Upload tpp/gemma-2-2b-res-matryoshka-dc_blocks.16.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-matryoshka-dc_blocks.16.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": 1745752774555,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.08475000113248826,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.09699999690055847,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.01224999576807022,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.15817500501871107,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.18429999947547915,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.02612499445676804,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.1898000121116638,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.25940001010894775,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.06959999799728393,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.21222500503063202,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.3135000050067902,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.10127499997615814,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.22897499948740005,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3743000030517578,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.14532500356435776,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.2468000203371048,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.41880002021789553,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.17199999988079073,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.23222503066062927,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.44910003542900084,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.21687500476837157
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.14404999911785127,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.16280000209808348,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.018750002980232237,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.24745001196861266,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.29060001373291017,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.04315000176429749,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.2685500204563141,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.38900002241134646,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.12045000195503235,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.2574500024318695,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.43420001268386843,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.1767500102519989,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.1989999979734421,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.45620001554489137,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.2572000175714493,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.1647000104188919,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.46540002822875975,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.3007000178098679,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.09330002963542938,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.4676000475883484,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.374300017952919
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.025450003147125245,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.031199991703033447,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.005749988555908203,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.0688999980688095,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.0779999852180481,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.009099987149238587,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.11105000376701354,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.12979999780654908,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.018749994039535523,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.16700000762939454,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.19279999732971193,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.025799989700317383,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.258950001001358,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.29239999055862426,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.03344998955726623,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.32890003025531767,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.37220001220703125,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.04329998195171356,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.37115003168582916,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.4306000232696533,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.059449991583824156
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.16.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.16.hook_resid_post",
|
| 163 |
+
"hook_layer": 16,
|
| 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": 0.24175004661083221,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.2800000309944153,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.03824998438358307,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.2762500196695328,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.3450000286102295,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.06875000894069672,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.30875004827976227,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.39100003242492676,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.08224998414516449,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.2612500339746475,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.4230000376701355,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.16175000369548798,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.18500003218650818,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.4410000443458557,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.25600001215934753,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.14850004017353058,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.44600003957748413,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.29749999940395355,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.07375006377696991,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.44800007343292236,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.37425000965595245
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.1287500560283661,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.15300005674362183,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.024250000715255737,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.18450000882148743,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.21700000762939453,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.032499998807907104,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.21275007724761963,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.3510000705718994,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.13824999332427979,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.2367500215768814,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.4140000343322754,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.17725001275539398,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.16175004839897156,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4500000476837158,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.28824999928474426,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.1275000274181366,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4590000510215759,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.33150002360343933,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.07400006055831909,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4620000720024109,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.3880000114440918
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.00824996829032898,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.02599996328353882,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.01774999499320984,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.15799996256828308,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.2149999737739563,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.05700001120567322,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.15824998915195465,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.3330000042915344,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.17475001513957977,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.16124999523162842,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.3880000114440918,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.22675001621246338,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.13124994933605194,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.4229999780654907,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.2917500287294388,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.12575000524520874,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.45100003480911255,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.3252500295639038,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.05700001120567322,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.45600003004074097,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.39900001883506775
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.3139999508857727,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.3219999670982361,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.008000016212463379,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.4385000169277191,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.4660000205039978,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.027500003576278687,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.39149998128414154,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.4869999885559082,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.09550000727176666,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.34074996411800385,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.4919999837875366,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.15125001966953278,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.2674999535083771,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.49299997091293335,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.22550001740455627,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.21799995005130768,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.49299997091293335,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.27500002086162567,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.14374999701976776,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49400001764297485,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.3502500206232071
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.0274999737739563,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.03299999237060547,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.00550001859664917,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.18000005185604095,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.21000003814697266,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.0299999862909317,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.27150000631809235,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.3830000162124634,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.11150000989437103,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.28724999725818634,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.45399999618530273,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.1667499989271164,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.24950000643730164,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4740000367164612,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.22450003027915955,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.20375002920627594,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.47800004482269287,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.27425001561641693,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.11800001561641693,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.47800004482269287,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.36000002920627594
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 1.4901161193847656e-08,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.0040000081062316895,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.003999993205070496,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.01000000536441803,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.013999998569488525,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.003999993205070496,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.03749997913837433,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.04799997806549072,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.010499998927116394,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.06575001776218414,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.09200000762939453,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.026249989867210388,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.18674996495246887,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.21599996089935303,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.029249995946884155,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.30900001525878906,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.3479999899864197,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.038999974727630615,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.4139999747276306,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4599999785423279,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.046000003814697266
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.07500001788139343,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.078000009059906,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.0029999911785125732,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.12599998712539673,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.1369999647140503,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.010999977588653564,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.1927500218153,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.22600001096725464,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.03324998915195465,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.25449998676776886,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.2979999780654907,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.04349999129772186,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.33674998581409454,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.38999998569488525,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.05324999988079071,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.37025003135204315,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.4300000071525574,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.05974997580051422,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.3592500388622284,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.44200003147125244,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.08274999260902405
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -0.010749995708465576,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": -0.0040000081062316895,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.006749987602233887,
|
| 352 |
+
"tpp_threshold_5_total_metric": 0.00899997353553772,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.0209999680519104,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.01199999451637268,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.03350000083446503,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.050999999046325684,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.017499998211860657,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.06175003945827484,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.08600002527236938,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.024249985814094543,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.15800002217292786,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.1850000023841858,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.026999980211257935,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.2522500157356262,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.29100000858306885,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.03874999284744263,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.3602500408887863,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.4320000410079956,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.07175000011920929
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.020749956369400024,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.02199995517730713,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.0012499988079071045,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.10750000178813934,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.11699998378753662,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.009499981999397278,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.1474999636411667,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.16499996185302734,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.017499998211860657,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.20599998533725739,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.218999981880188,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.012999996542930603,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.30275000631809235,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.3199999928474426,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.01724998652935028,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.37700001895427704,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.40700000524520874,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.0299999862909317,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.39375002682209015,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.42900002002716064,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.035249993205070496
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.042250022292137146,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.0559999942779541,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.013749971985816956,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.09200002253055573,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.10100001096725464,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.008999988436698914,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.14400005340576172,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.15900003910064697,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.014999985694885254,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.24700000882148743,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.26899999380111694,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.021999984979629517,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.31050002574920654,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.35100001096725464,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.040499985218048096,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.33600006997585297,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.3850000500679016,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.048999980092048645,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.32850007712841034,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.39000004529953003,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.06149996817111969
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|