Upload tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.7.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-snap-matryoshka-dc_blocks.7.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": 1745617723551,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.038850000500679015,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.049000000953674315,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.0101500004529953,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.12040000706911087,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.14330000281333924,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.022899995744228366,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.2019500106573105,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.2459000051021576,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.04394999444484711,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.2322000116109848,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.3064000129699707,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.0742000013589859,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.2570000171661377,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3782000124454498,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.12119999527931213,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.26517502069473264,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.4317000210285187,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.166525000333786,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.22257502973079682,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.4473000347614288,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.22472500503063203
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.05064999759197235,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.06540000438690186,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.014750006794929504,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.17250001132488252,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.20980000495910645,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.03729999363422394,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.29605001509189605,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3730000138282776,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.07694999873638153,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.29380002021789553,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.42300002574920653,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.12920000553131103,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.24560001492500305,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.4588000178337097,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.21320000290870667,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.1796500265598297,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.46660003662109373,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.286950010061264,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.09320002794265747,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.4698000431060791,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.3766000151634216
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.02705000340938568,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.032599997520446775,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.005549994111061096,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.06830000281333923,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.07680000066757202,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.008499997854232787,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.10785000622272492,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.1187999963760376,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.010949990153312683,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.1706000030040741,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.18980000019073487,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.019199997186660767,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.2684000194072723,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.2976000070571899,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.029199987649917603,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.35070001482963564,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.3968000054359436,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.04609999060630798,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.35195003151893617,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.42480002641677855,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.07284999489784241
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.7.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.7.hook_resid_post",
|
| 163 |
+
"hook_layer": 7,
|
| 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": 0.1797499656677246,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.23199999332427979,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.052250027656555176,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.21899999678134918,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.3059999942779541,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.08699999749660492,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.23475001752376556,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3410000205039978,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.10625000298023224,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.20399999618530273,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.4020000100135803,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.1980000138282776,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.16349998116493225,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.4399999976158142,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.27650001645088196,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.125249981880188,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.44999998807907104,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.32475000619888306,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.0637499988079071,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.4520000219345093,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.3882500231266022
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.018250003457069397,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.02799999713897705,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.009749993681907654,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.03999999165534973,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.05199998617172241,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.01199999451637268,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.23900000751018524,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.3190000057220459,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.07999999821186066,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.28049997985363007,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.3709999918937683,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.09050001204013824,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.19350004196166992,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4230000376701355,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.22949999570846558,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.1510000377893448,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4500000476837158,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.29900000989437103,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.0805000364780426,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.46400004625320435,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.38350000977516174
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.05224999785423279,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.06000000238418579,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.007750004529953003,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.22624994814395905,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.2669999599456787,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.040750011801719666,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.28224995732307434,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.36399996280670166,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.08175000548362732,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.3044999837875366,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.41200000047683716,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.10750001668930054,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.2827499657869339,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.4559999704360962,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.1732500046491623,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.19200000166893005,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.4570000171661377,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.26500001549720764,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.060749977827072144,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.4570000171661377,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.39625003933906555
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.002750009298324585,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.0040000081062316895,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.0012499988079071045,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.17850005626678467,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.19200003147125244,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.013499975204467773,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.43375004827976227,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.45100003480911255,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.01724998652935028,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.40250006318092346,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.48100006580352783,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.07850000262260437,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.34975002706050873,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.49400001764297485,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.14424999058246613,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.24100004136562347,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.49500006437301636,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.2540000230073929,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.15100005269050598,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49500006437301636,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.3440000116825104
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.000250011682510376,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.003000020980834961,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.002750009298324585,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.1987500637769699,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.23200005292892456,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.03324998915195465,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.29050004482269287,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.39000004529953003,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.09950000047683716,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.27750007808208466,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.4490000605583191,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.17149998247623444,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.23850005865097046,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.48100006580352783,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.24250000715255737,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.18900007009506226,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.48100006580352783,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.2919999957084656,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.11000007390975952,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.48100006580352783,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.3709999918937683
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.005000025033950806,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.008000016212463379,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.0029999911785125732,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.00925000011920929,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.013999998569488525,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.004749998450279236,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.025999993085861206,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.03299999237060547,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.006999999284744263,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.10500004887580872,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.12700003385543823,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.021999984979629517,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.25850003957748413,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.2900000214576721,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.03149998188018799,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.3562500327825546,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.3930000066757202,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.03674997389316559,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.38600000739097595,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4490000009536743,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.06299999356269836
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.07375000417232513,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.08300000429153442,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.00925000011920929,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.21250000596046448,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.23100000619888306,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.01850000023841858,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.29649998247623444,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.31699997186660767,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.02049998939037323,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.35749997198581696,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.38599997758865356,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.02850000560283661,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.3854999989271164,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.44099998474121094,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.05549998581409454,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.3775000125169754,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.44300001859664917,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.06550000607967377,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.34925001859664917,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.44300001859664917,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.09375
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": 0.004249989986419678,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": 0.004999995231628418,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.0007500052452087402,
|
| 352 |
+
"tpp_threshold_5_total_metric": 0.004999995231628418,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.009999990463256836,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.004999995231628418,
|
| 355 |
+
"tpp_threshold_10_total_metric": -0.0027499794960021973,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.0040000081062316895,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.006749987602233887,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.014249980449676514,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.02399998903274536,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.009750008583068848,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.06974999606609344,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.08399999141693115,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.014249995350837708,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.28700001537799835,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.3610000014305115,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.07399998605251312,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.2877500355243683,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.4190000295639038,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.13124999403953552
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.0059999823570251465,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.015999972820281982,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.009999990463256836,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.04324997961521149,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.05199998617172241,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.008750006556510925,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.10624997317790985,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.11399996280670166,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.007749989628791809,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.16899999976158142,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.1850000023841858,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.01600000262260437,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.3045000284910202,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.3230000138282776,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.018499985337257385,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.39149999618530273,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.4129999876022339,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.021499991416931152,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.39900003373622894,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.43300002813339233,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.03399999439716339
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.04625001549720764,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.050999999046325684,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.004749983549118042,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.0715000331401825,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.07700002193450928,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.005499988794326782,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.11325006186962128,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.1260000467300415,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.012749984860420227,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.20725001394748688,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.22699999809265137,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.01974998414516449,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.3237500339746475,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.3500000238418579,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.026249989867210388,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.341250017285347,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.37400001287460327,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.03274999558925629,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.33775006234645844,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.3800000548362732,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.04224999248981476
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|