Frozen Stage 1 error classifiers (combined/workspace/logprob) + meta
Browse files
classifiers/gemma-4-12b-it/frozen_meta.json
ADDED
|
@@ -0,0 +1,242 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "lightgbm",
|
| 3 |
+
"config": {
|
| 4 |
+
"n_estimators": 300,
|
| 5 |
+
"learning_rate": 0.05,
|
| 6 |
+
"num_leaves": 31,
|
| 7 |
+
"min_child_samples": 20,
|
| 8 |
+
"subsample": 0.9,
|
| 9 |
+
"colsample_bytree": 0.9,
|
| 10 |
+
"class_weight": "balanced",
|
| 11 |
+
"random_state": 0,
|
| 12 |
+
"deterministic": true
|
| 13 |
+
},
|
| 14 |
+
"n_train_rows": 16178,
|
| 15 |
+
"train_rows_by_dataset": {
|
| 16 |
+
"trivia_qa": 3000,
|
| 17 |
+
"popqa": 3000,
|
| 18 |
+
"squad_v2": 3000,
|
| 19 |
+
"halubench": 1692,
|
| 20 |
+
"medhallu": 487,
|
| 21 |
+
"esconv": 499,
|
| 22 |
+
"hotpotqa": 2500,
|
| 23 |
+
"drop": 2000
|
| 24 |
+
},
|
| 25 |
+
"train_error_rate": 0.3388552355050068,
|
| 26 |
+
"train_input": "stage1_features.jsonl",
|
| 27 |
+
"families": {
|
| 28 |
+
"combined": {
|
| 29 |
+
"feature_names": [
|
| 30 |
+
"bl_answer_len",
|
| 31 |
+
"bl_first_token_logprob",
|
| 32 |
+
"bl_mean_logprob",
|
| 33 |
+
"bl_min_logprob",
|
| 34 |
+
"ent_early",
|
| 35 |
+
"ent_late",
|
| 36 |
+
"ent_max_jump",
|
| 37 |
+
"ent_mid",
|
| 38 |
+
"ent_n_dir_changes",
|
| 39 |
+
"ent_slope",
|
| 40 |
+
"ent_std",
|
| 41 |
+
"pfx_ent_delta",
|
| 42 |
+
"pfx_ent_max",
|
| 43 |
+
"pfx_ign_delta",
|
| 44 |
+
"pfx_ign_min",
|
| 45 |
+
"pfx_rank_delta",
|
| 46 |
+
"pfx_rank_max",
|
| 47 |
+
"shape_eff_k20_late",
|
| 48 |
+
"shape_eff_k20_mean",
|
| 49 |
+
"shape_rival_mass_late",
|
| 50 |
+
"shape_rival_mass_mean",
|
| 51 |
+
"shape_tail_mass_late",
|
| 52 |
+
"shape_tail_mass_mean",
|
| 53 |
+
"shape_top1_p_late",
|
| 54 |
+
"shape_top1_p_mean",
|
| 55 |
+
"ws_band_agreement",
|
| 56 |
+
"ws_best_hedge_rank_log",
|
| 57 |
+
"ws_ignition_depth",
|
| 58 |
+
"ws_ignition_frac",
|
| 59 |
+
"ws_mean_entropy",
|
| 60 |
+
"ws_mean_log_rank_answer"
|
| 61 |
+
],
|
| 62 |
+
"scaler_mean": [
|
| 63 |
+
25.05315861045865,
|
| 64 |
+
-0.17759722063044153,
|
| 65 |
+
-0.08564728337698219,
|
| 66 |
+
-0.48999179390064196,
|
| 67 |
+
3.5097295571145843,
|
| 68 |
+
0.9656747148905885,
|
| 69 |
+
7.757507380393137,
|
| 70 |
+
3.03901574823834,
|
| 71 |
+
11.530226233156139,
|
| 72 |
+
-0.14915573044714436,
|
| 73 |
+
2.636601144511192,
|
| 74 |
+
1.8378779996397074,
|
| 75 |
+
4.923152761911201,
|
| 76 |
+
0.08033028392467027,
|
| 77 |
+
0.00029103309020480473,
|
| 78 |
+
-1.4201457888984148,
|
| 79 |
+
11.621967572434453,
|
| 80 |
+
1.158498578316299,
|
| 81 |
+
1.7895239986401315,
|
| 82 |
+
0.009992559648905977,
|
| 83 |
+
0.06649042364630987,
|
| 84 |
+
0.18729164729880263,
|
| 85 |
+
0.20976582916100103,
|
| 86 |
+
0.7987580541476109,
|
| 87 |
+
0.7068893354143501,
|
| 88 |
+
4.1208225161742274e-05,
|
| 89 |
+
6.883650860813349,
|
| 90 |
+
0.9967934849796067,
|
| 91 |
+
0.0019239090122388518,
|
| 92 |
+
2.5048066260266064,
|
| 93 |
+
11.43785334412124
|
| 94 |
+
],
|
| 95 |
+
"scaler_scale": [
|
| 96 |
+
25.45752716643542,
|
| 97 |
+
0.3038479080200517,
|
| 98 |
+
0.12534207789154353,
|
| 99 |
+
0.4248399498109823,
|
| 100 |
+
0.33524849721297684,
|
| 101 |
+
0.9562015363239469,
|
| 102 |
+
0.549182723788699,
|
| 103 |
+
0.8816967083352109,
|
| 104 |
+
1.3023491088733605,
|
| 105 |
+
0.06205867710386437,
|
| 106 |
+
0.18028191312289407,
|
| 107 |
+
1.2330816974614625,
|
| 108 |
+
1.247578195138596,
|
| 109 |
+
0.08776144271626778,
|
| 110 |
+
0.003823243723796051,
|
| 111 |
+
1.3042709158780594,
|
| 112 |
+
0.5031065175799508,
|
| 113 |
+
1.0173983493160113,
|
| 114 |
+
0.3744807895526155,
|
| 115 |
+
0.024155810061547933,
|
| 116 |
+
0.02068268434655756,
|
| 117 |
+
0.2670197887905251,
|
| 118 |
+
0.05641887335486946,
|
| 119 |
+
0.28315736218053245,
|
| 120 |
+
0.07617318205238348,
|
| 121 |
+
0.0019650885453565345,
|
| 122 |
+
2.060719749069122,
|
| 123 |
+
0.035426640578511824,
|
| 124 |
+
0.013619528088853259,
|
| 125 |
+
0.632161754132111,
|
| 126 |
+
0.7096851345082171
|
| 127 |
+
],
|
| 128 |
+
"model_sha256": "e5a8892789e1965f4e04c50430b1685db45b9393c72768f497d961b7f3efc859"
|
| 129 |
+
},
|
| 130 |
+
"logprob": {
|
| 131 |
+
"feature_names": [
|
| 132 |
+
"bl_answer_len",
|
| 133 |
+
"bl_first_token_logprob",
|
| 134 |
+
"bl_mean_logprob",
|
| 135 |
+
"bl_min_logprob"
|
| 136 |
+
],
|
| 137 |
+
"scaler_mean": [
|
| 138 |
+
25.05315861045865,
|
| 139 |
+
-0.17759722063044153,
|
| 140 |
+
-0.08564728337698219,
|
| 141 |
+
-0.48999179390064196
|
| 142 |
+
],
|
| 143 |
+
"scaler_scale": [
|
| 144 |
+
25.45752716643542,
|
| 145 |
+
0.3038479080200517,
|
| 146 |
+
0.12534207789154353,
|
| 147 |
+
0.4248399498109823
|
| 148 |
+
],
|
| 149 |
+
"model_sha256": "e77a3b8c23705787b5265068374dee71a065d38c366c2cb430fb5cafe497e29d"
|
| 150 |
+
},
|
| 151 |
+
"workspace": {
|
| 152 |
+
"feature_names": [
|
| 153 |
+
"ent_early",
|
| 154 |
+
"ent_late",
|
| 155 |
+
"ent_max_jump",
|
| 156 |
+
"ent_mid",
|
| 157 |
+
"ent_n_dir_changes",
|
| 158 |
+
"ent_slope",
|
| 159 |
+
"ent_std",
|
| 160 |
+
"pfx_ent_delta",
|
| 161 |
+
"pfx_ent_max",
|
| 162 |
+
"pfx_ign_delta",
|
| 163 |
+
"pfx_ign_min",
|
| 164 |
+
"pfx_rank_delta",
|
| 165 |
+
"pfx_rank_max",
|
| 166 |
+
"shape_eff_k20_late",
|
| 167 |
+
"shape_eff_k20_mean",
|
| 168 |
+
"shape_rival_mass_late",
|
| 169 |
+
"shape_rival_mass_mean",
|
| 170 |
+
"shape_tail_mass_late",
|
| 171 |
+
"shape_tail_mass_mean",
|
| 172 |
+
"shape_top1_p_late",
|
| 173 |
+
"shape_top1_p_mean",
|
| 174 |
+
"ws_band_agreement",
|
| 175 |
+
"ws_best_hedge_rank_log",
|
| 176 |
+
"ws_ignition_depth",
|
| 177 |
+
"ws_ignition_frac",
|
| 178 |
+
"ws_mean_entropy",
|
| 179 |
+
"ws_mean_log_rank_answer"
|
| 180 |
+
],
|
| 181 |
+
"scaler_mean": [
|
| 182 |
+
3.5097295571145843,
|
| 183 |
+
0.9656747148905885,
|
| 184 |
+
7.757507380393137,
|
| 185 |
+
3.03901574823834,
|
| 186 |
+
11.530226233156139,
|
| 187 |
+
-0.14915573044714436,
|
| 188 |
+
2.636601144511192,
|
| 189 |
+
1.8378779996397074,
|
| 190 |
+
4.923152761911201,
|
| 191 |
+
0.08033028392467027,
|
| 192 |
+
0.00029103309020480473,
|
| 193 |
+
-1.4201457888984148,
|
| 194 |
+
11.621967572434453,
|
| 195 |
+
1.158498578316299,
|
| 196 |
+
1.7895239986401315,
|
| 197 |
+
0.009992559648905977,
|
| 198 |
+
0.06649042364630987,
|
| 199 |
+
0.18729164729880263,
|
| 200 |
+
0.20976582916100103,
|
| 201 |
+
0.7987580541476109,
|
| 202 |
+
0.7068893354143501,
|
| 203 |
+
4.1208225161742274e-05,
|
| 204 |
+
6.883650860813349,
|
| 205 |
+
0.9967934849796067,
|
| 206 |
+
0.0019239090122388518,
|
| 207 |
+
2.5048066260266064,
|
| 208 |
+
11.43785334412124
|
| 209 |
+
],
|
| 210 |
+
"scaler_scale": [
|
| 211 |
+
0.33524849721297684,
|
| 212 |
+
0.9562015363239469,
|
| 213 |
+
0.549182723788699,
|
| 214 |
+
0.8816967083352109,
|
| 215 |
+
1.3023491088733605,
|
| 216 |
+
0.06205867710386437,
|
| 217 |
+
0.18028191312289407,
|
| 218 |
+
1.2330816974614625,
|
| 219 |
+
1.247578195138596,
|
| 220 |
+
0.08776144271626778,
|
| 221 |
+
0.003823243723796051,
|
| 222 |
+
1.3042709158780594,
|
| 223 |
+
0.5031065175799508,
|
| 224 |
+
1.0173983493160113,
|
| 225 |
+
0.3744807895526155,
|
| 226 |
+
0.024155810061547933,
|
| 227 |
+
0.02068268434655756,
|
| 228 |
+
0.2670197887905251,
|
| 229 |
+
0.05641887335486946,
|
| 230 |
+
0.28315736218053245,
|
| 231 |
+
0.07617318205238348,
|
| 232 |
+
0.0019650885453565345,
|
| 233 |
+
2.060719749069122,
|
| 234 |
+
0.035426640578511824,
|
| 235 |
+
0.013619528088853259,
|
| 236 |
+
0.632161754132111,
|
| 237 |
+
0.7096851345082171
|
| 238 |
+
],
|
| 239 |
+
"model_sha256": "b535d47fe7c9f0889fd1aee5ed1b7bc0fb0a5dc49fe37fb77ec36b485dd559d6"
|
| 240 |
+
}
|
| 241 |
+
}
|
| 242 |
+
}
|
classifiers/gemma-4-12b-it/lgbm_stage1_combined.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
classifiers/gemma-4-12b-it/lgbm_stage1_logprob.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
classifiers/gemma-4-12b-it/lgbm_stage1_workspace.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|