Upload tpp/gemma-2-2b-res-matryoshka-dc_blocks.5.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-matryoshka-dc_blocks.5.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": 1745752280605,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.06562500298023223,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.10120000243186951,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.035574999451637265,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.09397500455379486,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.1577000081539154,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.06372500360012054,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.12410000264644622,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.22310000658035278,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.09900000393390655,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.13942500799894333,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.2701000154018402,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.13067500740289686,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.19275001138448716,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.36670001745224,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.17395000606775282,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.21587502360343935,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.41510003209114077,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.19922500848770142,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.18927503973245619,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.43440004587173464,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.2451250061392784
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.11545001864433288,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.18160001039505005,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.06614999175071716,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.1583000034093857,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.2802000045776367,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.12190000116825103,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.17880000472068786,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.3688000082969666,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.1900000035762787,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.157150000333786,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.4096000075340271,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.2524500072002411,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.12085000574588775,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.4490000128746033,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.3281500071287155,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.09360001385211944,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.46260002851486204,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.3690000146627426,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.05980003476142883,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.46620004177093505,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.40640000700950624
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.01579998731613159,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.020799994468688965,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.005000007152557373,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.02965000569820404,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.03520001173019409,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.005550006031990051,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.06940000057220459,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.07740000486373902,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.008000004291534423,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.12170001566410064,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.1306000232696533,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.008900007605552674,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.26465001702308655,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.2844000220298767,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.01975000500679016,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.33815003335475924,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.36760003566741944,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.029450002312660217,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.31875004470348356,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.4026000499725342,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.08385000526905059
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.5.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.5.hook_resid_post",
|
| 163 |
+
"hook_layer": 5,
|
| 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": 0.24050001800060272,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.3140000104904175,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.07349999248981476,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.20600000023841858,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.3569999933242798,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.1509999930858612,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.1757500171661377,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3960000276565552,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.22025001049041748,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.1289999783039093,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.421999990940094,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.2930000126361847,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.1082499772310257,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.44099998474121094,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.33275000751018524,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.0820000171661377,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.44600003957748413,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.36400002241134644,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.053000032901763916,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.45100003480911255,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.39800000190734863
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.014249995350837708,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.0690000057220459,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.05475001037120819,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.028249964118003845,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.09399998188018799,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.06575001776218414,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.1304999589920044,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.3539999723434448,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.22350001335144043,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.1199999749660492,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.4129999876022339,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.2930000126361847,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.09524999558925629,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.45399999618530273,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.35875000059604645,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.07099996507167816,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4599999785423279,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.3890000134706497,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.04175002872943878,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4610000252723694,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.4192499965429306
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.04950001835823059,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.10100001096725464,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.05149999260902405,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.15900002419948578,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.24900001287460327,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.08999998867511749,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.16850005090236664,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.284000039100647,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.11549998819828033,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.17400002479553223,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.3400000333786011,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.16600000858306885,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.1220000684261322,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.3940000534057617,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.2719999849796295,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.09250003099441528,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.4410000443458557,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.34850001335144043,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.039750054478645325,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.4530000686645508,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.41325001418590546
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.19125007092952728,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.27400004863739014,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.08274997770786285,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.2340000569820404,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.3800000548362732,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.1459999978542328,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.2795000374317169,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.44700002670288086,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.16749998927116394,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.2420000433921814,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.46900004148483276,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.22699999809265137,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.15125000476837158,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.484000027179718,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.33275002241134644,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.13150005042552948,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.49100005626678467,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.3595000058412552,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.09250004589557648,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49100005626678467,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.3985000103712082
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.08174999058246613,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.1499999761581421,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.06824998557567596,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.1642499715089798,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.32099997997283936,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.15675000846385956,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.13974995911121368,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.36299997568130493,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.22325001657009125,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.12074998021125793,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.4039999842643738,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.28325000405311584,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.12749998271465302,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.47200000286102295,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.34450002014636993,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.09100000560283661,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4750000238418579,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.3840000182390213,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.07200001180171967,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4750000238418579,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.40300001204013824
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.006499961018562317,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.011999964714050293,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.005500003695487976,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.012749984860420227,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.018000006675720215,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.005250021815299988,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.04499998688697815,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.0559999942779541,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.011000007390975952,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.1352500021457672,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.14300000667572021,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.007750004529953003,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.27049998939037323,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.28200000524520874,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.01150001585483551,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.3675000071525574,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.3920000195503235,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.024500012397766113,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.34825003147125244,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.437000036239624,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.08875000476837158
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.035500019788742065,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.04100000858306885,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.005499988794326782,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.05700002610683441,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.06800001859664917,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.010999992489814758,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.09274999797344208,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.09700000286102295,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.004250004887580872,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.15775004029273987,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.16700005531311035,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.009250015020370483,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.3160000443458557,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.33900004625320435,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.023000001907348633,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.3765000253915787,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.4150000214576721,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.038499996066093445,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.31000006198883057,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.42100006341934204,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.11100000143051147
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": 0.0222499817609787,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": 0.02799999713897705,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.005750015377998352,
|
| 352 |
+
"tpp_threshold_5_total_metric": 0.02850000560283661,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.03200000524520874,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.0034999996423721313,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.07700000703334808,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.08700001239776611,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.01000000536441803,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.09650002419948578,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.11000001430511475,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.013499990105628967,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.22750000655651093,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.2590000033378601,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.03149999678134918,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.3250000476837158,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.362000048160553,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.03700000047683716,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.3045000582933426,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.3980000615119934,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.09350000321865082
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.016250014305114746,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.021000027656555176,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.00475001335144043,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.029250040650367737,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.0350000262260437,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.005749985575675964,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.06299999356269836,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.06999999284744263,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.006999999284744263,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.09975001215934753,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.10900002717971802,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.009250015020370483,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.20725004374980927,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.22800004482269287,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.020750001072883606,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.29475004971027374,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.3200000524520874,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.02525000274181366,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.32975004613399506,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.4030000567436218,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.07325001060962677
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": -0.0015000402927398682,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.001999974250793457,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.003500014543533325,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.020749971270561218,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.023000001907348633,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.0022500306367874146,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.06925001740455627,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.07700002193450928,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.007750004529953003,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.11924999952316284,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.12400001287460327,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.00475001335144043,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.3020000010728836,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.3140000104904175,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.012000009417533875,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.32700003683567047,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.3490000367164612,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.02199999988079071,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.30125002562999725,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.3540000319480896,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.052750006318092346
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|