Upload tpp/gemma-2-2b-res-matryoshka-dc_blocks.11.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-matryoshka-dc_blocks.11.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": 1745752550815,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.06897500157356262,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.10889999866485596,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.03992499709129333,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.10377500206232071,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.17730000019073486,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.07352499812841416,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.10297501534223558,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.22150000929832458,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.11852499395608902,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.11180000901222228,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.26480000615119936,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.15299999713897705,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.16922501027584075,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3569000124931335,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.18767500221729277,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.2017250180244446,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.40750002264976504,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.20577500462532042,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.20382503569126129,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.4442000389099121,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.24037500321865082
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.13230000436306,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.20859999656677247,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.07629999220371246,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.18720001578330994,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.32880001068115233,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.14159999489784242,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.15880001187324524,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.38560000658035276,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.22679999470710754,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.12654999494552613,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.41719999313354494,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.2906499981880188,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.10765000581741332,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.4532000064849854,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.345550000667572,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.09105001389980316,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.4632000207901001,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.3721500068902969,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.06250003278255463,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.4678000330924988,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.4053000003099442
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.005649998784065247,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.009200000762939453,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.0035500019788742066,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.020349988341331483,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.025799989700317383,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.005450001358985901,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.047150018811225894,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.05740001201629639,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.010249993205070496,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.09705002307891845,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.11240001916885375,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.015349996089935303,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.2308000147342682,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.2606000185012817,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.02980000376701355,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.312400022149086,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.35180002450942993,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.03940000236034393,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.34515003859996796,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.42060004472732543,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.07545000612735749
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.11.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.11.hook_resid_post",
|
| 163 |
+
"hook_layer": 11,
|
| 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": 0.1730000078678131,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.28600001335144043,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.11300000548362732,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.19100002944469452,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.3460000157356262,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.1549999862909317,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.16324999928474426,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.3709999918937683,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.20774999260902405,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.1382499784231186,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.3999999761581421,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.2617499977350235,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.07575000822544098,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.42500001192092896,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.349250003695488,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.06350000202655792,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.43400001525878906,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.37050001323223114,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.038750022649765015,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.44700002670288086,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.40825000405311584
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.16699999570846558,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.2889999747276306,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.12199997901916504,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.19850002229213715,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.3320000171661377,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.13349999487400055,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.1652500033378601,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.3799999952316284,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.2147499918937683,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.1054999977350235,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.40799999237060547,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.30249999463558197,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.10750001668930054,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4480000138282776,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.34049999713897705,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.09125001728534698,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4650000333786011,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.3737500160932541,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.06025002896785736,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4650000333786011,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.4047500044107437
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.041500017046928406,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.0690000057220459,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.027499988675117493,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.16250000894069672,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.24900001287460327,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.08650000393390656,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.10500001907348633,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.35600000619888306,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.25099998712539673,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.07450000941753387,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.3790000081062317,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.3044999986886978,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.08275000751018524,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.43400001525878906,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.3512500077486038,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.08575001358985901,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.453000009059906,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.367249995470047,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.04625004529953003,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.45500004291534424,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.4087499976158142
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.2030000388622284,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.26200002431869507,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.058999985456466675,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.2605000138282776,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.3700000047683716,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.109499990940094,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.2580000162124634,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.4480000138282776,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.1899999976158142,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.20600000023841858,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.47200000286102295,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.26600000262260437,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.15175001323223114,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.48100000619888306,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.3292499929666519,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.11875000596046448,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.48500001430511475,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.36625000834465027,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.09775003790855408,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.4930000305175781,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.39524999260902405
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.0769999623298645,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.1369999647140503,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.06000000238418579,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.12350000441074371,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.34700000286102295,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.22349999845027924,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.10250002145767212,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.37300002574920654,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.2705000042915344,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.10849998891353607,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.4269999861717224,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.31849999725818634,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.12049998342990875,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4779999852180481,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.35750000178813934,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.09600003063678741,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4790000319480896,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.3830000013113022,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.06950002908706665,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4790000319480896,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.40950000286102295
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.0059999823570251465,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.009999990463256836,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.0040000081062316895,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.005250021815299988,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.00700002908706665,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.0017500072717666626,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.006750017404556274,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.013999998569488525,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.007249981164932251,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.035750046372413635,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.053000032901763916,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.01724998652935028,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.14899998903274536,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.16699999570846558,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.018000006675720215,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.29100000858306885,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.31800001859664917,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.027000010013580322,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.39250002801418304,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4500000476837158,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.057500019669532776
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.013249948620796204,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.012999951839447021,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": -0.00024999678134918213,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.010499954223632812,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.0209999680519104,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.010500013828277588,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.03674997389316559,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.04799997806549072,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.011250004172325134,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.08449998497962952,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.10199999809265137,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.01750001311302185,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.30024999380111694,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.34700000286102295,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.046750009059906006,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.3817500174045563,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.43300002813339233,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.05125001072883606,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.3387500196695328,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.43800002336502075,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.09925000369548798
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -0.009999945759773254,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": -0.003999948501586914,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.00599999725818634,
|
| 352 |
+
"tpp_threshold_5_total_metric": -0.0027499794960021973,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.0,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.0027499794960021973,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.00700002908706665,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.012000024318695068,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.004999995231628418,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.00850003957748413,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.017000019550323486,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.008499979972839355,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.09550006687641144,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.11600005626678467,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.02049998939037323,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.19275005161762238,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.22800004482269287,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.035249993205070496,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.36200007796287537,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.42500007152557373,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.06299999356269836
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": -0.0007499903440475464,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.0040000081062316895,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.004749998450279236,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.04774996638298035,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.05799996852874756,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.010250002145767212,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.09475003182888031,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.11400002241134644,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.019249990582466125,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.17000001668930054,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.18800002336502075,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.018000006675720215,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.28425000607967377,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.31300002336502075,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.028750017285346985,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.3697500228881836,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.4050000309944153,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.03525000810623169,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.34700001776218414,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.4140000343322754,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.06700001657009125
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.019749999046325684,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.023000001907348633,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.0032500028610229492,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.04099997878074646,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.042999982833862305,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.0020000040531158447,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.09050004184246063,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.09900003671646118,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.00849999487400055,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.18650002777576447,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.20200002193450928,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.015499994158744812,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.32500001788139343,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.36000001430511475,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.034999996423721313,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.3267500102519989,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.375,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.0482499897480011,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.28550004959106445,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.3760000467300415,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.09049999713897705
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|