Upload tpp/gemma-2-2b-res-matryoshka-dc_blocks.20.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-matryoshka-dc_blocks.20.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": "91ffb92c-ec01-47b1-9f45-773ec12047b3",
|
| 73 |
+
"datetime_epoch_millis": 1745752949766,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.07627499699592591,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.08829999566078187,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.012024998664855957,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.17157500833272935,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.21660000681877134,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.04502499848604202,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.2017000034451485,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.28240000605583193,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.08070000261068344,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.22342500388622283,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.33520000576972964,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.11177500188350678,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.24690000712871552,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.40460000634193416,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.15769999921321867,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.25012500435113905,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.43720000982284546,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.1870750054717064,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.2165500357747078,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.4524000406265259,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.2358500048518181
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.1166500061750412,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.13100000619888305,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.014350000023841857,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.2583000034093857,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.33560000658035277,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.07730000317096711,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.26240001022815707,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.403000020980835,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.14060001075267792,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.2395999997854233,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.43760000467300414,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.19800000488758088,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.18254999816417694,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.46060000658035277,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.27805000841617583,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.1420999974012375,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.4660000205039978,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.32390002310276034,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.08645002841949463,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.46680004596710206,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.38035001754760744
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.035899987816810607,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.04559998512268067,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.009699997305870057,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.084850013256073,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.09760000705718994,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.012749993801116943,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.1409999966621399,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.16179999113082885,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.020799994468688965,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.2072500079870224,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.23280000686645508,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.025549998879432677,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.3112500160932541,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.3486000061035156,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.037349990010261534,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.3581500113010406,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.4083999991416931,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.05024998784065247,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.346650043129921,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.4380000352859497,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.09134999215602875
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.20.hook_resid_post",
|
| 151 |
+
"sae_lens_release_id": "gemma-2-2b-res-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.20.hook_resid_post",
|
| 163 |
+
"hook_layer": 20,
|
| 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": 0.13100004196166992,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.14600002765655518,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.014999985694885254,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.2407500296831131,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.362000048160553,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.12125001847743988,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.2460000365972519,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.38600003719329834,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.14000000059604645,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.2680000215768814,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.42000001668930054,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.15199999511241913,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.16574999690055847,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.44300001859664917,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.2772500216960907,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.12349998950958252,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.4480000138282776,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.32450002431869507,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.054250046610832214,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.4490000605583191,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.3947500139474869
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.12599998712539673,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.15700000524520874,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.03100001811981201,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.18074999749660492,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.2639999985694885,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.0832500010728836,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.21550004184246063,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.33900004625320435,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.12350000441074371,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.1912500262260437,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.4100000262260437,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.21875,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.16700004041194916,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4500000476837158,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.28300000727176666,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.13600000739097595,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4610000252723694,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.32500001788139343,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.07675004005432129,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4630000591278076,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.38625001907348633
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.08149999380111694,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.08799999952316284,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.0065000057220458984,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.24724997580051422,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.2979999780654907,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.0507500022649765,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.2412499636411667,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.38999998569488525,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.14875002205371857,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.22124995291233063,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.4139999747276306,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.1927500218153,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.16799995303153992,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.44199997186660767,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.27400001883506775,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.13099993765354156,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.4509999752044678,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.3200000375509262,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.050499990582466125,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.4520000219345093,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.40150003135204315
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.21650004386901855,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.2290000319480896,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.012499988079071045,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.40000005066394806,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.46000003814697266,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.0599999874830246,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.34925004839897156,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.49100005626678467,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.1417500078678131,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.2892500162124634,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.49400001764297485,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.20475000143051147,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.2187500298023224,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.49400001764297485,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.27524998784065247,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.1897500604391098,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.49500006437301636,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.30525000393390656,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.14875005185604095,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49500006437301636,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.3462500125169754
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.028249964118003845,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.034999966621398926,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.006750002503395081,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.2227499634027481,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.29399996995925903,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.07125000655651093,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.25999996066093445,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.4089999794960022,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.14900001883506775,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.22824998199939728,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.44999998807907104,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.22175000607967377,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.19324997067451477,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4739999771118164,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.28075000643730164,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.1302499920129776,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4750000238418579,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.3447500318288803,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.10200001299381256,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4750000238418579,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.37300001084804535
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": -0.003000035881996155,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.006999969482421875,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.01000000536441803,
|
| 306 |
+
"tpp_threshold_5_total_metric": -0.0034999847412109375,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.008000016212463379,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.011500000953674316,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.01524999737739563,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.03299999237060547,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.01774999499320984,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.0637500137090683,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.08700001239776611,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.023249998688697815,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.18424999713897705,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.21399998664855957,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.02974998950958252,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.33000001311302185,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.3659999966621399,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.03599998354911804,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.40550005435943604,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.45500004291534424,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.0494999885559082
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.11225000023841858,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.12999999523162842,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.01774999499320984,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.20250000059604645,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.22699999809265137,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.02449999749660492,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.2852499932050705,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.3100000023841858,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.024750009179115295,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.35625001788139343,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.38700002431869507,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.030750006437301636,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.40299999713897705,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.4399999976158142,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.03700000047683716,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.3985000550746918,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.45100003480911255,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.052499979734420776,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.39775003492832184,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.45100003480911255,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.05324999988079071
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -0.011750012636184692,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": -0.003000020980834961,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.008749991655349731,
|
| 352 |
+
"tpp_threshold_5_total_metric": -0.0007499754428863525,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.017000019550323486,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.01774999499320984,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.03149999678134918,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.05199998617172241,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.02049998939037323,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.10399998724460602,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.13499999046325684,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.031000003218650818,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.2565000355243683,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.30900001525878906,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.052499979734420776,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.32750000059604645,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.39499998092651367,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.06749998033046722,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.29575005173683167,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.437000036239624,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.14124998450279236
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.03724996745586395,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.04399996995925903,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.006750002503395081,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.15775001049041748,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.16600000858306885,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.008249998092651367,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.26075001060962677,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.2919999957084656,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.031249985098838806,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.30675002932548523,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.3360000252723694,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.029249995946884155,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.35850003361701965,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.3960000276565552,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.03749999403953552,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.3619999885559082,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.4229999780654907,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.06099998950958252,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.2955000251531601,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.43300002813339233,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.13750000298023224
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.044750019907951355,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.050000011920928955,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.0052499920129776,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.06825001537799835,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.06999999284744263,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.001749977469444275,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.11224998533725739,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.12199997901916504,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.009749993681907654,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.20549999177455902,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.218999981880188,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.013499990105628967,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.3540000170469284,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.3840000033378601,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.0299999862909317,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.372749999165535,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.40700000524520874,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.03425000607967377,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.33875004947185516,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.4140000343322754,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.07524998486042023
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|