Upload tpp/gemma-2-2b-res-matryoshka-dc_blocks.8.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-matryoshka-dc_blocks.8.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": 1745752417333,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.0504749983549118,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.06300000548362732,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.012525007128715515,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.12190000414848327,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.15250000953674317,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.03060000538825989,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.1714500069618225,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.23470001220703124,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.06325000524520874,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.20885000675916673,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.3016000151634216,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.09275000840425492,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.254925012588501,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3901000201702118,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.1351750075817108,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.2638250172138214,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.43110002875328063,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.16727501153945923,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.22835002839565277,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.44750004410743716,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.21915001571178438
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.07439999580383301,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.09720000028610229,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.022800004482269286,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.1940000146627426,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.253600013256073,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.059599998593330386,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.2573000192642212,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3788000226020813,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.12150000333786011,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.25490000247955324,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.42560001611709597,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.17070001363754272,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.21130000054836273,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.4582000136375427,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.24690001308918,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.1640000194311142,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.4686000347137451,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.3046000152826309,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.0841000258922577,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.4704000473022461,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.3863000214099884
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.0265500009059906,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.028800010681152344,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.0022500097751617433,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.04979999363422394,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.05140000581741333,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.001600012183189392,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.08559999465942383,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.0906000018119812,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.005000007152557373,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.16280001103878022,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.1776000142097473,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.014800003170967102,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.2985500246286392,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.32200002670288086,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.023450002074241638,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.3636500149965286,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.39360002279281614,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.029950007796287537,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.37260003089904786,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.42460004091262815,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.052000010013580324
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.8.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.8.hook_resid_post",
|
| 163 |
+
"hook_layer": 8,
|
| 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": 0.08900001645088196,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.10500001907348633,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.01600000262260437,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.17475003004074097,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.22100001573562622,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.046249985694885254,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.2537500262260437,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3370000123977661,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.08324998617172241,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.27125002443790436,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.38600003719329834,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.11475001275539398,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.2055000215768814,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.43300002813339233,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.22750000655651093,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.1510000228881836,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.45100003480911255,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.30000001192092896,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.07825006544589996,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.4530000686645508,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.3747500032186508
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.11750002205371857,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.1420000195503235,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.02449999749660492,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.1522500365972519,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.18700003623962402,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.03474999964237213,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.2157500684261322,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.3190000653266907,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.10324999690055847,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.23125004768371582,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.362000048160553,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.13075000047683716,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.20025001466274261,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4240000247955322,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.2237500101327896,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.15775004029273987,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.45500004291534424,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.29725000262260437,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.06100006401538849,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4620000720024109,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.4010000079870224
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.028749987483024597,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.041999995708465576,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.013250008225440979,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.15524999797344208,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.18900001049041748,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.0337500125169754,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.21850000321865082,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.3410000205039978,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.12250001728534698,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.20524998009204865,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.4169999957084656,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.21175001561641693,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.17899997532367706,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.45899999141693115,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.2800000160932541,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.13975003361701965,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.46000003814697266,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.320250004529953,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.05700001120567322,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.46000003814697266,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.40300002694129944
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.030749991536140442,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.041999995708465576,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.011250004172325134,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.2992500364780426,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.38600003719329834,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.08675000071525574,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.35850000381469727,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.48000001907348633,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.12150001525878906,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.31200000643730164,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.48900002241134644,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.1770000159740448,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.2685000002384186,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.4950000047683716,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.226500004529953,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.2120000123977661,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.4970000386238098,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.2850000262260437,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.1315000057220459,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.4970000386238098,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.3655000329017639
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.10599996149539948,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.1549999713897705,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.04900000989437103,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.18849997222423553,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.2849999666213989,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.09649999439716339,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.23999999463558197,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.4169999957084656,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.1770000010728836,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.25474995374679565,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.4739999771118164,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.21925002336502075,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.203249990940094,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.48000001907348633,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.27675002813339233,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.15949998795986176,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.48000001907348633,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.3205000311136246,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.09274998307228088,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.48000001907348633,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.38725003600120544
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.0065000057220458984,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.009000003337860107,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.002499997615814209,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.018000036478042603,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.021000027656555176,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.0029999911785125732,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.0350000262260437,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.04000002145767212,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.004999995231628418,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.09375004470348358,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.1170000433921814,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.023249998688697815,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.2795000523328781,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.3160000443458557,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.0364999920129776,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.364250048995018,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.40800005197525024,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.04375000298023224,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.39225007593631744,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.44800007343292236,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.05574999749660492
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.023249968886375427,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.02399998903274536,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.0007500201463699341,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.05349995195865631,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.06199997663497925,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.008500024676322937,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.09099997580051422,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.0989999771118164,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.008000001311302185,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.2732499837875366,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.3009999990463257,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.027750015258789062,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.38475003838539124,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.4230000376701355,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.03824999928474426,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.39525000751018524,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.4390000104904175,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.04375000298023224,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.37150003015995026,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.4410000443458557,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.06950001418590546
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -0.006499975919723511,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": -0.001999974250793457,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.004500001668930054,
|
| 352 |
+
"tpp_threshold_5_total_metric": 0.0017500072717666626,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": -0.001999974250793457,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": -0.0037499815225601196,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.02425001561641693,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.022000014781951904,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": -0.002250000834465027,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.04499998688697815,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.0559999942779541,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.011000007390975952,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.1677500456571579,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.1820000410079956,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.014249995350837708,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.2902500331401825,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.3160000443458557,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.025750011205673218,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.3682500123977661,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.42000001668930054,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.051750004291534424
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.07649999856948853,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.07700002193450928,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.000500023365020752,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.1250000149011612,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.12800002098083496,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.003000006079673767,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.179500013589859,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.18800002336502075,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.008500009775161743,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.24300001561641693,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.24900001287460327,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.00599999725818634,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.3557499796152115,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.3709999918937683,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.015250012278556824,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.4100000113248825,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.42900002002716064,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.019000008702278137,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.38575002551078796,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.4320000410079956,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.04625001549720764
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.03300000727176666,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.03600001335144043,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.003000006079673767,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.05074995756149292,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.04799997806549072,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": -0.0027499794960021973,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.09824994206428528,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.10399997234344482,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.005750030279159546,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.15900002419948578,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.16500002145767212,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.00599999725818634,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.3050000071525574,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.31800001859664917,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.013000011444091797,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.3584999740123749,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.37599998712539673,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.01750001311302185,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.3452500104904175,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.38200002908706665,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.03675001859664917
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|