Upload tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.18.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.18.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": 1745619623162,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.05692499876022339,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.06589999794960022,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.008974999189376831,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.1515250042080879,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.1965000033378601,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.04497499912977219,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.18625001311302186,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.2580000102519989,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.07174999713897705,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.23035001009702682,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.33450000882148745,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.10414999872446061,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.2436500072479248,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.39450001120567324,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.15085000395774842,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.25500001311302184,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.4299000144004822,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.17490000128746033,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.2356000304222107,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.4502000391483307,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.21460000872612
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.08565000593662261,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.0974000096321106,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.011750003695487976,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.2209500104188919,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.2974000215530396,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.07645001113414765,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.23805002272129058,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.36360002756118776,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.12555000483989714,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.2441500097513199,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.42920001745224,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.18505000770092012,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.18515000343322754,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.457800018787384,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.2726500153541565,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.1491000235080719,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.4654000401496887,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.31630001664161683,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.08970001637935639,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.4676000475883484,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.377900031208992
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.028199991583824156,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.034399986267089844,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.006199994683265686,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.08209999799728393,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.09559998512268067,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.01349998712539673,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.1344500035047531,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.15239999294281006,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.017949989438056944,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.21655001044273375,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.23980000019073486,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.023249989748001097,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.30215001106262207,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.3312000036239624,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.029049992561340332,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.3609000027179718,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.3943999886512756,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.03349998593330383,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.381500044465065,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.43280003070831297,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.05129998624324798
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.18.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.18.hook_resid_post",
|
| 163 |
+
"hook_layer": 18,
|
| 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": 0.06900003552436829,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.07600003480911255,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.006999999284744263,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.21800002455711365,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.2890000343322754,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.07100000977516174,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.22950004041194916,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3760000467300415,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.14650000631809235,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.2147500216960907,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.4190000295639038,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.2042500078678131,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.18800005316734314,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.440000057220459,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.25200000405311584,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.14325004816055298,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.4490000605583191,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.3057500123977661,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.08250002562999725,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.4540000557899475,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.37150003015995026
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.06499996781349182,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.08099997043609619,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.01600000262260437,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.18800000846385956,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.2850000262260437,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.09700001776218414,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.22500000894069672,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.328000009059906,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.10300000011920929,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.1667499989271164,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.39800000190734863,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.23125000298023224,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.12874998152256012,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.453000009059906,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.3242500275373459,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.10175001621246338,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4650000333786011,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.3632500171661377,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.06350000202655792,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4650000333786011,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.40150003135204315
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.18275001645088196,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.19300001859664917,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.010250002145767212,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.3080000728368759,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.35600006580352783,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.047999992966651917,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.2915000468492508,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.3960000276565552,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.10449998080730438,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.2990000396966934,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.4280000329017639,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.1289999932050705,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.2057500183582306,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.44200003147125244,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.23625001311302185,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.1615000218153,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.45100003480911255,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.28950001299381256,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.06200005114078522,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.4530000686645508,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.39100001752376556
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.022250041365623474,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.03400003910064697,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.011749997735023499,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.15275000035762787,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.22100001573562622,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.06825001537799835,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.22100003063678741,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.32100003957748413,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.10000000894069672,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.3200000077486038,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.4570000171661377,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.13700000941753387,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.22550001740455627,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.48000001907348633,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.25450000166893005,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.19100002944469452,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.487000048160553,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.29600001871585846,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.14150002598762512,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49100005626678467,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.34950003027915955
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.08924996852874756,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.1029999852180481,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.013750016689300537,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.23799994587898254,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.3359999656677246,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.09800001978874207,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.22324998676776886,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.3970000147819519,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.17375002801418304,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.2202499806880951,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.4440000057220459,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.2237500250339508,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.17774994671344757,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4739999771118164,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.29625003039836884,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.14800000190734863,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4750000238418579,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.3270000219345093,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.0989999771118164,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4750000238418579,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.3760000467300415
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.013249948620796204,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.01799994707107544,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.004749998450279236,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.014749959111213684,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.026999950408935547,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.012249991297721863,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.06425000727176666,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.08799999952316284,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.02374999225139618,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.21950000524520874,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.2549999952316284,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.03549998998641968,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.32649996876716614,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.36799997091293335,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.04150000214576721,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.3934999853372574,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.437999963760376,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.04449997842311859,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.4032500237226486,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4570000171661377,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.053749993443489075
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.015750035643577576,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.026000022888183594,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.010249987244606018,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.15200001001358032,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.1809999942779541,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.02899998426437378,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.23850001394748688,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.27300000190734863,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.034499987959861755,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.32900001108646393,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.3659999966621399,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.036999985575675964,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.3785000443458557,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.4280000329017639,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.0494999885559082,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.38700003921985626,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.44300001859664917,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.05599997937679291,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.36750006675720215,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.4450000524520874,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.07749998569488525
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": 0.00449998676776886,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": 0.004999995231628418,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.0005000084638595581,
|
| 352 |
+
"tpp_threshold_5_total_metric": -0.004000023007392883,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.001999974250793457,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.00599999725818634,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.006749957799911499,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.012999951839447021,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.0062499940395355225,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.028749987483024597,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.03700000047683716,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.008250012993812561,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.10825000703334808,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.12099999189376831,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.012749984860420227,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.26399996876716614,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.2759999632835388,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.01199999451637268,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.3967500329017639,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.42900002002716064,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.03224998712539673
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.04224999248981476,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.04799997806549072,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.005749985575675964,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.10975003242492676,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.12000000476837158,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.010249972343444824,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.16300004720687866,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.17400002479553223,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.010999977588653564,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.22175002098083496,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.24000000953674316,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.018249988555908203,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.3395000398159027,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.3630000352859497,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.023499995470046997,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.3947499990463257,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.421999990940094,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.02724999189376831,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.3850000500679016,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.437000036239624,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.05199998617172241
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.06524999439716339,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.07499998807907104,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.009749993681907654,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.1380000114440918,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.14800000190734863,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.009999990463256836,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.19974999129772186,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.21399998664855957,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.014249995350837708,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.2837500274181366,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.3009999990463257,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.017249971628189087,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.3579999953508377,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.37599998712539673,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.01799999177455902,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.36525002121925354,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.3930000066757202,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.027749985456466675,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.3550000488758087,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.3960000276565552,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.04099997878074646
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|