Upload tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.17.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.17.hook_resid_post_eval_results.json
ADDED
|
@@ -0,0 +1,419 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"eval_type_id": "tpp",
|
| 3 |
+
"eval_config": {
|
| 4 |
+
"random_seed": 42,
|
| 5 |
+
"dataset_names": [
|
| 6 |
+
"LabHC/bias_in_bios_class_set1",
|
| 7 |
+
"canrager/amazon_reviews_mcauley_1and5"
|
| 8 |
+
],
|
| 9 |
+
"perform_scr": false,
|
| 10 |
+
"early_stopping_patience": 20,
|
| 11 |
+
"train_set_size": 4000,
|
| 12 |
+
"test_set_size": 1000,
|
| 13 |
+
"context_length": 128,
|
| 14 |
+
"probe_train_batch_size": 16,
|
| 15 |
+
"probe_test_batch_size": 500,
|
| 16 |
+
"probe_epochs": 20,
|
| 17 |
+
"probe_lr": 0.001,
|
| 18 |
+
"probe_l1_penalty": 0.001,
|
| 19 |
+
"sae_batch_size": 125,
|
| 20 |
+
"llm_batch_size": 32,
|
| 21 |
+
"llm_dtype": "bfloat16",
|
| 22 |
+
"lower_vram_usage": false,
|
| 23 |
+
"model_name": "gemma-2-2b",
|
| 24 |
+
"n_values": [
|
| 25 |
+
2,
|
| 26 |
+
5,
|
| 27 |
+
10,
|
| 28 |
+
20,
|
| 29 |
+
50,
|
| 30 |
+
100,
|
| 31 |
+
500
|
| 32 |
+
],
|
| 33 |
+
"column1_vals_lookup": {
|
| 34 |
+
"LabHC/bias_in_bios_class_set1": [
|
| 35 |
+
[
|
| 36 |
+
"professor",
|
| 37 |
+
"nurse"
|
| 38 |
+
],
|
| 39 |
+
[
|
| 40 |
+
"architect",
|
| 41 |
+
"journalist"
|
| 42 |
+
],
|
| 43 |
+
[
|
| 44 |
+
"surgeon",
|
| 45 |
+
"psychologist"
|
| 46 |
+
],
|
| 47 |
+
[
|
| 48 |
+
"attorney",
|
| 49 |
+
"teacher"
|
| 50 |
+
]
|
| 51 |
+
],
|
| 52 |
+
"canrager/amazon_reviews_mcauley_1and5": [
|
| 53 |
+
[
|
| 54 |
+
"Books",
|
| 55 |
+
"CDs_and_Vinyl"
|
| 56 |
+
],
|
| 57 |
+
[
|
| 58 |
+
"Software",
|
| 59 |
+
"Electronics"
|
| 60 |
+
],
|
| 61 |
+
[
|
| 62 |
+
"Pet_Supplies",
|
| 63 |
+
"Office_Products"
|
| 64 |
+
],
|
| 65 |
+
[
|
| 66 |
+
"Industrial_and_Scientific",
|
| 67 |
+
"Toys_and_Games"
|
| 68 |
+
]
|
| 69 |
+
]
|
| 70 |
+
}
|
| 71 |
+
},
|
| 72 |
+
"eval_id": "65b37170-42a3-4f63-bfdc-b4728a912b3a",
|
| 73 |
+
"datetime_epoch_millis": 1745619438226,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.08117500543594361,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.09920001029968262,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.018025004863739015,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.14222501069307328,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.17980000972747803,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.037574999034404755,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.1682000070810318,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.24000000953674316,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.07180000245571136,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.18482501655817032,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.2980000197887421,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.11317500323057174,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.22115001976490023,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3771000266075134,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.15595000684261323,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.2330000028014183,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.4121000111103058,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.1791000083088875,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.22052503228187562,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.44700004458427434,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.2264750123023987
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.1408000111579895,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.17180001735687256,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.031000006198883056,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.23640001714229583,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.305400013923645,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.06899999678134919,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.24945000410079957,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3842000126838684,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.13475000858306885,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.22425001859664917,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.4284000277519226,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.20415000915527343,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.17915000915527343,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.459000027179718,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.2798500180244446,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.14975000023841858,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.46560001373291016,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.3158500134944916,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.08950002193450927,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.46680004596710206,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.3773000240325928
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.021549999713897705,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.026600003242492676,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.005050003528594971,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.04805000424385071,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.05420000553131103,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.006150001287460327,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.08695001006126404,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.09580000638961791,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.008849996328353881,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.14540001451969148,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.16760001182556153,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.022199997305870058,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.263150030374527,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.2952000260353088,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.03204999566078186,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.31625000536441805,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.3586000084877014,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.04235000312328339,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.35155004262924194,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.42720004320144656,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.07565000057220458
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.17.hook_resid_post",
|
| 151 |
+
"sae_lens_release_id": "gemma-2-2b-res-snap-matryoshka-dc",
|
| 152 |
+
"sae_lens_version": "5.9.1",
|
| 153 |
+
"sae_cfg_dict": {
|
| 154 |
+
"architecture": "jumprelu",
|
| 155 |
+
"d_in": 2304,
|
| 156 |
+
"d_sae": 32768,
|
| 157 |
+
"activation_fn_str": "relu",
|
| 158 |
+
"apply_b_dec_to_input": true,
|
| 159 |
+
"finetuning_scaling_factor": false,
|
| 160 |
+
"context_size": 1024,
|
| 161 |
+
"model_name": "gemma-2-2b",
|
| 162 |
+
"hook_name": "blocks.17.hook_resid_post",
|
| 163 |
+
"hook_layer": 17,
|
| 164 |
+
"hook_head_index": null,
|
| 165 |
+
"prepend_bos": true,
|
| 166 |
+
"dataset_path": "chanind/pile-uncopyrighted-gemma-1024-abbrv-1B",
|
| 167 |
+
"dataset_trust_remote_code": true,
|
| 168 |
+
"normalize_activations": "none",
|
| 169 |
+
"dtype": "torch.bfloat16",
|
| 170 |
+
"device": "cuda",
|
| 171 |
+
"sae_lens_training_version": "5.5.1",
|
| 172 |
+
"activation_fn_kwargs": {
|
| 173 |
+
"k": 40
|
| 174 |
+
},
|
| 175 |
+
"neuronpedia_id": null,
|
| 176 |
+
"model_from_pretrained_kwargs": {
|
| 177 |
+
"center_writing_weights": false
|
| 178 |
+
},
|
| 179 |
+
"seqpos_slice": [
|
| 180 |
+
null
|
| 181 |
+
]
|
| 182 |
+
},
|
| 183 |
+
"eval_result_unstructured": {
|
| 184 |
+
"LabHC/bias_in_bios_class_set1": {
|
| 185 |
+
"0": {
|
| 186 |
+
"tpp_threshold_2_total_metric": 0.21550005674362183,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.2640000581741333,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.048500001430511475,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.23600007593631744,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.3190000653266907,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.08299998939037323,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.22100003063678741,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3540000319480896,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.13300000131130219,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.19575007259845734,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.4030000567436218,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.2072499841451645,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.1637500524520874,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.43400007486343384,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.27025002241134644,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.13875003159046173,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.44600003957748413,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.3072500079870224,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.06450006365776062,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.44800007343292236,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.38350000977516174
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.18425001204013824,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.24900001287460327,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.06475000083446503,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.22200007736682892,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.28700006008148193,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.06499998271465302,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.19175003468990326,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.3540000319480896,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.16224999725818634,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.20000006258487701,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.40800005197525024,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.20799998939037323,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.1250000298023224,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4500000476837158,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.32500001788139343,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.09750001132488251,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.45600003004074097,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.35850001871585846,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.06275005638599396,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4580000638961792,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.39525000751018524
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.138000026345253,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.1690000295639038,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.031000003218650818,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.2239999920129776,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.33399999141693115,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.10999999940395355,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.22874999046325684,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.3889999985694885,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.1602500081062317,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.20675000548362732,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.4190000295639038,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.2122500240802765,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.1694999784231186,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.4440000057220459,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.2745000272989273,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.12824997305870056,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.45399999618530273,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.3257500231266022,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.06000000238418579,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.45500004291534424,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.39500004053115845
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.14649999141693115,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.15700000524520874,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.010500013828277588,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.3764999657869339,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.4509999752044678,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.07450000941753387,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.3559999614953995,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.4829999804496765,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.12700001895427704,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.3149999678134918,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.4909999966621399,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.17600002884864807,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.2502500116825104,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.49400001764297485,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.24375000596046448,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.2110000103712082,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.49400001764297485,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.28300000727176666,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.14274998009204865,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49400001764297485,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.3512500375509262
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.019749969244003296,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.019999980926513672,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.000250011682510376,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.12349997460842133,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.13599997758865356,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.012500002980232239,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.24975000321865082,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.3410000205039978,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.09125001728534698,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.20374998450279236,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.42100000381469727,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.2172500193119049,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.18724997341632843,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4729999899864197,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.28575001657009125,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.1732499748468399,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4779999852180481,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.3047500103712082,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.11750000715255737,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4790000319480896,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.3615000247955322
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.003500029444694519,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.008000016212463379,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.00449998676776886,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.01250004768371582,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.01900005340576172,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.0065000057220458984,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.021250024437904358,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.027000010013580322,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.005749985575675964,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.08750005066394806,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.1170000433921814,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.029499992728233337,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.24875007569789886,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.2820000648498535,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.03324998915195465,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.32975004613399506,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.3760000467300415,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.04625000059604645,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.39975006878376007,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4530000686645508,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.05324999988079071
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.04174995422363281,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.04999995231628418,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.008249998092651367,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.07874998450279236,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.08399999141693115,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.005250006914138794,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.12474998831748962,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.13499999046325684,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.010250002145767212,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.19350001215934753,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.23500001430511475,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.04150000214576721,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.3062500059604645,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.37400001287460327,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.0677500069141388,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.35124996304512024,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.4229999780654907,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.07175001502037048,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.3460000306367874,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.43800002336502075,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.09199999272823334
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -0.001749977469444275,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": -0.001999974250793457,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": -0.00024999678134918213,
|
| 352 |
+
"tpp_threshold_5_total_metric": -0.008999958634376526,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": -0.0029999613761901855,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.00599999725818634,
|
| 355 |
+
"tpp_threshold_10_total_metric": -0.003749936819076538,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.001000046730041504,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.004749983549118042,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.014500007033348083,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.018999993801116943,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.00449998676776886,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.07850006222724915,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.08900004625320435,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.0104999840259552,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.18375001847743988,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.203000009059906,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.019249990582466125,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.3147500604391098,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.43000006675720215,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.11525000631809235
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": -0.010500013828277588,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": -0.004999995231628418,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.00550001859664917,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.06024998426437378,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.06699997186660767,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.006749987602233887,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.13624994456768036,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.14699995517730713,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.01075001060962677,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.18274998664855957,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.19999998807907104,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.017250001430511475,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.3580000102519989,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.375,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.0169999897480011,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.3799999803304672,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.4089999794960022,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.028999999165534973,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.3722500056028366,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.42900002002716064,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.056750014424324036
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.07475000619888306,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.0820000171661377,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.007250010967254639,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.09774996340274811,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.10399997234344482,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.006250008940696716,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.1562500298023224,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.1690000295639038,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.012749999761581421,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.2487500160932541,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.2670000195503235,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.018250003457069397,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.3242499977350235,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.35600000619888306,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.03175000846385956,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.33650001883506775,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.38200002908706665,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.0455000102519989,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.3250000476837158,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.38600003719329834,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.06099998950958252
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|