Upload tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.21.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.21.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": 1745620172671,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.05669999420642853,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.06759998798370362,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.010899993777275085,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.1655000016093254,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.19699999690055847,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.03149999529123306,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.22512501329183576,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.27330000400543214,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.04817499071359635,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.25917501002550125,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.33780000805854793,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.07862499803304672,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.27317501157522206,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.41500000953674315,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.14182499796152115,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.26687500923872,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.43790000677108765,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.17102499753236772,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.22522503435611724,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.45050003528594973,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.22527500092983246
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.042599993944168094,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.05339999198913574,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.010799998044967651,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.21880000829696655,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.2648000121116638,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.046000003814697266,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.2970000237226486,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3642000198364258,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.06719999611377717,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.31435000598430635,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.43160001039505,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.11725000441074371,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.23465001285076142,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.46040002107620237,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.22575000822544097,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.1905500113964081,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.4652000188827515,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.2746500074863434,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.10780002772808076,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.46620004177093505,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.3584000140428543
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.07079999446868897,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.08179998397827148,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.01099998950958252,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.11219999492168427,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.12919998168945312,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.01699998676776886,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.15325000286102294,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.18239998817443848,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.029149985313415526,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.20400001406669616,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.2440000057220459,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.03999999165534973,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.31170001029968264,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.36959999799728394,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.05789998769760132,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.3432000070810318,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.4105999946594238,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.06739998757839202,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.34265004098415375,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.43480002880096436,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.09214998781681061
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.21.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.21.hook_resid_post",
|
| 163 |
+
"hook_layer": 21,
|
| 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": 0.036999985575675964,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.04799997806549072,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.010999992489814758,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.20125004649162292,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.27900004386901855,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.07774999737739563,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.2460000365972519,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3400000333786011,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.09399999678134918,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.2514999955892563,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.40299999713897705,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.15150000154972076,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.22175000607967377,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.4399999976158142,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.21824999153614044,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.18199999630451202,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.4490000009536743,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.2670000046491623,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.09700003266334534,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.4500000476837158,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.3530000150203705
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.07875001430511475,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.10500001907348633,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.026250004768371582,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.17100003361701965,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.21100002527236938,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.03999999165534973,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.24425004422664642,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.33500003814697266,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.09074999392032623,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.3345000445842743,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.409000039100647,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.07449999451637268,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.19025002419948578,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.45500004291534424,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.26475001871585846,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.14900000393390656,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4610000252723694,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.31200002133846283,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.07950003445148468,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.46400004625320435,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.38450001180171967
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.03249995410442352,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.035999953746795654,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.0034999996423721313,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.23124997317790985,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.2789999842643738,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.04775001108646393,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.3034999817609787,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.3569999933242798,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.053500011563301086,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.3467499613761902,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.4229999780654907,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.07625001668930054,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.25324998795986176,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.4449999928474426,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.19175000488758087,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.19824998080730438,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.4509999752044678,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.2527499943971634,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.06849999725818634,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.4520000219345093,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.38350002467632294
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.048500001430511475,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.050999999046325684,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.002499997615814209,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.3032499849796295,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.3240000009536743,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.0207500159740448,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.42875002324581146,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.4710000157356262,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.04224999248981476,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.31700003147125244,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.4920000433921814,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.17500001192092896,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.25600001215934753,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.4930000305175781,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.2370000183582306,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.21575002372264862,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.4920000433921814,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.2762500196695328,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.16575004160404205,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.4920000433921814,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.32625000178813934
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.016250014305114746,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.027000010013580322,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.010749995708465576,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.18725000321865082,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.23100000619888306,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.04375000298023224,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.2625000327825546,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.31800001859664917,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.05549998581409454,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.32199999690055847,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.4309999942779541,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.10899999737739563,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.25200003385543823,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.46900004148483276,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.21700000762939453,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.20775005221366882,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.47300004959106445,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.26524999737739563,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.12825003266334534,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.47300004959106445,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.3447500169277191
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.015499964356422424,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.01799994707107544,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.002499982714653015,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.016499951481819153,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.026999950408935547,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.010499998927116394,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.027749955654144287,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.04599994421005249,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.018249988555908203,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.0664999932050705,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.0989999771118164,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.03249998390674591,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.20225000381469727,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.23199999332427979,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.02974998950958252,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.3194999694824219,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.34999996423721313,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.03049999475479126,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.4087500274181366,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4570000171661377,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.0482499897480011
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.08300000429153442,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.1029999852180481,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.019999980926513672,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.17800001800060272,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.19499999284744263,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.016999974846839905,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.2522500157356262,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.3059999942779541,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.05374997854232788,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.3282500058412552,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.4020000100135803,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.07375000417232513,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.3422500044107437,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.44099998474121094,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.09874998033046722,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.3397500514984131,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.44200003147125244,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.10224997997283936,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.32350002229213715,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.44200003147125244,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.1185000091791153
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -0.0025000423192977905,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": -0.001000046730041504,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.0014999955892562866,
|
| 352 |
+
"tpp_threshold_5_total_metric": 0.0012499690055847168,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.011999964714050293,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.010749995708465576,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.022499993443489075,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.042999982833862305,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.02049998939037323,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.08125001192092896,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.10699999332427979,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.02574998140335083,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.31974998116493225,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.4039999842643738,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.08425000309944153,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.3232499808073044,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.42799997329711914,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.10474999248981476,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.27400003373622894,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.42900002002716064,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.1549999862909317
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.14000003039836884,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.15500003099441528,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.015000000596046448,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.216000035405159,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.24000000953674316,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.023999974131584167,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.254750058054924,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.27900004386901855,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.024249985814094543,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.28450004756450653,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.3160000443458557,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.03149999678134918,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.37025004625320435,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.4010000228881836,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.030749976634979248,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.38600003719329834,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.43300002813339233,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.046999990940093994,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.3770000785589218,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.4360000491142273,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.05899997055530548
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.11800001561641693,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.1340000033378601,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.015999987721443176,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.14925000071525574,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.171999990940094,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.022749990224838257,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.20899999141693115,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.23799997568130493,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.02899998426437378,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.25950001180171967,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.29600000381469727,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.0364999920129776,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.3240000158548355,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.3700000047683716,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.04599998891353607,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.3474999964237213,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.3999999761581421,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.052499979734420776,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.33000004291534424,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.4100000262260437,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.07999998331069946
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|