Upload tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.8.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.8.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": 1745617889770,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.034275001287460326,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.04880000948905945,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.01452500820159912,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.09335000514984132,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.12610000967979432,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.032750004529953004,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.15155001133680346,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.22580001950263978,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.07425000816583634,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.17910000532865522,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.2895000159740448,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.11040001064538957,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.23420000821352005,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3863000214099884,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.15210001319646835,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.2484750121831894,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.42960002422332766,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.18112501204013823,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.21202503293752673,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.4476000487804413,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.23557501584291457
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.04569999873638153,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.07200000286102295,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.026300004124641417,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.1410499930381775,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.20260000228881836,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.06155000925064087,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.22605000138282777,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3658000111579895,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.13975000977516175,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.2153500109910965,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.4216000199317932,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.20625000894069673,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.1738000065088272,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.456600022315979,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.2828000158071518,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.13885000348091125,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.46760002374649046,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.3287500202655792,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.08020002245903016,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.4704000473022461,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.3902000248432159
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.022850003838539124,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.025600016117095947,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.002750012278556824,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.045650017261505124,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.04960001707077026,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.003949999809265137,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.07705002129077912,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.08580002784729004,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.008750006556510925,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.14284999966621398,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.1574000120162964,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.014550012350082398,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.2946000099182129,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.3160000205039978,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.02140001058578491,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.35810002088546755,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.3916000247001648,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.03350000381469727,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.34385004341602327,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.4248000502586365,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.08095000684261322
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.8.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.8.hook_resid_post",
|
| 163 |
+
"hook_layer": 8,
|
| 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": 0.12300002574920654,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.19600003957748413,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.07300001382827759,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.16875003278255463,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.26500004529953003,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.0962500125169754,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.19075001776218414,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3370000123977661,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.14624999463558197,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.17150002717971802,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.38700002431869507,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.21549999713897705,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.14350001513957977,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.43300002813339233,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.28950001299381256,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.12699998915195465,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.44700002670288086,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.3200000375509262,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.05675004422664642,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.4530000686645508,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.39625002443790436
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.07025003433227539,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.10000002384185791,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.02974998950958252,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.11925002932548523,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.1510000228881836,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.031749993562698364,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.22475002706050873,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.2990000247955322,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.0742499977350235,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.26175007224082947,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.3890000581741333,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.12724998593330383,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.20500002801418304,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.44200003147125244,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.2370000034570694,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.1220000684261322,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4580000638961792,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.335999995470047,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.06700006127357483,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4620000720024109,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.39500001072883606
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.00875002145767212,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.017000019550323486,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.008249998092651367,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.0742499828338623,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.12099999189376831,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.046750009059906006,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.14274998009204865,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.3100000023841858,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.16725002229213715,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.13974995911121368,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.38099998235702515,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.24125002324581146,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.13649998605251312,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.4390000104904175,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.30250002443790436,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.10924999415874481,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.4570000171661377,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.3477500230073929,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.05300001800060272,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.46000003814697266,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.40700002014636993
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.013999968767166138,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.0339999794960022,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.02000001072883606,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.1484999805688858,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.19099998474121094,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.042500004172325134,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.3632500320672989,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.46900004148483276,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.10575000941753387,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.29200002551078796,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.4830000400543213,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.19100001454353333,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.2172500044107437,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.48900002241134644,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.2717500180006027,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.19999997317790985,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.4959999918937683,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.29600001871585846,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.13450001180171967,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.4970000386238098,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.36250002682209015
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.012499943375587463,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.012999951839447021,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.0005000084638595581,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.1944999396800995,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.2849999666213989,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.09050002694129944,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.2087499499320984,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.4139999747276306,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.20525002479553223,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.21174997091293335,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.46799999475479126,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.2562500238418579,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.1667499989271164,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.48000001907348633,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.31325002014636993,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.13599999248981476,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.48000001907348633,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.34400002658367157,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.08974997699260712,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.48000001907348633,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.3902500420808792
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.0022500604391098022,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.001000046730041504,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": -0.0012500137090682983,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.003750056028366089,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.005000054836273193,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.0012499988079071045,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.021250054240226746,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.028000056743621826,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.006750002503395081,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.07750004529953003,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.08800005912780762,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.010500013828277588,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.23975002765655518,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.26100003719329834,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.021250009536743164,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.35475003719329834,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.39500004053115845,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.04025000333786011,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.39375007152557373,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.44800007343292236,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.05425000190734863
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.003000020980834961,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.012000024318695068,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.009000003337860107,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.02799999713897705,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.03799998760223389,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.009999990463256836,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.07700000703334808,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.09000003337860107,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.01300002634525299,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.14699998497962952,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.17100000381469727,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.02400001883506775,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.37975001335144043,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.4180000424385071,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.03825002908706665,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.39100004732608795,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.4410000443458557,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.04999999701976776,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.3580000251531601,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.4410000443458557,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.08300001919269562
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": 0.001999974250793457,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": 0.0,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": -0.001999974250793457,
|
| 352 |
+
"tpp_threshold_5_total_metric": 0.012000054121017456,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.01500004529953003,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.0029999911785125732,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.02100004255771637,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.029000043869018555,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.008000001311302185,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.07100000977516174,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.08300000429153442,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.01199999451637268,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.1847500205039978,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.19700002670288086,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.012250006198883057,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.294500008225441,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.3190000057220459,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.02449999749660492,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.26575006544589996,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.42100006341934204,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.15524999797344208
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.0755000114440918,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.08100003004074097,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.00550001859664917,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.11924998462200165,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.12099999189376831,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.0017500072717666626,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.18275001645088196,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.18800002336502075,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.005250006914138794,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.24925000965595245,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.26200002431869507,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.012750014662742615,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.37824998795986176,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.39399999380111694,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.015750005841255188,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.3997500091791153,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.42900002002716064,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.02925001084804535,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.36400003731250763,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.4320000410079956,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.06800000369548798
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.0314999520778656,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.0339999794960022,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.0025000274181365967,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.06524999439716339,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.0690000057220459,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.0037500113248825073,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.08324998617172241,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.09399998188018799,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.010749995708465576,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.1694999486207962,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.18299996852874756,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.013500019907951355,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.2905000001192093,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.3100000023841858,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.0195000022649765,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.3505000025033951,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.37400001287460327,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.02350001037120819,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.33775001764297485,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.38200002908706665,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.0442500114440918
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|