Upload tpp/gemma-2-2b-res-matryoshka-dc_blocks.14.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-matryoshka-dc_blocks.14.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": 1745752684918,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.048400001227855684,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.06080000400543213,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.012400002777576446,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.1058000087738037,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.1442000091075897,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.03840000033378601,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.1433750107884407,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.22120001316070556,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.07782500237226486,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.1572250097990036,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.27350001931190493,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.11627500951290132,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.1887250140309334,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3616000235080719,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.17287500947713852,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.20962500721216204,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.40600002408027647,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.19637501686811448,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.20867503732442855,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.44170004725456236,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.2330250099301338
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.07679998874664307,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.0997999906539917,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.023000001907348633,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.1800000101327896,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.2520000100135803,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.07199999988079071,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.22305001616477965,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3720000147819519,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.14894999861717223,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.20475001335144044,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.4224000215530396,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.21765000820159913,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.1442500203847885,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.45700002908706666,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.31275000870227815,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.11514999866485595,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.46440001726150515,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.34925001859664917,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.06950003206729889,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.4670000433921814,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.3975000113248825
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.020000013709068298,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.021800017356872557,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.0018000036478042602,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.03160000741481781,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.036400008201599124,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.004800000786781311,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.06370000541210175,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.07040001153945923,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.006700006127357483,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.10970000624656677,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.12460001707077026,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.014900010824203492,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.23320000767707824,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.26620001792907716,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.033000010251998904,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.3041000157594681,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.34760003089904784,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.043500015139579774,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.3478500425815582,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.41640005111694334,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.06855000853538513
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.14.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.14.hook_resid_post",
|
| 163 |
+
"hook_layer": 14,
|
| 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": 0.06824998557567596,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.09200000762939453,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.023750022053718567,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.09974999725818634,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.1290000081062317,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.02925001084804535,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.2435000091791153,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.34700000286102295,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.10349999368190765,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.20500002801418304,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.4050000309944153,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.20000000298023224,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.15525001287460327,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.44700002670288086,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.2917500138282776,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.12324999272823334,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.45600003004074097,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.33275003731250763,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.05775001645088196,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.45600003004074097,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.398250013589859
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.1782499998807907,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.23600000143051147,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.057750001549720764,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.18949998915195465,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.2879999876022339,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.09849999845027924,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.20524999499320984,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.33799999952316284,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.132750004529953,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.19575002789497375,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.3960000276565552,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.20024999976158142,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.12450000643730164,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4350000023841858,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.31049999594688416,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.09799997508525848,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.45399999618530273,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.35600002110004425,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.05950002372264862,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.46000003814697266,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.40050001442432404
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.01925000548362732,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.02399998903274536,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.004749983549118042,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.2200000286102295,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.2900000214576721,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.06999999284744263,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.17750003933906555,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.34400004148483276,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.1665000021457672,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.17200003564357758,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.3810000419616699,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.20900000631809235,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.130000039935112,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.44600003957748413,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.31599999964237213,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.09649999439716339,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.453000009059906,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.3565000146627426,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.04675005376338959,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.4540000557899475,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.4072500020265579
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.008499979972839355,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.0059999823570251465,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": -0.002499997615814209,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.18775001168251038,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.22200000286102295,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.03424999117851257,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.3137500286102295,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.43800002336502075,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.12424999475479126,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.23750001192092896,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.4750000238418579,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.23750001192092896,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.16175003349781036,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.4830000400543213,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.3212500065565109,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.1432500034570694,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.48500001430511475,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.34175001084804535,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.1040000468492508,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49100005626678467,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.3870000094175339
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.10974997282028198,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.14099997282028198,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.03125,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.2030000239610672,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.33100003004074097,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.12800000607967377,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.17525000870227814,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.3930000066757202,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.21774999797344208,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.21349996328353882,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.45499998331069946,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.24150002002716064,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.1497500091791153,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4740000367164612,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.3242500275373459,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.11475002765655518,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4740000367164612,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.359250009059906,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.07950001955032349,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4740000367164612,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.3945000171661377
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.006999984383583069,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.009999990463256836,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.003000006079673767,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.003999948501586914,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.007999956607818604,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.0040000081062316895,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.018499940633773804,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.02599996328353882,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.007500022649765015,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.040249958634376526,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.05299997329711914,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.012750014662742615,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.12424996495246887,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.1589999794960022,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.034750014543533325,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.23399995267391205,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.2789999842643738,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.04500003159046173,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.3877499848604202,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4480000138282776,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.06025002896785736
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.03375004231929779,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.03400003910064697,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.00024999678134918213,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.04600006341934204,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.061000049114227295,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.014999985694885254,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.09799998998641968,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.11100000143051147,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.013000011444091797,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.15725001692771912,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.18800002336502075,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.030750006437301636,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.335750013589859,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.4140000343322754,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.07825002074241638,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.3370000123977661,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.42900002002716064,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.09200000762939453,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.29575005173683167,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.43000006675720215,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.13425001502037048
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -0.0007499754428863525,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": 0.0020000338554382324,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.002750009298324585,
|
| 352 |
+
"tpp_threshold_5_total_metric": -0.005499973893165588,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": -0.0059999823570251465,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": -0.0005000084638595581,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.015750035643577576,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.016000032424926758,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.00024999678134918213,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.03975002467632294,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.04700005054473877,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.0072500258684158325,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.1720000058412552,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.18400001525878906,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.012000009417533875,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.2835000455379486,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.3060000538825989,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.02250000834465027,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.3760000616312027,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.42100006341934204,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.04500000178813934
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.018250003457069397,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.018000006675720215,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": -0.00024999678134918213,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.03275001049041748,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.03600001335144043,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.0032500028610229492,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.04950004816055298,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.057000041007995605,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.007499992847442627,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.1352500319480896,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.15000003576278687,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.014750003814697266,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.22475004196166992,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.2420000433921814,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.017250001430511475,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.34150004386901855,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.362000048160553,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.020500004291534424,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.3692500740289688,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.4160000681877136,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.04674999415874481
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.04175001382827759,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.04500001668930054,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.0032500028610229492,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.0807499885559082,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.08300000429153442,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.0022500157356262207,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.1367500126361847,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.1420000195503235,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.005250006914138794,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.17599999904632568,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.1850000023841858,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.009000003337860107,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.30925001204013824,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.3320000171661377,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.02275000512599945,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.32450002431869507,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.362000048160553,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.03750002384185791,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.31050004065036774,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.3670000433921814,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.05650000274181366
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|