002562c99cc71e2b31050d37714772ef996eee --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/59_wikidata_occupation_isjournalist_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7479165438498813, + "test_acc": 0.7558974358974359, + "test_auc": 0.7667468946712895, + "val_auc": 0.7990616075071401, + "k": 1, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468 + ] + }, + { + "test_f1": 0.8363791473324184, + "test_acc": 0.837948717948718, + "test_auc": 0.8675775069848857, + "val_auc": 0.8380252957976336, + "k": 2, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351 + ] + }, + { + "test_f1": 0.8545806535704583, + "test_acc": 0.8553846153846154, + "test_auc": 0.9007809607163295, + "val_auc": 0.8884128926968584, + "k": 5, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 6475, + 1132 + ] + }, + { + "test_f1": 0.8576527010505786, + "test_acc": 0.8584615384615385, + "test_auc": 0.9112877436294475, + "val_auc": 0.9010607915136679, + "k": 10, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 6475, + 1132, + 343, + 1474, + 930, + 805, + 3417 + ] + }, + { + "test_f1": 0.8630969587105646, + "test_acc": 0.8635897435897436, + "test_auc": 0.9190195071868583, + "val_auc": 0.9018767849857201, + "k": 16, + "dataset": "59_wikidata_occupation_isjournalist", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 468, + 1351, + 1804, + 6475, + 1132, + 343, + 1474, + 930, + 805, + 3417, + 1305, + 7595, + 27324, + 21810, + 1551, + 52 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ab4eded5b67e4c211167a3c8872568e24864f5af --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/5_hist_fig_ismale_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8834895900842332, + "test_acc": 0.8840251572327044, + "test_auc": 0.8978984627442561, + "val_auc": 0.9403761504601841, + "k": 1, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731 + ] + }, + { + "test_f1": 0.8810919070610992, + "test_acc": 0.8812578616352201, + "test_auc": 0.9460268151435032, + "val_auc": 0.9783913565426171, + "k": 2, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279 + ] + }, + { + "test_f1": 0.9121624995018124, + "test_acc": 0.9122012578616352, + "test_auc": 0.9649593585671047, + "val_auc": 0.9923969587835134, + "k": 5, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279, + 1447, + 7740, + 238 + ] + }, + { + "test_f1": 0.92477573227406, + "test_acc": 0.9247798742138365, + "test_auc": 0.9783841448287106, + "val_auc": 0.9891956782713085, + "k": 10, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279, + 1447, + 7740, + 238, + 1556, + 1012, + 4877, + 6828, + 11381 + ] + }, + { + "test_f1": 0.9366024013604014, + "test_acc": 0.9366037735849057, + "test_auc": 0.9821077952364412, + "val_auc": 0.9915966386554622, + "k": 16, + "dataset": "5_hist_fig_ismale", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1731, + 1279, + 1447, + 7740, + 238, + 1556, + 1012, + 4877, + 6828, + 11381, + 27068, + 88, + 582, + 2860, + 1478, + 26750 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..252329c91f1de8da51d41223170975b0fb71af35 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/60_wikidata_occupation_isathlete_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8576591391702336, + "test_acc": 0.8594871794871795, + "test_auc": 0.8607356851920422, + "val_auc": 0.8931048551611587, + "k": 1, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628 + ] + }, + { + "test_f1": 0.9011451993508887, + "test_acc": 0.9015384615384615, + "test_auc": 0.9194003096913186, + "val_auc": 0.9267645858833129, + "k": 2, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416 + ] + }, + { + "test_f1": 0.8939231461725682, + "test_acc": 0.8943589743589744, + "test_auc": 0.9169555996903085, + "val_auc": 0.9265605875152998, + "k": 5, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 737 + ] + }, + { + "test_f1": 0.9001023803761079, + "test_acc": 0.9005128205128206, + "test_auc": 0.9322550240684014, + "val_auc": 0.9359445124439004, + "k": 10, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 737, + 238, + 4211, + 805, + 930, + 1132 + ] + }, + { + "test_f1": 0.9011190938662798, + "test_acc": 0.9015384615384615, + "test_auc": 0.9296377958057023, + "val_auc": 0.9330885352917178, + "k": 16, + "dataset": "60_wikidata_occupation_isathlete", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1628, + 416, + 5964, + 9381, + 737, + 238, + 4211, + 805, + 930, + 1132, + 1524, + 2062, + 334, + 1274, + 1288, + 878 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..f207ae5a99a990dc5009468690213f27affe4f49 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/62_wikidata_occupation_ispolitician_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8195834808390846, + "test_acc": 0.8215384615384616, + "test_auc": 0.8384766216716598, + "val_auc": 0.799469604243166, + "k": 1, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776 + ] + }, + { + "test_f1": 0.8132673773305001, + "test_acc": 0.8164102564102564, + "test_auc": 0.8691406907462889, + "val_auc": 0.8151774785801715, + "k": 2, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815 + ] + }, + { + "test_f1": 0.8588305504606473, + "test_acc": 0.8594871794871795, + "test_auc": 0.918243175009257, + "val_auc": 0.8992248062015503, + "k": 5, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 1866, + 334, + 1524 + ] + }, + { + "test_f1": 0.8829037412943967, + "test_acc": 0.8830769230769231, + "test_auc": 0.9457661493923991, + "val_auc": 0.8851489188086495, + "k": 10, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 1866, + 334, + 1524, + 6758, + 6935, + 487, + 930, + 1221 + ] + }, + { + "test_f1": 0.8881345245593405, + "test_acc": 0.8882051282051282, + "test_auc": 0.952069377587774, + "val_auc": 0.934312525499796, + "k": 16, + "dataset": "62_wikidata_occupation_ispolitician", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 776, + 1815, + 1866, + 334, + 1524, + 6758, + 6935, + 487, + 930, + 1221, + 913, + 1744, + 1158, + 26988, + 2046, + 508 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..45742b6cc081efdc6aff68259ef84bef3c020292 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/63_wikidata_occupation_issinger_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8371683801949218, + "test_acc": 0.837948717948718, + "test_auc": 0.883322954185882, + "val_auc": 0.8771929824561404, + "k": 1, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52 + ] + }, + { + "test_f1": 0.8465490172787636, + "test_acc": 0.8471794871794872, + "test_auc": 0.8831840980240347, + "val_auc": 0.8812729498164015, + "k": 2, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230 + ] + }, + { + "test_f1": 0.845498584166062, + "test_acc": 0.8461538461538461, + "test_auc": 0.8867354495573433, + "val_auc": 0.8806609547123623, + "k": 5, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 2131, + 2884, + 1274 + ] + }, + { + "test_f1": 0.8506685067477902, + "test_acc": 0.8512820512820513, + "test_auc": 0.9011806981519508, + "val_auc": 0.9279885760913913, + "k": 10, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 2131, + 2884, + 1274, + 888, + 2046, + 238, + 821, + 1193 + ] + }, + { + "test_f1": 0.8486861122590452, + "test_acc": 0.8492307692307692, + "test_auc": 0.9037011478776047, + "val_auc": 0.9461444308445532, + "k": 16, + "dataset": "63_wikidata_occupation_issinger", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 52, + 19230, + 2131, + 2884, + 1274, + 888, + 2046, + 238, + 821, + 1193, + 2454, + 1815, + 18545, + 1682, + 2484, + 17546 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..498d76bd714274e332636ce33fe194d07048261e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/64_wikidata_occupation_isresearcher_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8790156864071803, + "test_acc": 0.879144385026738, + "test_auc": 0.923916524826589, + "val_auc": 0.9313131313131313, + "k": 1, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839 + ] + }, + { + "test_f1": 0.8779181896870455, + "test_acc": 0.8780748663101604, + "test_auc": 0.9378511685792199, + "val_auc": 0.9381818181818182, + "k": 2, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682 + ] + }, + { + "test_f1": 0.8844845791280591, + "test_acc": 0.8844919786096257, + "test_auc": 0.9465079887992094, + "val_auc": 0.9563636363636364, + "k": 5, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 343 + ] + }, + { + "test_f1": 0.9015894191618603, + "test_acc": 0.9016042780748663, + "test_auc": 0.9534535771152474, + "val_auc": 0.96, + "k": 10, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 343, + 930, + 3417, + 888, + 340, + 1895 + ] + }, + { + "test_f1": 0.9101357579020791, + "test_acc": 0.9101604278074866, + "test_auc": 0.9581526016215524, + "val_auc": 0.9494949494949495, + "k": 16, + "dataset": "64_wikidata_occupation_isresearcher", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1839, + 1682, + 1274, + 1305, + 343, + 930, + 3417, + 888, + 340, + 1895, + 874, + 238, + 1441, + 456, + 1744, + 1776 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5b6a567a5a6b14c9e878641e99d876f21354c78e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/65_high-school_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9443656756819676, + "test_acc": 0.944367816091954, + "test_auc": 0.9859012257156772, + "val_auc": 0.9955357142857142, + "k": 1, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554 + ] + }, + { + "test_f1": 0.9595373046273171, + "test_acc": 0.9595402298850575, + "test_auc": 0.9923815547378104, + "val_auc": 0.9995941558441559, + "k": 2, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072 + ] + }, + { + "test_f1": 0.9733332995110996, + "test_acc": 0.9733333333333334, + "test_auc": 0.9954779749986471, + "val_auc": 1.0, + "k": 5, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 4323, + 1101, + 760 + ] + }, + { + "test_f1": 0.9760919439152115, + "test_acc": 0.9760919540229885, + "test_auc": 0.9954686739001027, + "val_auc": 1.0, + "k": 10, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 4323, + 1101, + 760, + 192, + 1065, + 1682, + 1622, + 1876 + ] + }, + { + "test_f1": 0.9811494252873563, + "test_acc": 0.9811494252873563, + "test_auc": 0.9949926267655176, + "val_auc": 1.0, + "k": 16, + "dataset": "65_high-school", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 7554, + 1072, + 4323, + 1101, + 760, + 192, + 1065, + 1682, + 1622, + 1876, + 763, + 1872, + 1684, + 770, + 652, + 416 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..61f3934b6bfc6853ed1778b4a48bd9e248c446a2 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/66_living-room_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9789766707548201, + "test_acc": 0.9789775441949355, + "test_auc": 0.9867407744242406, + "val_auc": 0.9978549978549979, + "k": 1, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286 + ] + }, + { + "test_f1": 0.9775439796359937, + "test_acc": 0.9775441949354993, + "test_auc": 0.9867407744242406, + "val_auc": 0.9978549978549979, + "k": 2, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905 + ] + }, + { + "test_f1": 0.9861432086882203, + "test_acc": 0.9861442904921166, + "test_auc": 0.9934511971744819, + "val_auc": 1.0, + "k": 5, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581 + ] + }, + { + "test_f1": 0.9866215177530673, + "test_acc": 0.9866220735785953, + "test_auc": 0.9956873960199495, + "val_auc": 1.0, + "k": 10, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581, + 304, + 1079, + 6080, + 2874, + 1682 + ] + }, + { + "test_f1": 0.9890101310285884, + "test_acc": 0.989010989010989, + "test_auc": 0.9975236540347456, + "val_auc": 1.0, + "k": 16, + "dataset": "66_living-room", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18286, + 1905, + 1233, + 1874, + 581, + 304, + 1079, + 6080, + 2874, + 1682, + 948, + 1684, + 760, + 239, + 7003, + 1596 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..3ccc947210962e84472223665fa976c92f9c194c --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/67_social-security_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9885054263892054, + "test_acc": 0.9885057471264368, + "test_auc": 0.9972887297743385, + "val_auc": 1.0, + "k": 1, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559 + ] + }, + { + "test_f1": 0.9898850531948972, + "test_acc": 0.9898850574712643, + "test_auc": 0.9987367417068024, + "val_auc": 1.0, + "k": 2, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300 + ] + }, + { + "test_f1": 0.991264367816092, + "test_acc": 0.991264367816092, + "test_auc": 0.9996034349802478, + "val_auc": 1.0, + "k": 5, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 334, + 982 + ] + }, + { + "test_f1": 0.9903448030942082, + "test_acc": 0.9903448275862069, + "test_auc": 0.9996152727420314, + "val_auc": 1.0, + "k": 10, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 334, + 982, + 2236, + 724, + 3825, + 1158, + 1268 + ] + }, + { + "test_f1": 0.9935632183908046, + "test_acc": 0.9935632183908046, + "test_auc": 0.9990022457925213, + "val_auc": 1.0, + "k": 16, + "dataset": "67_social-security", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14559, + 300, + 1789, + 334, + 982, + 2236, + 724, + 3825, + 1158, + 1268, + 896, + 1872, + 913, + 1769, + 1592, + 1305 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..886c852cbacbce26331e25dfd435ef3bdb300b04 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/68_credit-card_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9420688185654673, + "test_acc": 0.9420689655172414, + "test_auc": 0.9765789883110557, + "val_auc": 0.9573863636363636, + "k": 1, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095 + ] + }, + { + "test_f1": 0.9622964117306805, + "test_acc": 0.9622988505747126, + "test_auc": 0.9863696628605444, + "val_auc": 0.9756493506493507, + "k": 2, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023 + ] + }, + { + "test_f1": 0.9632173644504474, + "test_acc": 0.9632183908045977, + "test_auc": 0.9884911588830565, + "val_auc": 0.9756493506493505, + "k": 5, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 494 + ] + }, + { + "test_f1": 0.9756304221188694, + "test_acc": 0.975632183908046, + "test_auc": 0.9952716597218464, + "val_auc": 0.987012987012987, + "k": 10, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 494, + 1301, + 1682, + 1701, + 1079, + 211 + ] + }, + { + "test_f1": 0.9747115210855107, + "test_acc": 0.9747126436781609, + "test_auc": 0.9955227893825424, + "val_auc": 0.9857954545454545, + "k": 16, + "dataset": "68_credit-card", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 16095, + 1023, + 495, + 1789, + 494, + 1301, + 1682, + 1701, + 1079, + 211, + 586, + 6679, + 1193, + 235, + 1627, + 2011 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..75d81292de295bbb28c8c7ac260341d75f4dc26d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/6_hist_fig_isamerican_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.828971750216821, + "test_acc": 0.8291823899371069, + "test_auc": 0.8715296307285079, + "val_auc": 0.8983593437374949, + "k": 1, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508 + ] + }, + { + "test_f1": 0.8932651063285781, + "test_acc": 0.8933333333333333, + "test_auc": 0.9482536892213875, + "val_auc": 0.9899959983993598, + "k": 2, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028 + ] + }, + { + "test_f1": 0.9096747188223325, + "test_acc": 0.9096855345911949, + "test_auc": 0.963455620486887, + "val_auc": 0.9951980792316927, + "k": 5, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 3417, + 778, + 1839 + ] + }, + { + "test_f1": 0.9053773291211025, + "test_acc": 0.9054088050314465, + "test_auc": 0.961030399812058, + "val_auc": 0.9951980792316927, + "k": 10, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 3417, + 778, + 1839, + 1305, + 1602, + 4663, + 1895, + 654 + ] + }, + { + "test_f1": 0.9053773291211025, + "test_acc": 0.9054088050314465, + "test_auc": 0.9610458422401545, + "val_auc": 0.9951980792316927, + "k": 16, + "dataset": "6_hist_fig_isamerican", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 508, + 1028, + 3417, + 778, + 1839, + 1305, + 1602, + 4663, + 1895, + 654, + 1478, + 1279, + 988, + 343, + 1815, + 30649 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ed2e21357411ad1353fd30916f9aefb89142e95e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9990804593813543, + "test_acc": 0.999080459770115, + "test_auc": 0.9999890077926296, + "val_auc": 1.0, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686 + ] + }, + { + "test_f1": 0.9981609187627085, + "test_acc": 0.9981609195402299, + "test_auc": 0.9999560311705178, + "val_auc": 1.0, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379 + ] + }, + { + "test_f1": 0.9986206896551724, + "test_acc": 0.9986206896551724, + "test_auc": 0.9991333067265544, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379, + 682, + 4787, + 547 + ] + }, + { + "test_f1": 0.9977011465095805, + "test_acc": 0.9977011494252873, + "test_auc": 0.9990986389956167, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379, + 682, + 4787, + 547, + 1362, + 4276, + 24163, + 1840, + 1245 + ] + }, + { + "test_f1": 0.9981609187627085, + "test_acc": 0.9981609195402299, + "test_auc": 0.9994275596623194, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1686, + 1379, + 682, + 4787, + 547, + 1362, + 4276, + 24163, + 1840, + 1245, + 1931, + 1584, + 1234, + 613, + 119, + 6777 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b3dea0c8d6e2b0ba51ec1e734086a3da2d537883 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/73_control-group_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9001158234350929, + "test_acc": 0.9002298850574713, + "test_auc": 0.9433994331403215, + "val_auc": 0.9435876623376623, + "k": 1, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961 + ] + }, + { + "test_f1": 0.9415969483484435, + "test_acc": 0.9416091954022988, + "test_auc": 0.9778397099410142, + "val_auc": 0.9910714285714286, + "k": 2, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 8734 + ] + }, + { + "test_f1": 0.9452809713183624, + "test_acc": 0.9452873563218391, + "test_auc": 0.9829121908653067, + "val_auc": 0.9914772727272727, + "k": 5, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 8734, + 7182, + 248, + 732 + ] + }, + { + "test_f1": 0.9576992716224334, + "test_acc": 0.9577011494252874, + "test_auc": 0.9848180705124736, + "val_auc": 0.9963474025974026, + "k": 10, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 8734, + 7182, + 248, + 732, + 1264, + 1536, + 1441, + 859, + 656 + ] + }, + { + "test_f1": 0.9668964117695942, + "test_acc": 0.9668965517241379, + "test_auc": 0.9921963783213377, + "val_auc": 0.9991883116883117, + "k": 16, + "dataset": "73_control-group", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 1961, + 8734, + 7182, + 248, + 732, + 1264, + 1536, + 1441, + 859, + 656, + 1482, + 1682, + 653, + 1960, + 306, + 4861 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..79e7a0b897f03070ec7e1d36df775cf4a6ddf81e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/78_north-america_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.954936336905543, + "test_acc": 0.9549425287356322, + "test_auc": 0.9773416783916878, + "val_auc": 1.0, + "k": 1, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6161 + ] + }, + { + "test_f1": 0.9636781609195403, + "test_acc": 0.9636781609195403, + "test_auc": 0.9856048588938796, + "val_auc": 1.0, + "k": 2, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6161, + 9439 + ] + }, + { + "test_f1": 0.9710342255915481, + "test_acc": 0.9710344827586207, + "test_auc": 0.9858581024406081, + "val_auc": 1.0, + "k": 5, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6161, + 9439, + 26232, + 1121, + 1795 + ] + }, + { + "test_f1": 0.9714942408219831, + "test_acc": 0.9714942528735632, + "test_auc": 0.9879660695925105, + "val_auc": 1.0, + "k": 10, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6161, + 9439, + 26232, + 1121, + 1795, + 5535, + 802, + 5907, + 5672, + 493 + ] + }, + { + "test_f1": 0.9811494013785613, + "test_acc": 0.9811494252873563, + "test_auc": 0.9980518426321772, + "val_auc": 1.0, + "k": 16, + "dataset": "78_north-america", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6161, + 9439, + 26232, + 1121, + 1795, + 5535, + 802, + 5907, + 5672, + 493, + 1441, + 345, + 1631, + 5797, + 2020, + 822 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..792a6dae640985580af75682db5d92601c0b7555 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9558544914524637, + "test_acc": 0.9558620689655173, + "test_auc": 0.9824166959792195, + "val_auc": 0.9939123376623377, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272 + ] + }, + { + "test_f1": 0.9540100390313877, + "test_acc": 0.9540229885057471, + "test_auc": 0.9824166959792195, + "val_auc": 0.9939123376623377, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866 + ] + }, + { + "test_f1": 0.9641377794170602, + "test_acc": 0.9641379310344828, + "test_auc": 0.9940845013258293, + "val_auc": 0.9979707792207793, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 5703, + 165, + 1072 + ] + }, + { + "test_f1": 0.9820689200321012, + "test_acc": 0.9820689655172414, + "test_auc": 0.9974810934033227, + "val_auc": 0.9991883116883117, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 5703, + 165, + 1072, + 3198, + 1682, + 918, + 1726, + 508 + ] + }, + { + "test_f1": 0.9834482338758478, + "test_acc": 0.983448275862069, + "test_auc": 0.9978818861951404, + "val_auc": 0.9995941558441559, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18272, + 1866, + 5703, + 165, + 1072, + 3198, + 1682, + 918, + 1726, + 508, + 495, + 1715, + 1079, + 6099, + 1645, + 1833 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..cded273cae725e41b93c646284aa46dd5bddca53 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9466379020072656, + "test_acc": 0.9466666666666667, + "test_auc": 0.9666771233562422, + "val_auc": 0.9752435064935066, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772 + ] + }, + { + "test_f1": 0.9632091359065256, + "test_acc": 0.9632183908045977, + "test_auc": 0.9795409654202067, + "val_auc": 0.9876217532467533, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963 + ] + }, + { + "test_f1": 0.9737923721673882, + "test_acc": 0.9737931034482759, + "test_auc": 0.9918344810325235, + "val_auc": 0.9991883116883117, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963, + 217, + 352, + 1942 + ] + }, + { + "test_f1": 0.9788505747126437, + "test_acc": 0.9788505747126437, + "test_auc": 0.9932414835759511, + "val_auc": 0.9983766233766234, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963, + 217, + 352, + 1942, + 1682, + 111, + 393, + 1898, + 581 + ] + }, + { + "test_f1": 0.9839076717870553, + "test_acc": 0.9839080459770115, + "test_auc": 0.9944573908220141, + "val_auc": 0.9983766233766234, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 3772, + 10963, + 217, + 352, + 1942, + 1682, + 111, + 393, + 1898, + 581, + 1261, + 760, + 8111, + 1441, + 1726, + 1221 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..b7da66dc4f2c78cef317815fca907d9f10af91d5 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/82_federal-government_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.858311374462362, + "test_acc": 0.8583908045977011, + "test_auc": 0.9033501711402132, + "val_auc": 0.9312094155844156, + "k": 1, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020 + ] + }, + { + "test_f1": 0.9149387527421725, + "test_acc": 0.9149425287356322, + "test_auc": 0.9613116578278046, + "val_auc": 0.9853896103896104, + "k": 2, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760 + ] + }, + { + "test_f1": 0.9429635482981767, + "test_acc": 0.9429885057471264, + "test_auc": 0.983806787434385, + "val_auc": 0.9849837662337662, + "k": 5, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1682, + 913, + 1482 + ] + }, + { + "test_f1": 0.9503267461272392, + "test_acc": 0.9503448275862069, + "test_auc": 0.9881385626927863, + "val_auc": 0.9914772727272727, + "k": 10, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1682, + 913, + 1482, + 49, + 18571, + 170, + 2023, + 1960 + ] + }, + { + "test_f1": 0.9572306206483003, + "test_acc": 0.9572413793103448, + "test_auc": 0.9926758076735753, + "val_auc": 0.9963474025974026, + "k": 16, + "dataset": "82_federal-government", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2020, + 760, + 1682, + 913, + 1482, + 49, + 18571, + 170, + 2023, + 1960, + 1264, + 364, + 2817, + 2084, + 901, + 1928 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0bafafa190310205a6219beb8f3422cc8734a16b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/89_glue_mrpc_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6920545449945608, + "test_acc": 0.6933235509904622, + "test_auc": 0.7459672036551388, + "val_auc": 0.7760416666666666, + "k": 1, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394 + ] + }, + { + "test_f1": 0.6844941280762071, + "test_acc": 0.6867204695524578, + "test_auc": 0.7459672036551388, + "val_auc": 0.7760416666666666, + "k": 2, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041 + ] + }, + { + "test_f1": 0.6920545449945608, + "test_acc": 0.6933235509904622, + "test_auc": 0.7429969727113397, + "val_auc": 0.7772435897435899, + "k": 5, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 584, + 28053 + ] + }, + { + "test_f1": 0.6896277173728873, + "test_acc": 0.6903888481291269, + "test_auc": 0.7501808621959254, + "val_auc": 0.7980769230769231, + "k": 10, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 584, + 28053, + 1928, + 1737, + 1551, + 1257, + 7649 + ] + }, + { + "test_f1": 0.7027490138591429, + "test_acc": 0.7028613352898019, + "test_auc": 0.7608463489520758, + "val_auc": 0.813701923076923, + "k": 16, + "dataset": "89_glue_mrpc", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5394, + 2041, + 5133, + 584, + 28053, + 1928, + 1737, + 1551, + 1257, + 7649, + 31, + 1697, + 28397, + 465, + 2367, + 943 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..58d2023d3cdd6d8665898476121c9b9b45d77f41 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6573583604540816, + "test_acc": 0.6573584905660378, + "test_auc": 0.6860027806496756, + "val_auc": 0.6952781112444978, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520 + ] + }, + { + "test_f1": 0.694081428786699, + "test_acc": 0.6940880503144654, + "test_auc": 0.770954362308729, + "val_auc": 0.7482993197278912, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1011 + ] + }, + { + "test_f1": 0.734285943604907, + "test_acc": 0.7343396226415094, + "test_auc": 0.8194590289598689, + "val_auc": 0.7663065226090436, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1011, + 1927, + 30127, + 4674 + ] + }, + { + "test_f1": 0.7657437320529938, + "test_acc": 0.7657861635220126, + "test_auc": 0.8415819527127536, + "val_auc": 0.8291316526610645, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1011, + 1927, + 30127, + 4674, + 24630, + 1848, + 3801, + 7614, + 1036 + ] + }, + { + "test_f1": 0.7666895044073928, + "test_acc": 0.7667924528301887, + "test_auc": 0.84527016148223, + "val_auc": 0.8159263705482193, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4520, + 1011, + 1927, + 30127, + 4674, + 24630, + 1848, + 3801, + 7614, + 1036, + 1520, + 1928, + 6417, + 14940, + 5653, + 635 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..982197f090f258a6cb3320a3683794294c125e94 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/94_ai_gen_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.876704205273842, + "test_acc": 0.8782389937106918, + "test_auc": 0.8817335315364762, + "val_auc": 0.8695478191276511, + "k": 1, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480 + ] + }, + { + "test_f1": 0.8766760387937521, + "test_acc": 0.8782389937106918, + "test_auc": 0.9445621894426448, + "val_auc": 0.9443777511004401, + "k": 2, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433 + ] + }, + { + "test_f1": 0.92074669355166, + "test_acc": 0.9210062893081761, + "test_auc": 0.9680965764390064, + "val_auc": 0.98359343737495, + "k": 5, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 7666 + ] + }, + { + "test_f1": 0.917986476709057, + "test_acc": 0.9179874213836478, + "test_auc": 0.9797718115436453, + "val_auc": 0.9971988795518207, + "k": 10, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 7666, + 6553, + 1335, + 3481, + 1927, + 1928 + ] + }, + { + "test_f1": 0.9192369003664036, + "test_acc": 0.9192452830188679, + "test_auc": 0.9805314017978024, + "val_auc": 0.9975990396158463, + "k": 16, + "dataset": "94_ai_gen", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 18480, + 433, + 1116, + 444, + 7666, + 6553, + 1335, + 3481, + 1927, + 1928, + 1064, + 4110, + 31245, + 988, + 1500, + 1441 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..d2a6c0a8f80f25dc325b9080590aa7980e227e8d --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/96_spam_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.8848237049633199, + "test_acc": 0.8848614072494669, + "test_auc": 0.9363884342607747, + "val_auc": 0.9605999999999999, + "k": 1, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655 + ] + }, + { + "test_f1": 0.9188762218158011, + "test_acc": 0.9189765458422174, + "test_auc": 0.976577559556283, + "val_auc": 0.992, + "k": 2, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 1417 + ] + }, + { + "test_f1": 0.953090404838086, + "test_acc": 0.9530916844349681, + "test_auc": 0.9867430441898527, + "val_auc": 0.9952000000000001, + "k": 5, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 1417, + 28188, + 12485, + 4080 + ] + }, + { + "test_f1": 0.959477956014937, + "test_acc": 0.9594882729211087, + "test_auc": 0.9885615566466631, + "val_auc": 0.9928, + "k": 10, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 1417, + 28188, + 12485, + 4080, + 727, + 1927, + 5406, + 1472, + 1625 + ] + }, + { + "test_f1": 0.9594882729211087, + "test_acc": 0.9594882729211087, + "test_auc": 0.9901800327332242, + "val_auc": 0.9936, + "k": 16, + "dataset": "96_spam_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 32655, + 1417, + 28188, + 12485, + 4080, + 727, + 1927, + 5406, + 1472, + 1625, + 6096, + 330, + 1118, + 19585, + 984, + 610 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/scr/saebench_scr_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/scr/saebench_scr_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..9aa96762676cce7943afd9de2a993e5625676e9a --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/scr/saebench_scr_custom_sae_eval_results.json @@ -0,0 +1,326 @@ +{ + "eval_type_id": "scr", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": true, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "b5079288-c93b-49c7-a66b-059c9a0a5aa3", + "datetime_epoch_millis": 1776879736952, + "eval_result_metrics": { + "scr_metrics": { + "scr_dir1_threshold_2": 0.25580976118569804, + "scr_metric_threshold_2": 0.1843316007574383, + "scr_dir2_threshold_2": 0.17422691752734093, + "scr_dir1_threshold_5": 0.0808702078158759, + "scr_metric_threshold_5": 0.2439554180118887, + "scr_dir2_threshold_5": 0.2266687756213812, + "scr_dir1_threshold_10": 0.1445751847860934, + "scr_metric_threshold_10": 0.295437944766743, + "scr_dir2_threshold_10": 0.27283981873157087, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results", + "scr_dir1_threshold_2": 0.5128207283533913, + "scr_metric_threshold_2": 0.06476677137111979, + "scr_dir2_threshold_2": 0.06476677137111979, + "scr_dir1_threshold_5": 0.5512825313325533, + "scr_metric_threshold_5": 0.21761653710635845, + "scr_dir2_threshold_5": 0.21761653710635845, + "scr_dir1_threshold_10": 0.5384614208981502, + "scr_metric_threshold_10": 0.22538861143735814, + "scr_dir2_threshold_10": 0.22538861143735814, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results", + "scr_dir1_threshold_2": 0.23484857036936357, + "scr_metric_threshold_2": 0.31804288982255313, + "scr_dir2_threshold_2": 0.31804288982255313, + "scr_dir1_threshold_5": -0.16666629037480024, + "scr_metric_threshold_5": 0.31804288982255313, + "scr_dir2_threshold_5": 0.31804288982255313, + "scr_dir1_threshold_10": -0.13636365688864727, + "scr_metric_threshold_10": 0.4128439798401429, + "scr_dir2_threshold_10": 0.4128439798401429, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results", + "scr_dir1_threshold_2": 0.3749997671693945, + "scr_metric_threshold_2": 0.05025135035200431, + "scr_dir2_threshold_2": 0.05025135035200431, + "scr_dir1_threshold_5": -0.2968748544808716, + "scr_metric_threshold_5": 0.11809041124652613, + "scr_dir2_threshold_5": 0.11809041124652613, + "scr_dir1_threshold_10": -0.06250011641530274, + "scr_metric_threshold_10": 0.17336682175353543, + "scr_dir2_threshold_10": 0.17336682175353543, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results", + "scr_dir1_threshold_2": 0.17449641072003616, + "scr_metric_threshold_2": 0.06369439446808171, + "scr_dir2_threshold_2": 0.06369439446808171, + "scr_dir1_threshold_5": -0.5503357261866425, + "scr_metric_threshold_5": -0.09872616905774931, + "scr_dir2_threshold_5": -0.09872616905774931, + "scr_dir1_threshold_10": -0.5838927436340035, + "scr_metric_threshold_10": -0.14649682254103324, + "scr_dir2_threshold_10": -0.14649682254103324, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results", + "scr_dir1_threshold_2": 0.047368439214711934, + "scr_metric_threshold_2": 0.49603167939138937, + "scr_dir2_threshold_2": 0.49603167939138937, + "scr_dir1_threshold_5": 0.03157916861558172, + "scr_metric_threshold_5": 0.5595238630238142, + "scr_dir2_threshold_5": 0.5595238630238142, + "scr_dir1_threshold_10": 0.08421069803000372, + "scr_metric_threshold_10": 0.5992063595309115, + "scr_dir2_threshold_10": 0.5992063595309115, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results", + "scr_dir1_threshold_2": 0.2786885673060324, + "scr_metric_threshold_2": 0.11450392446499717, + "scr_dir2_threshold_2": 0.11450392446499717, + "scr_dir1_threshold_5": 0.3989072479911928, + "scr_metric_threshold_5": 0.19465646684173443, + "scr_dir2_threshold_5": 0.19465646684173443, + "scr_dir1_threshold_10": 0.502732368584764, + "scr_metric_threshold_10": 0.2671755204201627, + "scr_dir2_threshold_10": 0.2671755204201627, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results", + "scr_dir1_threshold_2": 0.2790697352057136, + "scr_metric_threshold_2": 0.22317592504242034, + "scr_dir2_threshold_2": 0.22317592504242034, + "scr_dir1_threshold_5": 0.41860446419313896, + "scr_metric_threshold_5": 0.38197422367501804, + "scr_dir2_threshold_5": 0.38197422367501804, + "scr_dir1_threshold_10": 0.48837210591771457, + "scr_metric_threshold_10": 0.506437685896798, + "scr_dir2_threshold_10": 0.506437685896798, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results", + "scr_dir1_threshold_2": 0.14418587114694043, + "scr_metric_threshold_2": 0.14418587114694043, + "scr_dir2_threshold_2": 0.06334840530616152, + "scr_dir1_threshold_5": 0.2604651214368547, + "scr_metric_threshold_5": 0.2604651214368547, + "scr_dir2_threshold_5": 0.1221719823127949, + "scr_dir1_threshold_10": 0.32558140179606837, + "scr_metric_threshold_10": 0.32558140179606837, + "scr_dir2_threshold_10": 0.14479639351469145, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_scr", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": null +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/tpp/saebench_tpp_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/tpp/saebench_tpp_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..994aaf8ff4aefb26f4ec231feba546ef1a845344 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-183105-tokens-750000000/tpp/saebench_tpp_custom_sae_eval_results.json @@ -0,0 +1,297 @@ +{ + "eval_type_id": "tpp", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": false, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "4c5d8f8e-fb8f-44e7-898d-7b10dcd8e627", + "datetime_epoch_millis": 1776879909140, + "eval_result_metrics": { + "tpp_metrics": { + "tpp_threshold_2_total_metric": 0.06044999659061432, + "tpp_threshold_2_intended_diff_only": 0.07430000305175781, + "tpp_threshold_2_unintended_diff_only": 0.013850006461143495, + "tpp_threshold_5_total_metric": 0.11140000075101852, + "tpp_threshold_5_intended_diff_only": 0.1346000075340271, + "tpp_threshold_5_unintended_diff_only": 0.023200006783008577, + "tpp_threshold_10_total_metric": 0.14835001081228255, + "tpp_threshold_10_intended_diff_only": 0.19780001640319822, + "tpp_threshold_10_unintended_diff_only": 0.049450005590915676, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results", + "tpp_threshold_2_total_metric": 0.1104999989271164, + "tpp_threshold_2_intended_diff_only": 0.13660001754760742, + "tpp_threshold_2_unintended_diff_only": 0.02610001862049103, + "tpp_threshold_5_total_metric": 0.1994500070810318, + "tpp_threshold_5_intended_diff_only": 0.24140002727508544, + "tpp_threshold_5_unintended_diff_only": 0.04195002019405365, + "tpp_threshold_10_total_metric": 0.24950001537799835, + "tpp_threshold_10_intended_diff_only": 0.3412000298500061, + "tpp_threshold_10_unintended_diff_only": 0.09170001447200775, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results", + "tpp_threshold_2_total_metric": 0.010399994254112244, + "tpp_threshold_2_intended_diff_only": 0.011999988555908203, + "tpp_threshold_2_unintended_diff_only": 0.0015999943017959595, + "tpp_threshold_5_total_metric": 0.02334999442100525, + "tpp_threshold_5_intended_diff_only": 0.02779998779296875, + "tpp_threshold_5_unintended_diff_only": 0.004449993371963501, + "tpp_threshold_10_total_metric": 0.04720000624656677, + "tpp_threshold_10_intended_diff_only": 0.05440000295639038, + "tpp_threshold_10_unintended_diff_only": 0.007199996709823608, + "tpp_threshold_20_total_metric": null, + "tpp_threshold_20_intended_diff_only": null, + "tpp_threshold_20_unintended_diff_only": null, + "tpp_threshold_50_total_metric": null, + "tpp_threshold_50_intended_diff_only": null, + "tpp_threshold_50_unintended_diff_only": null, + "tpp_threshold_100_total_metric": null, + "tpp_threshold_100_intended_diff_only": null, + "tpp_threshold_100_unintended_diff_only": null, + "tpp_threshold_500_total_metric": null, + "tpp_threshold_500_intended_diff_only": null, + "tpp_threshold_500_unintended_diff_only": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_tpp", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": { + "LabHC/bias_in_bios_class_set1": { + "0": { + "tpp_threshold_2_total_metric": 0.1759999692440033, + "tpp_threshold_2_intended_diff_only": 0.2409999966621399, + "tpp_threshold_2_unintended_diff_only": 0.0650000274181366, + "tpp_threshold_5_total_metric": 0.2122499793767929, + "tpp_threshold_5_intended_diff_only": 0.29100000858306885, + "tpp_threshold_5_unintended_diff_only": 0.07875002920627594, + "tpp_threshold_10_total_metric": 0.2212499976158142, + "tpp_threshold_10_intended_diff_only": 0.35100001096725464, + "tpp_threshold_10_unintended_diff_only": 0.12975001335144043 + }, + "1": { + "tpp_threshold_2_total_metric": 0.1107499897480011, + "tpp_threshold_2_intended_diff_only": 0.15200001001358032, + "tpp_threshold_2_unintended_diff_only": 0.041250020265579224, + "tpp_threshold_5_total_metric": 0.12575000524520874, + "tpp_threshold_5_intended_diff_only": 0.17400002479553223, + "tpp_threshold_5_unintended_diff_only": 0.048250019550323486, + "tpp_threshold_10_total_metric": 0.24324998259544373, + "tpp_threshold_10_intended_diff_only": 0.2919999957084656, + "tpp_threshold_10_unintended_diff_only": 0.04875001311302185 + }, + "2": { + "tpp_threshold_2_total_metric": 0.1482500433921814, + "tpp_threshold_2_intended_diff_only": 0.16200006008148193, + "tpp_threshold_2_unintended_diff_only": 0.013750016689300537, + "tpp_threshold_5_total_metric": 0.20325003564357758, + "tpp_threshold_5_intended_diff_only": 0.22300004959106445, + "tpp_threshold_5_unintended_diff_only": 0.019750013947486877, + "tpp_threshold_10_total_metric": 0.20775005221366882, + "tpp_threshold_10_intended_diff_only": 0.31000006198883057, + "tpp_threshold_10_unintended_diff_only": 0.10225000977516174 + }, + "6": { + "tpp_threshold_2_total_metric": 0.10450001060962677, + "tpp_threshold_2_intended_diff_only": 0.11400002241134644, + "tpp_threshold_2_unintended_diff_only": 0.009500011801719666, + "tpp_threshold_5_total_metric": 0.29899999499320984, + "tpp_threshold_5_intended_diff_only": 0.32200002670288086, + "tpp_threshold_5_unintended_diff_only": 0.02300003170967102, + "tpp_threshold_10_total_metric": 0.3765000253915787, + "tpp_threshold_10_intended_diff_only": 0.4270000457763672, + "tpp_threshold_10_unintended_diff_only": 0.05050002038478851 + }, + "9": { + "tpp_threshold_2_total_metric": 0.01299998164176941, + "tpp_threshold_2_intended_diff_only": 0.013999998569488525, + "tpp_threshold_2_unintended_diff_only": 0.0010000169277191162, + "tpp_threshold_5_total_metric": 0.15700002014636993, + "tpp_threshold_5_intended_diff_only": 0.19700002670288086, + "tpp_threshold_5_unintended_diff_only": 0.040000006556510925, + "tpp_threshold_10_total_metric": 0.19875001907348633, + "tpp_threshold_10_intended_diff_only": 0.32600003480911255, + "tpp_threshold_10_unintended_diff_only": 0.12725001573562622 + } + }, + "canrager/amazon_reviews_mcauley_1and5": { + "1": { + "tpp_threshold_2_total_metric": 0.010749965906143188, + "tpp_threshold_2_intended_diff_only": 0.006999969482421875, + "tpp_threshold_2_unintended_diff_only": -0.0037499964237213135, + "tpp_threshold_5_total_metric": 0.010249987244606018, + "tpp_threshold_5_intended_diff_only": 0.010999977588653564, + "tpp_threshold_5_unintended_diff_only": 0.0007499903440475464, + "tpp_threshold_10_total_metric": 0.01299998164176941, + "tpp_threshold_10_intended_diff_only": 0.019999980926513672, + "tpp_threshold_10_unintended_diff_only": 0.006999999284744263 + }, + "2": { + "tpp_threshold_2_total_metric": 0.006750017404556274, + "tpp_threshold_2_intended_diff_only": 0.013000011444091797, + "tpp_threshold_2_unintended_diff_only": 0.0062499940395355225, + "tpp_threshold_5_total_metric": 0.029750019311904907, + "tpp_threshold_5_intended_diff_only": 0.04100000858306885, + "tpp_threshold_5_unintended_diff_only": 0.01124998927116394, + "tpp_threshold_10_total_metric": 0.07425004243850708, + "tpp_threshold_10_intended_diff_only": 0.08600002527236938, + "tpp_threshold_10_unintended_diff_only": 0.011749982833862305 + }, + "3": { + "tpp_threshold_2_total_metric": 0.017499983310699463, + "tpp_threshold_2_intended_diff_only": 0.015999972820281982, + "tpp_threshold_2_unintended_diff_only": -0.0015000104904174805, + "tpp_threshold_5_total_metric": 0.025250017642974854, + "tpp_threshold_5_intended_diff_only": 0.027000010013580322, + "tpp_threshold_5_unintended_diff_only": 0.0017499923706054688, + "tpp_threshold_10_total_metric": 0.01975002884864807, + "tpp_threshold_10_intended_diff_only": 0.026000022888183594, + "tpp_threshold_10_unintended_diff_only": 0.0062499940395355225 + }, + "5": { + "tpp_threshold_2_total_metric": -0.006500020623207092, + "tpp_threshold_2_intended_diff_only": 0.001999974250793457, + "tpp_threshold_2_unintended_diff_only": 0.00849999487400055, + "tpp_threshold_5_total_metric": 0.013999953866004944, + "tpp_threshold_5_intended_diff_only": 0.01699995994567871, + "tpp_threshold_5_unintended_diff_only": 0.003000006079673767, + "tpp_threshold_10_total_metric": 0.029749974608421326, + "tpp_threshold_10_intended_diff_only": 0.0339999794960022, + "tpp_threshold_10_unintended_diff_only": 0.004250004887580872 + }, + "6": { + "tpp_threshold_2_total_metric": 0.023500025272369385, + "tpp_threshold_2_intended_diff_only": 0.022000014781951904, + "tpp_threshold_2_unintended_diff_only": -0.0015000104904174805, + "tpp_threshold_5_total_metric": 0.03749999403953552, + "tpp_threshold_5_intended_diff_only": 0.042999982833862305, + "tpp_threshold_5_unintended_diff_only": 0.005499988794326782, + "tpp_threshold_10_total_metric": 0.09925000369548798, + "tpp_threshold_10_intended_diff_only": 0.10600000619888306, + "tpp_threshold_10_unintended_diff_only": 0.006750002503395081 + } + } + } +} \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..0db3b94d1cbbf03bc73a204b2ccf86eb70a5139e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/100_news_fake_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9732636468748819, + "test_acc": 0.9732696897374702, + "test_auc": 0.9787123515387475, + "val_auc": 0.9851994851994852, + "k": 1, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431 + ] + }, + { + "test_f1": 0.975648934477878, + "test_acc": 0.9756563245823389, + "test_auc": 0.9927756148063899, + "val_auc": 0.9922779922779923, + "k": 2, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928 + ] + }, + { + "test_f1": 0.9756518200292755, + "test_acc": 0.9756563245823389, + "test_auc": 0.9971064181922906, + "val_auc": 0.9974259974259974, + "k": 5, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1758, + 1500, + 126 + ] + }, + { + "test_f1": 0.9918852381515783, + "test_acc": 0.9918854415274463, + "test_auc": 0.9992144039312613, + "val_auc": 0.9978549978549979, + "k": 10, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1758, + 1500, + 126, + 461, + 1472, + 1411, + 5951, + 786 + ] + }, + { + "test_f1": 0.9961813738077403, + "test_acc": 0.99618138424821, + "test_auc": 0.9994714086776467, + "val_auc": 0.9978549978549979, + "k": 16, + "dataset": "100_news_fake", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 431, + 1928, + 1758, + 1500, + 126, + 461, + 1472, + 1411, + 5951, + 786, + 542, + 1319, + 1634, + 1848, + 7026, + 26453 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e402383c20441a3068812201286717853302775e --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/105_click_bait_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9196289689409414, + "test_acc": 0.9196370706416073, + "test_auc": 0.9514122699139131, + "val_auc": 0.9392310872261266, + "k": 1, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838 + ] + }, + { + "test_f1": 0.9513922461931187, + "test_acc": 0.9513933895009722, + "test_auc": 0.9654736463646566, + "val_auc": 0.9578338156262919, + "k": 2, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441 + ] + }, + { + "test_f1": 0.9630454720797459, + "test_acc": 0.9630589760207389, + "test_auc": 0.9676417814156972, + "val_auc": 0.965274906986358, + "k": 5, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 943, + 1411 + ] + }, + { + "test_f1": 0.9610995162141833, + "test_acc": 0.9611147116007777, + "test_auc": 0.982189539189398, + "val_auc": 0.9838776353865234, + "k": 10, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 943, + 1411, + 1417, + 1697, + 1500, + 1180, + 601 + ] + }, + { + "test_f1": 0.9636946819065899, + "test_acc": 0.9637070641607258, + "test_auc": 0.9855883281923079, + "val_auc": 0.9818106655642828, + "k": 16, + "dataset": "105_click_bait", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6838, + 1441, + 4110, + 943, + 1411, + 1417, + 1697, + 1500, + 1180, + 601, + 654, + 908, + 30737, + 1703, + 698, + 302 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..98da5437b6c3b49ac2b634964dded1b513fc0a23 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/106_hate_hate_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5635188682639607, + "test_acc": 0.5844025157232704, + "test_auc": 0.5900346720483951, + "val_auc": 0.6038415366146458, + "k": 1, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921 + ] + }, + { + "test_f1": 0.5682799806216777, + "test_acc": 0.5876729559748428, + "test_auc": 0.6077284036377296, + "val_auc": 0.6074429771908764, + "k": 2, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777 + ] + }, + { + "test_f1": 0.5866085695927861, + "test_acc": 0.5979874213836478, + "test_auc": 0.6208168740677583, + "val_auc": 0.6066426570628252, + "k": 5, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 2046 + ] + }, + { + "test_f1": 0.5960416493847434, + "test_acc": 0.5984905660377359, + "test_auc": 0.6244034412818127, + "val_auc": 0.6674669867947178, + "k": 10, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 2046, + 1380, + 969, + 559, + 505, + 1335 + ] + }, + { + "test_f1": 0.5992487043038841, + "test_acc": 0.6050314465408805, + "test_auc": 0.6276615404556174, + "val_auc": 0.6542617046818727, + "k": 16, + "dataset": "106_hate_hate", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 2921, + 777, + 7284, + 1438, + 2046, + 1380, + 969, + 559, + 505, + 1335, + 6411, + 1178, + 1319, + 7396, + 521, + 394 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..9789f7090c32074890e9d2d115e4646451744e08 --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/107_hate_offensive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6689725579354602, + "test_acc": 0.6767295597484276, + "test_auc": 0.6871375459602103, + "val_auc": 0.7460984393757504, + "k": 1, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567 + ] + }, + { + "test_f1": 0.677172733069669, + "test_acc": 0.6792452830188679, + "test_auc": 0.7242691174728291, + "val_auc": 0.7579031612645059, + "k": 2, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 1164 + ] + }, + { + "test_f1": 0.7055829397227843, + "test_acc": 0.7079245283018868, + "test_auc": 0.7644990476325493, + "val_auc": 0.7931172468987595, + "k": 5, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 1164, + 1345, + 4470, + 4196 + ] + }, + { + "test_f1": 0.7281145219382662, + "test_acc": 0.730062893081761, + "test_auc": 0.7789873361963428, + "val_auc": 0.7975190076030412, + "k": 10, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 1164, + 1345, + 4470, + 4196, + 4505, + 1629, + 6411, + 28711, + 559 + ] + }, + { + "test_f1": 0.7303966151465403, + "test_acc": 0.7315723270440252, + "test_auc": 0.7808519461003565, + "val_auc": 0.802721088435374, + "k": 16, + "dataset": "107_hate_offensive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 567, + 1164, + 1345, + 4470, + 4196, + 4505, + 1629, + 6411, + 28711, + 559, + 1731, + 794, + 1335, + 993, + 19328, + 433 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/scr/saebench_scr_custom_sae_eval_results.json b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/scr/saebench_scr_custom_sae_eval_results.json new file mode 100644 index 0000000000000000000000000000000000000000..1f1520c6bbefb476d194564e6bb967ec95522f2b --- /dev/null +++ b/gemma-2-2b-btk-mat/k-100/seed-0/snapshots/step-195312-tokens-800000000/scr/saebench_scr_custom_sae_eval_results.json @@ -0,0 +1,326 @@ +{ + "eval_type_id": "scr", + "eval_config": { + "random_seed": 42, + "dataset_names": [ + "LabHC/bias_in_bios_class_set1", + "canrager/amazon_reviews_mcauley_1and5" + ], + "perform_scr": true, + "early_stopping_patience": 20, + "train_set_size": 4000, + "test_set_size": 1000, + "context_length": 128, + "probe_train_batch_size": 16, + "probe_test_batch_size": 500, + "probe_epochs": 20, + "probe_lr": 0.001, + "probe_l1_penalty": 0.001, + "sae_batch_size": 32, + "llm_batch_size": 128, + "llm_dtype": "float32", + "lower_vram_usage": false, + "model_name": "gemma-2-2b", + "n_values": [ + 2, + 5, + 10 + ], + "column1_vals_lookup": { + "LabHC/bias_in_bios_class_set1": [ + [ + "professor", + "nurse" + ], + [ + "architect", + "journalist" + ], + [ + "surgeon", + "psychologist" + ], + [ + "attorney", + "teacher" + ] + ], + "canrager/amazon_reviews_mcauley_1and5": [ + [ + "Books", + "CDs_and_Vinyl" + ], + [ + "Software", + "Electronics" + ], + [ + "Pet_Supplies", + "Office_Products" + ], + [ + "Industrial_and_Scientific", + "Toys_and_Games" + ] + ] + } + }, + "eval_id": "99789dba-7ea7-4745-901c-d3a198956604", + "datetime_epoch_millis": 1776879220450, + "eval_result_metrics": { + "scr_metrics": { + "scr_dir1_threshold_2": 0.24127658442703298, + "scr_metric_threshold_2": 0.1761015801770613, + "scr_dir2_threshold_2": 0.16945369041379896, + "scr_dir1_threshold_5": -0.05499874126302609, + "scr_metric_threshold_5": 0.2558299446096355, + "scr_dir2_threshold_5": 0.24432568207791602, + "scr_dir1_threshold_10": -0.013002158635238507, + "scr_metric_threshold_10": 0.307922923608865, + "scr_dir2_threshold_10": 0.28882897888563647, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + }, + "eval_result_details": [ + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_professor_nurse_results", + "scr_dir1_threshold_2": 0.5000003820810118, + "scr_metric_threshold_2": 0.06217613139950766, + "scr_dir2_threshold_2": 0.06217613139950766, + "scr_dir1_threshold_5": 0.5897435701496917, + "scr_metric_threshold_5": 0.21502589713474632, + "scr_dir2_threshold_5": 0.21502589713474632, + "scr_dir1_threshold_10": 0.48718003580863245, + "scr_metric_threshold_10": 0.23056989138058243, + "scr_dir2_threshold_10": 0.23056989138058243, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_architect_journalist_results", + "scr_dir1_threshold_2": 0.24242422874090183, + "scr_metric_threshold_2": 0.2813455702086162, + "scr_dir2_threshold_2": 0.2813455702086162, + "scr_dir1_threshold_5": -0.5454541760043493, + "scr_metric_threshold_5": 0.29969413887699026, + "scr_dir2_threshold_5": 0.29969413887699026, + "scr_dir1_threshold_10": -0.5833329194122803, + "scr_metric_threshold_10": 0.3822629713004566, + "scr_dir2_threshold_10": 0.3822629713004566, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_surgeon_psychologist_results", + "scr_dir1_threshold_2": 0.3749997671693945, + "scr_metric_threshold_2": 0.037688550204100946, + "scr_dir2_threshold_2": 0.037688550204100946, + "scr_dir1_threshold_5": -0.7343747380655689, + "scr_metric_threshold_5": 0.1457286913802262, + "scr_dir2_threshold_5": 0.1457286913802262, + "scr_dir1_threshold_10": -0.5468753201420825, + "scr_metric_threshold_10": 0.1457286913802262, + "scr_dir2_threshold_10": 0.1457286913802262, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "LabHC/bias_in_bios_class_set1_scr_attorney_teacher_results", + "scr_dir1_threshold_2": 0.18120773420326786, + "scr_metric_threshold_2": 0.07324841127051662, + "scr_dir2_threshold_2": 0.07324841127051662, + "scr_dir1_threshold_5": -0.7516782309020104, + "scr_metric_threshold_5": 0.06050953237690028, + "scr_dir2_threshold_5": 0.06050953237690028, + "scr_dir1_threshold_10": -0.7449669074187787, + "scr_metric_threshold_10": 0.044585981215805634, + "scr_dir2_threshold_10": 0.044585981215805634, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Books_CDs_and_Vinyl_results", + "scr_dir1_threshold_2": 0.03684225881529179, + "scr_metric_threshold_2": 0.5039683206086106, + "scr_dir2_threshold_2": 0.5039683206086106, + "scr_dir1_threshold_5": 0.052631529414422, + "scr_metric_threshold_5": 0.5753969089319204, + "scr_dir2_threshold_5": 0.5753969089319204, + "scr_dir1_threshold_10": 0.07368420392192293, + "scr_metric_threshold_10": 0.6269840124753451, + "scr_dir2_threshold_10": 0.6269840124753451, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Software_Electronics_results", + "scr_dir1_threshold_2": 0.245901447122854, + "scr_metric_threshold_2": 0.11450392446499717, + "scr_dir2_threshold_2": 0.11450392446499717, + "scr_dir1_threshold_5": 0.377049276438573, + "scr_metric_threshold_5": 0.17938926174649328, + "scr_dir2_threshold_5": 0.17938926174649328, + "scr_dir1_threshold_10": 0.4754099855711136, + "scr_metric_threshold_10": 0.25954203162185363, + "scr_dir2_threshold_10": 0.25954203162185363, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Pet_Supplies_Office_Products_results", + "scr_dir1_threshold_2": 0.22325567727222573, + "scr_metric_threshold_2": 0.21030055324882435, + "scr_dir2_threshold_2": 0.21030055324882435, + "scr_dir1_threshold_5": 0.3488370996994263, + "scr_metric_threshold_5": 0.3476393872642003, + "scr_dir2_threshold_5": 0.3476393872642003, + "scr_dir1_threshold_10": 0.44186052958857275, + "scr_metric_threshold_10": 0.48068668649565843, + "scr_dir2_threshold_10": 0.48068668649565843, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + }, + { + "dataset_name": "canrager/amazon_reviews_mcauley_1and5_scr_Industrial_and_Scientific_Toys_and_Games_results", + "scr_dir1_threshold_2": 0.12558118001131657, + "scr_metric_threshold_2": 0.12558118001131657, + "scr_dir2_threshold_2": 0.07239806190521783, + "scr_dir1_threshold_5": 0.22325573916560695, + "scr_metric_threshold_5": 0.22325573916560695, + "scr_dir2_threshold_5": 0.1312216389118512, + "scr_dir1_threshold_10": 0.29302312300099165, + "scr_metric_threshold_10": 0.29302312300099165, + "scr_dir2_threshold_10": 0.14027156521516332, + "scr_dir1_threshold_20": null, + "scr_metric_threshold_20": null, + "scr_dir2_threshold_20": null, + "scr_dir1_threshold_50": null, + "scr_metric_threshold_50": null, + "scr_dir2_threshold_50": null, + "scr_dir1_threshold_100": null, + "scr_metric_threshold_100": null, + "scr_dir2_threshold_100": null, + "scr_dir1_threshold_500": null, + "scr_metric_threshold_500": null, + "scr_dir2_threshold_500": null + } + ], + "sae_bench_commit_hash": "Unknown", + "sae_lens_id": "custom_sae", + "sae_lens_release_id": "saebench_scr", + "sae_lens_version": "6.39.0", + "sae_cfg_dict": { + "d_in": 2304, + "d_sae": 32768, + "dtype": "float32", + "device": "cuda", + "apply_b_dec_to_input": true, + "normalize_activations": "none", + "reshape_activations": "none", + "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, + "hook_layer": 12, + "dataset_trust_remote_code": null + }, + "architecture": "jumprelu" + }, + "eval_result_unstructured": null +} \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2759cf681191c3492e6066812e2fbfec7427961c --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/147_disease_class_nervous system diseases_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5616269904487614, + "test_acc": 0.5656565656565656, + "test_auc": 0.5787755102040816, + "val_auc": 0.553939393939394, + "k": 1, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5736 + ] + }, + { + "test_f1": 0.5801977526115457, + "test_acc": 0.5858585858585859, + "test_auc": 0.5857142857142856, + "val_auc": 0.5696969696969697, + "k": 2, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5736, + 12000 + ] + }, + { + "test_f1": 0.5412375481440526, + "test_acc": 0.5454545454545454, + "test_auc": 0.5424489795918367, + "val_auc": 0.6173737373737374, + "k": 5, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5736, + 12000, + 28699, + 17153, + 25334 + ] + }, + { + "test_f1": 0.5344029492965663, + "test_acc": 0.5353535353535354, + "test_auc": 0.5608163265306123, + "val_auc": 0.6363636363636364, + "k": 10, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5736, + 12000, + 28699, + 17153, + 25334, + 18045, + 10212, + 23815, + 18869, + 24057 + ] + }, + { + "test_f1": 0.5451761671012463, + "test_acc": 0.5454545454545454, + "test_auc": 0.5612244897959183, + "val_auc": 0.6577777777777778, + "k": 16, + "dataset": "147_disease_class_nervous system diseases", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5736, + 12000, + 28699, + 17153, + 25334, + 18045, + 10212, + 23815, + 18869, + 24057, + 29738, + 27750, + 18897, + 11008, + 11371, + 19959 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..5019a8e154463f54f5b2eafd2c3641004c24465d --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/157_amazon_5star_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7716399729302954, + "test_acc": 0.7784615384615384, + "test_auc": 0.7873649308243849, + "val_auc": 0.7413300693594451, + "k": 1, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15970 + ] + }, + { + "test_f1": 0.7725420942519388, + "test_acc": 0.7805128205128206, + "test_auc": 0.8397936513279698, + "val_auc": 0.7974296205630355, + "k": 2, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15970, + 20140 + ] + }, + { + "test_f1": 0.781889518141571, + "test_acc": 0.7887179487179488, + "test_auc": 0.8553413336923956, + "val_auc": 0.8076295389636883, + "k": 5, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15970, + 20140, + 9609, + 12674, + 18064 + ] + }, + { + "test_f1": 0.7878802216134486, + "test_acc": 0.7938461538461539, + "test_auc": 0.8578701989430101, + "val_auc": 0.8561811505507956, + "k": 10, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15970, + 20140, + 9609, + 12674, + 18064, + 19655, + 6720, + 4282, + 11408, + 10377 + ] + }, + { + "test_f1": 0.7964289627205229, + "test_acc": 0.801025641025641, + "test_auc": 0.8742888881408422, + "val_auc": 0.8655650754793962, + "k": 16, + "dataset": "157_amazon_5star", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 15970, + 20140, + 9609, + 12674, + 18064, + 19655, + 6720, + 4282, + 11408, + 10377, + 28056, + 19698, + 14074, + 23530, + 7305, + 23430 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..ead5825d66372b3cdd8d29a713d68cb432745452 --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/159_code_Python_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7231133806930882, + "test_acc": 0.7297478991596639, + "test_auc": 0.7562429496496518, + "val_auc": 0.8069048574869531, + "k": 1, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12475 + ] + }, + { + "test_f1": 0.7226202489703148, + "test_acc": 0.7307563025210084, + "test_auc": 0.7562429496496518, + "val_auc": 0.8069048574869531, + "k": 2, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12475, + 25694 + ] + }, + { + "test_f1": 0.7940504786824634, + "test_acc": 0.7966386554621848, + "test_auc": 0.8908052132821369, + "val_auc": 0.9020473705339223, + "k": 5, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12475, + 25694, + 13090, + 22652, + 16980 + ] + }, + { + "test_f1": 0.7997737420541741, + "test_acc": 0.801344537815126, + "test_auc": 0.9024746729722107, + "val_auc": 0.939783219590526, + "k": 10, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12475, + 25694, + 13090, + 22652, + 16980, + 12447, + 32259, + 3916, + 22573, + 17641 + ] + }, + { + "test_f1": 0.8151476720701379, + "test_acc": 0.8157983193277311, + "test_auc": 0.9164217121866209, + "val_auc": 0.9522280208751506, + "k": 16, + "dataset": "159_code_Python", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12475, + 25694, + 13090, + 22652, + 16980, + 12447, + 32259, + 3916, + 22573, + 17641, + 3728, + 14170, + 8680, + 27718, + 8670, + 11085 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..37f368bcdfa1e60863c148fcd12beaa808d2f8df --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/160_code_HTML_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7502197295491544, + "test_acc": 0.7596638655462185, + "test_auc": 0.7554877486604334, + "val_auc": 0.7045363307908471, + "k": 1, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9823 + ] + }, + { + "test_f1": 0.7468240309630676, + "test_acc": 0.7573109243697479, + "test_auc": 0.7858767924159924, + "val_auc": 0.7482938578883983, + "k": 2, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9823, + 19144 + ] + }, + { + "test_f1": 0.7723050212819547, + "test_acc": 0.7757983193277311, + "test_auc": 0.8657380541756152, + "val_auc": 0.8627057406663992, + "k": 5, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9823, + 19144, + 22573, + 20861, + 13662 + ] + }, + { + "test_f1": 0.771322598310516, + "test_acc": 0.773781512605042, + "test_auc": 0.8828692756578518, + "val_auc": 0.8936170212765957, + "k": 10, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9823, + 19144, + 22573, + 20861, + 13662, + 23321, + 3377, + 8626, + 5265, + 32259 + ] + }, + { + "test_f1": 0.8271664994556434, + "test_acc": 0.8275630252100841, + "test_auc": 0.9219702475215306, + "val_auc": 0.9213167402649538, + "k": 16, + "dataset": "160_code_HTML", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 9823, + 19144, + 22573, + 20861, + 13662, + 23321, + 3377, + 8626, + 5265, + 32259, + 2289, + 2139, + 17641, + 14255, + 26430, + 25694 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..15c3f32263fabf8f3519d4e6d4ed588eeadb6870 --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/163_agnews_2_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9468903242538758, + "test_acc": 0.9468907563025211, + "test_auc": 0.9896253642681012, + "val_auc": 0.9859494179044561, + "k": 1, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10212 + ] + }, + { + "test_f1": 0.9768032316395986, + "test_acc": 0.9768067226890756, + "test_auc": 0.9981086079354404, + "val_auc": 0.9971898835808912, + "k": 2, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10212, + 21960 + ] + }, + { + "test_f1": 0.9774752199329724, + "test_acc": 0.9774789915966386, + "test_auc": 0.9981800153299926, + "val_auc": 0.9971898835808912, + "k": 5, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10212, + 21960, + 13279, + 12301, + 23872 + ] + }, + { + "test_f1": 0.9956302422263833, + "test_acc": 0.9956302521008403, + "test_auc": 0.9998956005813827, + "val_auc": 1.0, + "k": 10, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10212, + 21960, + 13279, + 12301, + 23872, + 9262, + 25694, + 21666, + 7814, + 17641 + ] + }, + { + "test_f1": 0.998655462184874, + "test_acc": 0.998655462184874, + "test_auc": 0.9999873455250161, + "val_auc": 1.0, + "k": 16, + "dataset": "163_agnews_2", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 10212, + 21960, + 13279, + 12301, + 23872, + 9262, + 25694, + 21666, + 7814, + 17641, + 22652, + 23110, + 13090, + 29219, + 29019, + 12475 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..dd918e342838acc891deac994799bd62f64f446c --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/22_headline_isobama_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6391723175193075, + "test_acc": 0.6465927099841522, + "test_auc": 0.6555505324492665, + "val_auc": 0.6828731492597039, + "k": 1, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23846 + ] + }, + { + "test_f1": 0.6546354241020873, + "test_acc": 0.6656101426307448, + "test_auc": 0.7458107293550331, + "val_auc": 0.7651060424169668, + "k": 2, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23846, + 168 + ] + }, + { + "test_f1": 0.7385089876089024, + "test_acc": 0.7385103011093502, + "test_auc": 0.8100813743218807, + "val_auc": 0.8191276510604242, + "k": 5, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23846, + 168, + 19998, + 8754, + 19830 + ] + }, + { + "test_f1": 0.7729444469190712, + "test_acc": 0.7733755942947702, + "test_auc": 0.8541390395820776, + "val_auc": 0.8859543817527011, + "k": 10, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23846, + 168, + 19998, + 8754, + 19830, + 16457, + 10377, + 564, + 3748, + 23023 + ] + }, + { + "test_f1": 0.7717013576139543, + "test_acc": 0.7717908082408875, + "test_auc": 0.873789431384368, + "val_auc": 0.8775510204081632, + "k": 16, + "dataset": "22_headline_isobama", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23846, + 168, + 19998, + 8754, + 19830, + 16457, + 10377, + 564, + 3748, + 23023, + 21960, + 6184, + 24791, + 29359, + 31583, + 26576 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..a24ace6821bbd3b6cc277b21b591ef9110683cbd --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/24_headline_isiran_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5883443534349138, + "test_acc": 0.5895316804407713, + "test_auc": 0.5996903648837351, + "val_auc": 0.6038415366146459, + "k": 1, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6141 + ] + }, + { + "test_f1": 0.5864315861581296, + "test_acc": 0.5867768595041323, + "test_auc": 0.6245825997207213, + "val_auc": 0.657062825130052, + "k": 2, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6141, + 16457 + ] + }, + { + "test_f1": 0.6162879204203152, + "test_acc": 0.6170798898071626, + "test_auc": 0.6902738145832069, + "val_auc": 0.7302921168467387, + "k": 5, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6141, + 16457, + 24500, + 168, + 19998 + ] + }, + { + "test_f1": 0.6852195098648067, + "test_acc": 0.6859504132231405, + "test_auc": 0.7554489709185842, + "val_auc": 0.7759103641456582, + "k": 10, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6141, + 16457, + 24500, + 168, + 19998, + 6653, + 7507, + 23772, + 3981, + 11658 + ] + }, + { + "test_f1": 0.6960934069785891, + "test_acc": 0.696969696969697, + "test_auc": 0.7664076255236475, + "val_auc": 0.7907162865146059, + "k": 16, + "dataset": "24_headline_isiran", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 6141, + 16457, + 24500, + 168, + 19998, + 6653, + 7507, + 23772, + 3981, + 11658, + 3753, + 26576, + 4342, + 16385, + 564, + 7492 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..124e2df9185e2b84fb6099e395e56899344870e7 --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/36_sciq_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7517861082201058, + "test_acc": 0.7564779874213836, + "test_auc": 0.7731933878054436, + "val_auc": 0.795718287314926, + "k": 1, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688 + ] + }, + { + "test_f1": 0.7835245464171122, + "test_acc": 0.7849056603773585, + "test_auc": 0.829544959743362, + "val_auc": 0.870748299319728, + "k": 2, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688, + 6993 + ] + }, + { + "test_f1": 0.8087870968254766, + "test_acc": 0.8088050314465409, + "test_auc": 0.8896353713625487, + "val_auc": 0.9315726290516206, + "k": 5, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688, + 6993, + 2836, + 26457, + 1469 + ] + }, + { + "test_f1": 0.8260315084539296, + "test_acc": 0.8261635220125786, + "test_auc": 0.9089511907884144, + "val_auc": 0.9643857543017207, + "k": 10, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688, + 6993, + 2836, + 26457, + 1469, + 24666, + 22490, + 8811, + 28267, + 28408 + ] + }, + { + "test_f1": 0.8303360727986769, + "test_acc": 0.830440251572327, + "test_auc": 0.9103446800582051, + "val_auc": 0.9659863945578232, + "k": 16, + "dataset": "36_sciq_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688, + 6993, + 2836, + 26457, + 1469, + 24666, + 22490, + 8811, + 28267, + 28408, + 22184, + 8, + 16319, + 29558, + 31057, + 16088 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..891145f3997aee18df8d56a3d725c3c67922428e --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/42_temp_sense_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6582678595988606, + "test_acc": 0.6613844870725605, + "test_auc": 0.6784585420144686, + "val_auc": 0.7049114331723026, + "k": 1, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688 + ] + }, + { + "test_f1": 0.6712389930825892, + "test_acc": 0.6755629691409508, + "test_auc": 0.7202671118530886, + "val_auc": 0.7765700483091788, + "k": 2, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688, + 1646 + ] + }, + { + "test_f1": 0.6742052979906918, + "test_acc": 0.6780650542118432, + "test_auc": 0.7476697273233166, + "val_auc": 0.8103864734299517, + "k": 5, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688, + 1646, + 10758, + 10826, + 267 + ] + }, + { + "test_f1": 0.7473462476971617, + "test_acc": 0.7481234361968306, + "test_auc": 0.8390915414579856, + "val_auc": 0.8645330112721417, + "k": 10, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688, + 1646, + 10758, + 10826, + 267, + 2946, + 25113, + 5800, + 8000, + 2836 + ] + }, + { + "test_f1": 0.7708536060840564, + "test_acc": 0.7714762301918265, + "test_auc": 0.8593183082915972, + "val_auc": 0.8872785829307568, + "k": 16, + "dataset": "42_temp_sense", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 22688, + 1646, + 10758, + 10826, + 267, + 2946, + 25113, + 5800, + 8000, + 2836, + 13231, + 24666, + 23805, + 14274, + 12592, + 8722 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..daf15897d499a9bba88301fbd474d87fc7ab64c0 --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/47_reasoning_tf_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5070882673576845, + "test_acc": 0.5451825254338719, + "test_auc": 0.5477265564564651, + "val_auc": 0.5600160578081093, + "k": 1, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12094 + ] + }, + { + "test_f1": 0.5668965151192307, + "test_acc": 0.5703171753441053, + "test_auc": 0.5795712402945306, + "val_auc": 0.6013649136892815, + "k": 2, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12094, + 6608 + ] + }, + { + "test_f1": 0.5666120710370223, + "test_acc": 0.5703171753441053, + "test_auc": 0.5948815288084119, + "val_auc": 0.660377358490566, + "k": 5, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12094, + 6608, + 7288, + 31406, + 11843 + ] + }, + { + "test_f1": 0.5659146361331393, + "test_acc": 0.5685218432076601, + "test_auc": 0.5977344354353494, + "val_auc": 0.5881172219991971, + "k": 10, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12094, + 6608, + 7288, + 31406, + 11843, + 3813, + 26062, + 14050, + 15840, + 27261 + ] + }, + { + "test_f1": 0.573632063868767, + "test_acc": 0.5739078396169958, + "test_auc": 0.5953900810818554, + "val_auc": 0.5981533520674428, + "k": 16, + "dataset": "47_reasoning_tf", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 12094, + 6608, + 7288, + 31406, + 11843, + 3813, + 26062, + 14050, + 15840, + 27261, + 30869, + 15351, + 26814, + 13851, + 14412, + 8178 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..2bb1e3daa06cf1f29062d1048a692edb7070245b --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/57_wikidatais_alive_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.761214063569384, + "test_acc": 0.7662893081761006, + "test_auc": 0.7860752841153615, + "val_auc": 0.7627050820328132, + "k": 1, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14321 + ] + }, + { + "test_f1": 0.793759631400134, + "test_acc": 0.7952201257861635, + "test_auc": 0.8614949637432041, + "val_auc": 0.8755502200880352, + "k": 2, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14321, + 26903 + ] + }, + { + "test_f1": 0.9035824322275798, + "test_acc": 0.9036477987421384, + "test_auc": 0.9726366502993806, + "val_auc": 0.9467787114845938, + "k": 5, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14321, + 26903, + 1372, + 30818, + 1918 + ] + }, + { + "test_f1": 0.9303085904041086, + "test_acc": 0.930314465408805, + "test_auc": 0.9813485847141226, + "val_auc": 0.9623849539815925, + "k": 10, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14321, + 26903, + 1372, + 30818, + 1918, + 31587, + 10747, + 5229, + 31305, + 97 + ] + }, + { + "test_f1": 0.9557179680312342, + "test_acc": 0.9557232704402516, + "test_auc": 0.9907417580470239, + "val_auc": 0.9727891156462585, + "k": 16, + "dataset": "57_wikidatais_alive", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 14321, + 26903, + 1372, + 30818, + 1918, + 31587, + 10747, + 5229, + 31305, + 97, + 2995, + 12075, + 16210, + 18958, + 27102, + 13754 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..c9f172a862239819b7039a40c48ff82efc9b8215 --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/70_prime-factors_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9779229982787728, + "test_acc": 0.9779310344827586, + "test_auc": 0.995354524054332, + "val_auc": 0.9991883116883117, + "k": 1, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17406 + ] + }, + { + "test_f1": 0.999080459770115, + "test_acc": 0.999080459770115, + "test_auc": 0.9999991544455868, + "val_auc": 1.0, + "k": 2, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17406, + 17959 + ] + }, + { + "test_f1": 0.999080459770115, + "test_acc": 0.999080459770115, + "test_auc": 0.9999991544455868, + "val_auc": 1.0, + "k": 5, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17406, + 17959, + 17629, + 1323, + 8087 + ] + }, + { + "test_f1": 1.0, + "test_acc": 1.0, + "test_auc": 1.0, + "val_auc": 1.0, + "k": 10, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17406, + 17959, + 17629, + 1323, + 8087, + 2044, + 23936, + 29869, + 15290, + 18548 + ] + }, + { + "test_f1": 0.9995402296906769, + "test_acc": 0.9995402298850574, + "test_auc": 0.9997463336760648, + "val_auc": 1.0, + "k": 16, + "dataset": "70_prime-factors", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 17406, + 17959, + 17629, + 1323, + 8087, + 2044, + 23936, + 29869, + 15290, + 18548, + 19455, + 158, + 28694, + 26731, + 26306, + 25113 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..1787e396fc35a987d616b61bff63ccaa9a9dc6c8 --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/77_second-derivative_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9931034045399655, + "test_acc": 0.993103448275862, + "test_auc": 0.9975694538394935, + "val_auc": 1.0, + "k": 1, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 934 + ] + }, + { + "test_f1": 0.9921838155197102, + "test_acc": 0.992183908045977, + "test_auc": 0.997771118567022, + "val_auc": 1.0, + "k": 2, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 934, + 13416 + ] + }, + { + "test_f1": 0.9944827096363509, + "test_acc": 0.9944827586206897, + "test_auc": 0.9966668245034904, + "val_auc": 1.0, + "k": 5, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 934, + 13416, + 24707, + 15290, + 2109 + ] + }, + { + "test_f1": 0.9958620584689465, + "test_acc": 0.9958620689655172, + "test_auc": 0.9972561759294334, + "val_auc": 1.0, + "k": 10, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 934, + 13416, + 24707, + 15290, + 2109, + 4667, + 18745, + 1572, + 23703, + 30294 + ] + }, + { + "test_f1": 0.9958620199810231, + "test_acc": 0.9958620689655172, + "test_auc": 0.9973474958060502, + "val_auc": 1.0, + "k": 16, + "dataset": "77_second-derivative", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 934, + 13416, + 24707, + 15290, + 2109, + 4667, + 18745, + 1572, + 23703, + 30294, + 3266, + 8087, + 18397, + 6524, + 17024, + 24168 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..35c81762577f8d8405c13b0185d8a09a98fb3d21 --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/79_human-rights_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9181593625274971, + "test_acc": 0.9181609195402299, + "test_auc": 0.9674782861626712, + "val_auc": 0.9732142857142856, + "k": 1, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25599 + ] + }, + { + "test_f1": 0.9383907785507376, + "test_acc": 0.9383908045977012, + "test_auc": 0.9873116104767574, + "val_auc": 0.9849837662337662, + "k": 2, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25599, + 15290 + ] + }, + { + "test_f1": 0.9383890854544993, + "test_acc": 0.9383908045977012, + "test_auc": 0.9890847380810649, + "val_auc": 0.984172077922078, + "k": 5, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25599, + 15290, + 5146, + 15524, + 15805 + ] + }, + { + "test_f1": 0.9397687144029784, + "test_acc": 0.9397701149425287, + "test_auc": 0.9911504275123113, + "val_auc": 0.9906655844155845, + "k": 10, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25599, + 15290, + 5146, + 15524, + 15805, + 26238, + 22733, + 18745, + 9374, + 18601 + ] + }, + { + "test_f1": 0.954482566183558, + "test_acc": 0.9544827586206897, + "test_auc": 0.9925092334541913, + "val_auc": 0.992288961038961, + "k": 16, + "dataset": "79_human-rights", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 25599, + 15290, + 5146, + 15524, + 15805, + 26238, + 22733, + 18745, + 9374, + 18601, + 10826, + 9157, + 30706, + 30757, + 657, + 1914 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..26e943cc67072acb9da5d498a77a3d811a8521f5 --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/80_side-effects_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.9650530388082914, + "test_acc": 0.9650574712643678, + "test_auc": 0.9873585387466854, + "val_auc": 1.0, + "k": 1, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4633 + ] + }, + { + "test_f1": 0.9562912928508356, + "test_acc": 0.9563218390804598, + "test_auc": 0.9873585387466854, + "val_auc": 1.0, + "k": 2, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4633, + 15290 + ] + }, + { + "test_f1": 0.9562912928508356, + "test_acc": 0.9563218390804598, + "test_auc": 0.9873585387466854, + "val_auc": 1.0, + "k": 5, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4633, + 15290, + 2836, + 24666, + 30759 + ] + }, + { + "test_f1": 0.9562912928508356, + "test_acc": 0.9563218390804598, + "test_auc": 0.9873585387466854, + "val_auc": 1.0, + "k": 10, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4633, + 15290, + 2836, + 24666, + 30759, + 5127, + 32095, + 7113, + 16088, + 16696 + ] + }, + { + "test_f1": 0.9562912928508356, + "test_acc": 0.9563218390804598, + "test_auc": 0.9873585387466854, + "val_auc": 1.0, + "k": 16, + "dataset": "80_side-effects", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 4633, + 15290, + 2836, + 24666, + 30759, + 5127, + 32095, + 7113, + 16088, + 16696, + 21919, + 25113, + 19495, + 8356, + 890, + 6993 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..e95549884178fd375ba28657885caca00a9f70ed --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/90_glue_qnli_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.6349893530888924, + "test_acc": 0.6354716981132076, + "test_auc": 0.6684282848576106, + "val_auc": 0.6762705082032814, + "k": 1, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23499 + ] + }, + { + "test_f1": 0.6659779398125026, + "test_acc": 0.6666666666666666, + "test_auc": 0.7320681765479644, + "val_auc": 0.7304921968787513, + "k": 2, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23499, + 25388 + ] + }, + { + "test_f1": 0.7032457883929025, + "test_acc": 0.7033962264150944, + "test_auc": 0.7889085899392327, + "val_auc": 0.7775110044017607, + "k": 5, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23499, + 25388, + 6110, + 3265, + 3212 + ] + }, + { + "test_f1": 0.7204362286724765, + "test_acc": 0.720503144654088, + "test_auc": 0.8041224194690033, + "val_auc": 0.7889155662264906, + "k": 10, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23499, + 25388, + 6110, + 3265, + 3212, + 8780, + 22826, + 10377, + 26903, + 3997 + ] + }, + { + "test_f1": 0.7272750575752299, + "test_acc": 0.7272955974842767, + "test_auc": 0.8113983346480493, + "val_auc": 0.7977190876350541, + "k": 16, + "dataset": "90_glue_qnli", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 23499, + 25388, + 6110, + 3265, + 3212, + 8780, + 22826, + 10377, + 26903, + 3997, + 5487, + 16180, + 29019, + 20081, + 17055, + 4342 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..683ca7dc2b288bf60f1d02f2b8a449d794cf0c1e --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/91_glue_qqp_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.7424352063970951, + "test_acc": 0.7426415094339622, + "test_auc": 0.7663649233093579, + "val_auc": 0.675670268107243, + "k": 1, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8005 + ] + }, + { + "test_f1": 0.7455669764244945, + "test_acc": 0.7464150943396226, + "test_auc": 0.8035756056216514, + "val_auc": 0.7382953181272509, + "k": 2, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8005, + 10709 + ] + }, + { + "test_f1": 0.7697741656409249, + "test_acc": 0.769811320754717, + "test_auc": 0.8325664859818195, + "val_auc": 0.8163265306122449, + "k": 5, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8005, + 10709, + 20043, + 32209, + 11203 + ] + }, + { + "test_f1": 0.7690115775885709, + "test_acc": 0.7690566037735849, + "test_auc": 0.8425140677988413, + "val_auc": 0.864545818327331, + "k": 10, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8005, + 10709, + 20043, + 32209, + 11203, + 30613, + 23634, + 1889, + 25894, + 21953 + ] + }, + { + "test_f1": 0.7778517654963643, + "test_acc": 0.7778616352201257, + "test_auc": 0.8524454477240899, + "val_auc": 0.8631452581032413, + "k": 16, + "dataset": "91_glue_qqp", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 8005, + 10709, + 20043, + 32209, + 11203, + 30613, + 23634, + 1889, + 25894, + 21953, + 27482, + 22989, + 12992, + 10377, + 12160, + 23805 + ] + } +] \ No newline at end of file diff --git a/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json new file mode 100644 index 0000000000000000000000000000000000000000..530a840d8b4d47817b8c66e3e87e6ac9cf925728 --- /dev/null +++ b/gemma-2-2b-btk/k-50/seed-0/snapshots/step-97656-tokens-400000000/sae_probes/sae_probes_gemma-2-2b/normal_setting/95_toxic_is_blocks.12.hook_resid_post_l1.json @@ -0,0 +1,101 @@ +[ + { + "test_f1": 0.5748169958696274, + "test_acc": 0.6262626262626263, + "test_auc": 0.626938775510204, + "val_auc": 0.6730769230769231, + "k": 1, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5334 + ] + }, + { + "test_f1": 0.6272499430394167, + "test_acc": 0.6363636363636364, + "test_auc": 0.6485714285714285, + "val_auc": 0.7708333333333334, + "k": 2, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5334, + 16088 + ] + }, + { + "test_f1": 0.7446113542392278, + "test_acc": 0.7474747474747475, + "test_auc": 0.7987755102040817, + "val_auc": 0.8557692307692307, + "k": 5, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5334, + 16088, + 19285, + 3881, + 10826 + ] + }, + { + "test_f1": 0.7473200928340257, + "test_acc": 0.7474747474747475, + "test_auc": 0.8277551020408163, + "val_auc": 0.8517628205128205, + "k": 10, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5334, + 16088, + 19285, + 3881, + 10826, + 10159, + 10223, + 7413, + 26731, + 10212 + ] + }, + { + "test_f1": 0.7473200928340257, + "test_acc": 0.7474747474747475, + "test_auc": 0.8391836734693877, + "val_auc": 0.8637820512820513, + "k": 16, + "dataset": "95_toxic_is", + "hook_name": "blocks.12.hook_resid_post", + "reg_type": "l1", + "binarize": false, + "indices": [ + 5334, + 16088, + 19285, + 3881, + 10826, + 10159, + 10223, + 7413, + 26731, + 10212, + 8754, + 23543, + 2615, + 1604, + 23799, + 23872 + ] + } +] \ No newline at end of file