Upload tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.19.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.19.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": 1745619808049,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.06397499144077301,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.07360000014305115,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.009625008702278138,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.16095000654459002,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.19380001425743104,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.03285000771284104,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.19397500455379485,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.2591000199317932,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.06512501537799835,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.21055000871419907,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.30880002379417415,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.09825001507997513,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.23247500360012055,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3744000196456909,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.14192501604557037,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.2522250100970268,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.4164000272750854,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.16417501717805863,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.22612503468990325,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.45150005221366885,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.22537501752376554
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.10694999396800994,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.1190000057220459,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.01205001175403595,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.25255001783370973,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.30580003261566163,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.053250014781951904,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.2869000047445297,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3998000264167786,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.11290002167224884,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.26420001685619354,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.43500003814697263,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.17080002129077912,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.21349999606609343,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.46140002012252807,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.24790002405643463,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.18240001797676086,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.46620004177093505,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.2838000237941742,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.09815002977848053,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.46780005693435667,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.36965002715587614
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.02099998891353607,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.028199994564056398,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.007200005650520325,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.06934999525547028,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.08179999589920044,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.012450000643730164,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.10105000436306,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.11840001344680787,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.017350009083747862,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.1569000005722046,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.18260000944137572,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.02570000886917114,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.25145001113414767,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.28740001916885377,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.03595000803470612,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.3220500022172928,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.3666000127792358,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.044550010561943056,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.35410003960132597,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.43520004749298097,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.08110000789165497
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.19.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.19.hook_resid_post",
|
| 163 |
+
"hook_layer": 19,
|
| 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": 0.14924998581409454,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.16699999570846558,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.017750009894371033,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.2567500025033951,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.32200002670288086,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.06525002419948578,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.2630000114440918,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3720000386238098,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.10900002717971802,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.20325002074241638,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.41300004720687866,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.20975002646446228,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.14949998259544373,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.4390000104904175,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.28950002789497375,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.11849997937679291,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.4480000138282776,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.3295000344514847,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.05925002694129944,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.4490000605583191,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.38975003361701965
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.11900004744529724,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.15300005674362183,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.034000009298324585,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.20725004374980927,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.28800004720687866,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.0807500034570694,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.24950000643730164,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.3460000157356262,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.09650000929832458,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.21600006520748138,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.4070000648498535,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.19099999964237213,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.17075000703334808,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4570000171661377,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.2862500101327896,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.14200004935264587,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4670000672340393,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.32500001788139343,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.0792500376701355,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4670000672340393,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.3877500295639038
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.0392499715089798,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.03799998760223389,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": -0.0012499839067459106,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.2134999930858612,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.22600001096725464,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.012500017881393433,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.2815000116825104,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.38200002908706665,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.10050001740455627,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.27025000751018524,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.4100000262260437,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.13975001871585846,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.25049997866153717,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.4440000057220459,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.19350002706050873,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.2302500456571579,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.4490000605583191,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.2187500149011612,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.06675003468990326,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.4540000557899475,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.38725002110004425
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.19249999523162842,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.19900000095367432,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.0065000057220458984,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.3432500511407852,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.4030000567436218,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.05975000560283661,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.36499999463558197,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.48000001907348633,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.11500002443790436,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.34175001084804535,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.4880000352859497,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.14625002443790436,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.257750004529953,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.4930000305175781,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.23525002598762512,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.2252500206232071,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.4930000305175781,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.26775000989437103,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.18575003743171692,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49500006437301636,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.30925002694129944
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.034749969840049744,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.03799998760223389,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.003250017762184143,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.24199999868869781,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.2900000214576721,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.048000022768974304,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.27549999952316284,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.4190000295639038,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.14350003004074097,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.28974997997283936,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.4570000171661377,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.16725003719329834,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.23900000751018524,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4740000367164612,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.23500002920627594,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.19599999487400055,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4740000367164612,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.27800004184246063,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.09975001215934753,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4740000367164612,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.37425002455711365
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.005750000476837158,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.009999990463256836,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.004249989986419678,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.004249989986419678,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.014999985694885254,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.010749995708465576,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.006999984383583069,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.02399998903274536,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.017000004649162292,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.05425003170967102,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.09000003337860107,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.035750001668930054,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.16875003278255463,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.21500003337860107,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.04625000059604645,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.26999999582767487,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.3230000138282776,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.05300001800060272,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.38075006008148193,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4540000557899475,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.07324999570846558
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.053749993443489075,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.06400001049041748,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.010250017046928406,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.13125000894069672,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.1420000195503235,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.01075001060962677,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.18050003051757812,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.20000004768371582,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.019500017166137695,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.27650000154972076,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.2990000247955322,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.022500023245811462,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.3475000113248825,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.3920000195503235,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.04450000822544098,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.3945000320672989,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.4410000443458557,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.046500012278556824,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.3717500567436218,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.4440000653266907,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.07225000858306885
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -0.006750017404556274,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": -0.0040000081062316895,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.002750009298324585,
|
| 352 |
+
"tpp_threshold_5_total_metric": -0.01199999451637268,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": -0.0009999871253967285,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.011000007390975952,
|
| 355 |
+
"tpp_threshold_10_total_metric": -0.0002499818801879883,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.012000024318695068,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.012250006198883057,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.02924998104572296,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.046999990940093994,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.017750009894371033,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.14075002074241638,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.16500002145767212,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.024250000715255737,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.25049999356269836,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.28299999237060547,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.032499998807907104,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.3070000261068344,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.4320000410079956,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.1250000149011612
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.02899995446205139,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.04399996995925903,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.015000015497207642,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.16324998438358307,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.18699997663497925,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.02374999225139618,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.21774998307228088,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.24199998378753662,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.024250000715255737,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.24124999344348907,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.27799999713897705,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.036750003695487976,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.3099999725818634,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.3479999899864197,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.038000017404556274,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.34050001204013824,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.3960000276565552,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.05550001561641693,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.3670000284910202,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.437000036239624,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.07000000774860382
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.02325001358985901,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.027000010013580322,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.0037499964237213135,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.0599999874830246,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.06599998474121094,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.00599999725818634,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.1002500057220459,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.11400002241134644,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.013750016689300537,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.18324999511241913,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.19900000095367432,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.015750005841255188,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.2902500182390213,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.31700003147125244,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.02675001323223114,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.35474997758865356,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.38999998569488525,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.03525000810623169,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.34400002658367157,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.409000039100647,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.0650000125169754
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|