Upload tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.20.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.20.hook_resid_post_eval_results.json
ADDED
|
@@ -0,0 +1,419 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"eval_type_id": "tpp",
|
| 3 |
+
"eval_config": {
|
| 4 |
+
"random_seed": 42,
|
| 5 |
+
"dataset_names": [
|
| 6 |
+
"LabHC/bias_in_bios_class_set1",
|
| 7 |
+
"canrager/amazon_reviews_mcauley_1and5"
|
| 8 |
+
],
|
| 9 |
+
"perform_scr": false,
|
| 10 |
+
"early_stopping_patience": 20,
|
| 11 |
+
"train_set_size": 4000,
|
| 12 |
+
"test_set_size": 1000,
|
| 13 |
+
"context_length": 128,
|
| 14 |
+
"probe_train_batch_size": 16,
|
| 15 |
+
"probe_test_batch_size": 500,
|
| 16 |
+
"probe_epochs": 20,
|
| 17 |
+
"probe_lr": 0.001,
|
| 18 |
+
"probe_l1_penalty": 0.001,
|
| 19 |
+
"sae_batch_size": 125,
|
| 20 |
+
"llm_batch_size": 32,
|
| 21 |
+
"llm_dtype": "bfloat16",
|
| 22 |
+
"lower_vram_usage": false,
|
| 23 |
+
"model_name": "gemma-2-2b",
|
| 24 |
+
"n_values": [
|
| 25 |
+
2,
|
| 26 |
+
5,
|
| 27 |
+
10,
|
| 28 |
+
20,
|
| 29 |
+
50,
|
| 30 |
+
100,
|
| 31 |
+
500
|
| 32 |
+
],
|
| 33 |
+
"column1_vals_lookup": {
|
| 34 |
+
"LabHC/bias_in_bios_class_set1": [
|
| 35 |
+
[
|
| 36 |
+
"professor",
|
| 37 |
+
"nurse"
|
| 38 |
+
],
|
| 39 |
+
[
|
| 40 |
+
"architect",
|
| 41 |
+
"journalist"
|
| 42 |
+
],
|
| 43 |
+
[
|
| 44 |
+
"surgeon",
|
| 45 |
+
"psychologist"
|
| 46 |
+
],
|
| 47 |
+
[
|
| 48 |
+
"attorney",
|
| 49 |
+
"teacher"
|
| 50 |
+
]
|
| 51 |
+
],
|
| 52 |
+
"canrager/amazon_reviews_mcauley_1and5": [
|
| 53 |
+
[
|
| 54 |
+
"Books",
|
| 55 |
+
"CDs_and_Vinyl"
|
| 56 |
+
],
|
| 57 |
+
[
|
| 58 |
+
"Software",
|
| 59 |
+
"Electronics"
|
| 60 |
+
],
|
| 61 |
+
[
|
| 62 |
+
"Pet_Supplies",
|
| 63 |
+
"Office_Products"
|
| 64 |
+
],
|
| 65 |
+
[
|
| 66 |
+
"Industrial_and_Scientific",
|
| 67 |
+
"Toys_and_Games"
|
| 68 |
+
]
|
| 69 |
+
]
|
| 70 |
+
}
|
| 71 |
+
},
|
| 72 |
+
"eval_id": "65b37170-42a3-4f63-bfdc-b4728a912b3a",
|
| 73 |
+
"datetime_epoch_millis": 1745619989291,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.06834999024868012,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.0835999846458435,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.015249994397163392,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.15797501504421235,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.19630001187324525,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.0383249968290329,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.18905001282691958,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.2542000114917755,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.06514999866485596,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.18907500952482223,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.3008000075817108,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.11172499805688858,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.22695002406835557,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.38240002393722533,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.15544999986886976,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.24377500563859938,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.4292000114917755,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.18542500585317612,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.21352503299713135,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.45220003724098207,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.2386750042438507
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.11694998741149902,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.13879998922348022,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.021850001811981202,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.2637000232934952,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.32420002222061156,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.0604999989271164,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.2847500115633011,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3958000183105469,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.11105000674724579,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.23505001068115233,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.43280001878738406,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.1977500081062317,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.1851000279188156,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.4618000388145447,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.27670001089572904,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.15394999384880065,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.46560001373291016,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.3116500198841095,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.09100003242492676,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.46680004596710206,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.3758000135421753
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.019749993085861207,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.028399980068206786,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.008649986982345582,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.052250006794929506,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.06840000152587891,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.016149994730949403,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.09335001409053803,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.11260000467300416,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.019249990582466125,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.14310000836849213,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.1687999963760376,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.02569998800754547,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.2688000202178955,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.303000009059906,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.0341999888420105,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.3336000174283981,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.39280000925064085,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.05919999182224274,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.33605003356933594,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.43760002851486207,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.10154999494552612
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.20.hook_resid_post",
|
| 151 |
+
"sae_lens_release_id": "gemma-2-2b-res-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.20.hook_resid_post",
|
| 163 |
+
"hook_layer": 20,
|
| 164 |
+
"hook_head_index": null,
|
| 165 |
+
"prepend_bos": true,
|
| 166 |
+
"dataset_path": "chanind/pile-uncopyrighted-gemma-1024-abbrv-1B",
|
| 167 |
+
"dataset_trust_remote_code": true,
|
| 168 |
+
"normalize_activations": "none",
|
| 169 |
+
"dtype": "torch.bfloat16",
|
| 170 |
+
"device": "cuda",
|
| 171 |
+
"sae_lens_training_version": "5.5.1",
|
| 172 |
+
"activation_fn_kwargs": {
|
| 173 |
+
"k": 40
|
| 174 |
+
},
|
| 175 |
+
"neuronpedia_id": null,
|
| 176 |
+
"model_from_pretrained_kwargs": {
|
| 177 |
+
"center_writing_weights": false
|
| 178 |
+
},
|
| 179 |
+
"seqpos_slice": [
|
| 180 |
+
null
|
| 181 |
+
]
|
| 182 |
+
},
|
| 183 |
+
"eval_result_unstructured": {
|
| 184 |
+
"LabHC/bias_in_bios_class_set1": {
|
| 185 |
+
"0": {
|
| 186 |
+
"tpp_threshold_2_total_metric": 0.2072499841451645,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.2639999985694885,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.056750014424324036,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.2617500126361847,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.3460000157356262,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.08425000309944153,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.2797500491142273,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3760000467300415,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.09624999761581421,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.20875000953674316,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.42000001668930054,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.21125000715255737,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.14775000512599945,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.44300001859664917,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.2952500134706497,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.1184999942779541,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.4480000138282776,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.3295000195503235,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.055750057101249695,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.4490000605583191,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.3932500034570694
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.12600000202655792,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.15799999237060547,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.031999990344047546,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.1975000500679016,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.2510000467300415,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.05349999666213989,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.21300001442432404,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.34200000762939453,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.1289999932050705,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.19225002825260162,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.3960000276565552,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.20374999940395355,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.17050005495548248,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4490000605583191,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.2785000056028366,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.11775000393390656,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4610000252723694,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.34325002133846283,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.06300005316734314,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4630000591278076,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.4000000059604645
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.06599998474121094,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.07499998807907104,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.009000003337860107,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.229249969124794,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.32999998331069946,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.10075001418590546,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.2177499681711197,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.38499999046325684,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.16725002229213715,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.18924999237060547,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.41200000047683716,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.2227500081062317,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.14800001680850983,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.44700002670288086,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.29900000989437103,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.11549995839595795,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.44999998807907104,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.3345000296831131,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.056249991059303284,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.4520000219345093,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.395750030875206
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.16249999403953552,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.16699999570846558,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.004500001668930054,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.39225006103515625,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.41700005531311035,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.0247499942779541,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.41325005888938904,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.48100006580352783,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.0677500069141388,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.3202500492334366,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.49500006437301636,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.17475001513957977,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.24375005066394806,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.49500006437301636,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.2512500137090683,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.24575002491474152,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.49400001764297485,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.24824999272823334,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.16350005567073822,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49500006437301636,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.33150000870227814
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.022999972105026245,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.029999971389770508,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.006999999284744263,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.23775002360343933,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.2770000100135803,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.03924998641014099,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.2999999672174454,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.39499998092651367,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.0950000137090683,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.2647499740123749,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.44099998474121094,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.17625001072883606,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.21550001204013824,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4750000238418579,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.25950001180171967,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.17224998772144318,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4750000238418579,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.30275003612041473,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.11650000512599945,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4750000238418579,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.35850001871585846
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": -0.0045000165700912476,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.001999974250793457,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.006499990820884705,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.00024999678134918213,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.009999990463256836,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.009749993681907654,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.009000018239021301,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.023000001907348633,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.013999983668327332,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.036250039935112,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.050000011920928955,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.013749971985816956,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.1352500021457672,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.15799999237060547,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.022749990224838257,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.2795000374317169,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.30400002002716064,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.024499982595443726,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.4072500318288803,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.453000009059906,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.045749977231025696
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.04825003445148468,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.06800001859664917,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.01974998414516449,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.12025003135204315,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.1420000195503235,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.021749988198280334,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.2045000195503235,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.23000001907348633,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.025499999523162842,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.2879999876022339,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.3199999928474426,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.03200000524520874,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.4022500216960907,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.44300001859664917,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.04074999690055847,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.3904999941587448,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.44999998807907104,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.05949999392032623,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.3707500398159027,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.45100003480911255,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.08024999499320984
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -1.4901161193847656e-08,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": 0.001999974250793457,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.001999989151954651,
|
| 352 |
+
"tpp_threshold_5_total_metric": -0.001749977469444275,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.009000003337860107,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.010749980807304382,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.007249996066093445,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.018999993801116943,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.011749997735023499,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.05824999511241913,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.07999998331069946,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.021749988198280334,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.20674999058246613,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.23799997568130493,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.031249985098838806,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.2902500033378601,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.40700000524520874,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.11675000190734863,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.2900000214576721,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.437000036239624,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.1470000147819519
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.035499975085258484,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.04499995708465576,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.009499981999397278,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.10374999046325684,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.12099999189376831,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.017250001430511475,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.1744999885559082,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.19099998474121094,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.016499996185302734,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.2122500240802765,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.24000000953674316,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.027749985456466675,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.302000030875206,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.3360000252723694,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.03399999439716339,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.36225004494190216,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.4050000309944153,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.04274998605251312,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.29500003159046173,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.43300002813339233,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.1379999965429306
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.019499987363815308,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.02499997615814209,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.005499988794326782,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.03874999284744263,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.06000000238418579,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.021250009536743164,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.07150004804134369,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.10000002384185791,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.02849997580051422,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.12074999511241913,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.15399998426437378,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.03324998915195465,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.2977500557899475,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.3400000333786011,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.042249977588653564,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.34550000727176666,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.39800000190734863,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.05249999463558197,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.3172500431537628,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.4140000343322754,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.09674999117851257
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|