Upload tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.11.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.11.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": 1745618394683,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.0765000119805336,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.12230001091957092,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.04579999893903732,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.10167500376701355,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.19480000138282777,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.0931249976158142,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.11500000208616257,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.23920000195503233,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.12419999986886979,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.13050001859664917,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.28660001754760744,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.15609999895095825,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.18260000348091127,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3718000054359436,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.18920000195503234,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.2084250032901764,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.412500011920929,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.20407500863075256,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.2076750323176384,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.4441000342369079,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.23642500191926955
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.13225001692771912,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.22060000896453857,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.08834999203681945,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.15930000245571135,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.3381999969482422,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.17889999449253083,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.15724999606609344,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3937999963760376,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.23655000030994416,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.13380002081394196,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.4248000144958496,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.29099999368190765,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.10550000369548798,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.454200005531311,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.34870000183582306,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.0913000077009201,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.4626000165939331,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.371300008893013,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.06810002028942108,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.46760002374649046,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.3995000034570694
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.020750007033348082,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.024000012874603273,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.003250005841255188,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.04405000507831573,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.05140000581741333,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.007350000739097595,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.0727500081062317,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.0846000075340271,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.01184999942779541,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.12720001637935638,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.14840002059936525,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.02120000422000885,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.2597000032663345,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.2894000053405762,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.029700002074241637,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.3255499988794327,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.3624000072479248,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.03685000836849213,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.3472500443458557,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.42060004472732543,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.07335000038146973
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.11.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.11.hook_resid_post",
|
| 163 |
+
"hook_layer": 11,
|
| 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": 0.1692500114440918,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.28200000524520874,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.11274999380111694,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.18074999749660492,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.33399999141693115,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.15324999392032623,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.13349996507167816,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.37699997425079346,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.2435000091791153,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.08249999582767487,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.40799999237060547,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.3254999965429306,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.06825003027915955,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.4240000247955322,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.3557499945163727,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.07075002789497375,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.43800002336502075,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.367249995470047,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.041250020265579224,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.44700002670288086,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.40575000643730164
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.1445000022649765,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.2459999918937683,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.10149998962879181,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.15825000405311584,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.296999990940094,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.13874998688697815,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.1652500182390213,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.35600000619888306,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.19074998795986176,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.11925004422664642,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.38700002431869507,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.26774998009204865,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.09349997341632843,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.44599997997283936,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.3525000065565109,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.07325001060962677,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4610000252723694,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.3877500146627426,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.05925002694129944,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4650000333786011,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.40575000643730164
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.09100005030632019,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.21500003337860107,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.12399998307228088,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.09900003671646118,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.32600003480911255,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.22699999809265137,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.09549997746944427,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.37599998712539673,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.28050000965595245,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.09925003349781036,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.41100001335144043,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.31174997985363007,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.07325002551078796,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.44300001859664917,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.3697499930858612,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.07049998641014099,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.45399999618530273,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.38350000977516174,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.046750038862228394,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.45500004291534424,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.40825000405311584
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.14174999296665192,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.20399999618530273,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.06225000321865082,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.22974999248981476,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.3579999804496765,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.12824998795986176,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.2695000022649765,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.4449999928474426,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.17549999058246613,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.2395000010728836,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.4710000157356262,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.23150001466274261,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.1600000113248825,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.48000001907348633,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.3200000077486038,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.1262499988079071,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.48100000619888306,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.35475000739097595,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.10799998044967651,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.4919999837875366,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.3840000033378601
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.11475002765655518,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.156000018119812,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.041249990463256836,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.12874998152256012,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.37599998712539673,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.2472500056028366,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.12250001728534698,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.4150000214576721,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.29250000417232513,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.12850002944469452,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.44700002670288086,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.31849999725818634,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.13249997794628143,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4779999852180481,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.34550000727176666,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.1157500147819519,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4790000319480896,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.3632500171661377,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.08525003492832184,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4790000319480896,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.39374999701976776
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.013000011444091797,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.013000011444091797,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.0,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.010749995708465576,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.013999998569488525,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.0032500028610229492,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.011999979615211487,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.02499997615814209,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.012999996542930603,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.0492500364780426,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.07600003480911255,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.026749998331069946,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.19025000929832458,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.22100001573562622,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.030750006437301636,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.28975000977516174,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.3270000219345093,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.037250012159347534,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.39500005543231964,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4500000476837158,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.05499999225139618
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.017499953508377075,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.019999980926513672,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.0025000274181365967,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.026499956846237183,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.030999958515167236,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.004500001668930054,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.051499977707862854,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.06199997663497925,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.010499998927116394,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.08224999904632568,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.11100000143051147,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.02875000238418579,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.2789999842643738,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.3219999670982361,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.042999982833862305,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.3699999451637268,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.41899996995925903,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.04900002479553223,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.3347500264644623,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.43800002336502075,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.10324999690055847
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": 0.0017500221729278564,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": 0.003000020980834961,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.0012499988079071045,
|
| 352 |
+
"tpp_threshold_5_total_metric": 0.005250036716461182,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.016000032424926758,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.010749995708465576,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.017000049352645874,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.029000043869018555,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.01199999451637268,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.05550001561641693,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.06800001859664917,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.012500002980232239,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.1497500091791153,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.16600000858306885,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.016249999403953552,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.26200005412101746,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.28300005197525024,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.020999997854232788,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.3590000867843628,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.42500007152557373,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.06599998474121094
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.015000015497207642,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.022000014781951904,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.006999999284744263,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.07625001668930054,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.08700001239776611,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.010749995708465576,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.12475000321865082,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.1380000114440918,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.013250008225440979,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.20225001871585846,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.2250000238418579,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.02275000512599945,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.3435000032186508,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.37300002574920654,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.029500022530555725,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.36274999380111694,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.40799999237060547,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.045249998569488525,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.33650003373622894,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.4140000343322754,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.07750000059604645
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.05650003254413605,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.06200003623962402,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.005500003695487976,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.1015000194311142,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.10900002717971802,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.007500007748603821,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.15850003063678741,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.1690000295639038,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.010499998927116394,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.24675001204013824,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.26200002431869507,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.015250012278556824,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.3360000103712082,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.36500000953674316,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.028999999165534973,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.34324999153614044,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.375,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.03175000846385956,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.3110000193119049,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.3760000467300415,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.0650000274181366
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|