Upload tpp/gemma-2-2b-res-matryoshka-dc_blocks.12.hook_resid_post_eval_results.json with huggingface_hub
Browse files
tpp/gemma-2-2b-res-matryoshka-dc_blocks.12.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": 1745752595297,
|
| 74 |
+
"eval_result_metrics": {
|
| 75 |
+
"tpp_metrics": {
|
| 76 |
+
"tpp_threshold_2_total_metric": 0.09577501267194748,
|
| 77 |
+
"tpp_threshold_2_intended_diff_only": 0.13370001316070557,
|
| 78 |
+
"tpp_threshold_2_unintended_diff_only": 0.03792500048875809,
|
| 79 |
+
"tpp_threshold_5_total_metric": 0.1066250130534172,
|
| 80 |
+
"tpp_threshold_5_intended_diff_only": 0.19700002074241638,
|
| 81 |
+
"tpp_threshold_5_unintended_diff_only": 0.09037500768899917,
|
| 82 |
+
"tpp_threshold_10_total_metric": 0.10720000118017196,
|
| 83 |
+
"tpp_threshold_10_intended_diff_only": 0.23860000371932985,
|
| 84 |
+
"tpp_threshold_10_unintended_diff_only": 0.13140000253915787,
|
| 85 |
+
"tpp_threshold_20_total_metric": 0.11999999731779099,
|
| 86 |
+
"tpp_threshold_20_intended_diff_only": 0.2837000072002411,
|
| 87 |
+
"tpp_threshold_20_unintended_diff_only": 0.1637000098824501,
|
| 88 |
+
"tpp_threshold_50_total_metric": 0.17365001738071442,
|
| 89 |
+
"tpp_threshold_50_intended_diff_only": 0.3700000286102295,
|
| 90 |
+
"tpp_threshold_50_unintended_diff_only": 0.1963500112295151,
|
| 91 |
+
"tpp_threshold_100_total_metric": 0.20192500501871108,
|
| 92 |
+
"tpp_threshold_100_intended_diff_only": 0.4146000146865845,
|
| 93 |
+
"tpp_threshold_100_unintended_diff_only": 0.21267500966787337,
|
| 94 |
+
"tpp_threshold_500_total_metric": 0.19970003366470337,
|
| 95 |
+
"tpp_threshold_500_intended_diff_only": 0.4451000452041626,
|
| 96 |
+
"tpp_threshold_500_unintended_diff_only": 0.24540001153945923
|
| 97 |
+
}
|
| 98 |
+
},
|
| 99 |
+
"eval_result_details": [
|
| 100 |
+
{
|
| 101 |
+
"dataset_name": "LabHC/bias_in_bios_class_set1_tpp_results",
|
| 102 |
+
"tpp_threshold_2_total_metric": 0.18645002841949462,
|
| 103 |
+
"tpp_threshold_2_intended_diff_only": 0.25760003328323366,
|
| 104 |
+
"tpp_threshold_2_unintended_diff_only": 0.07115000486373901,
|
| 105 |
+
"tpp_threshold_5_total_metric": 0.18865001499652861,
|
| 106 |
+
"tpp_threshold_5_intended_diff_only": 0.360200035572052,
|
| 107 |
+
"tpp_threshold_5_unintended_diff_only": 0.17155002057552338,
|
| 108 |
+
"tpp_threshold_10_total_metric": 0.14434999525547026,
|
| 109 |
+
"tpp_threshold_10_intended_diff_only": 0.39360001087188723,
|
| 110 |
+
"tpp_threshold_10_unintended_diff_only": 0.24925001561641694,
|
| 111 |
+
"tpp_threshold_20_total_metric": 0.11660000085830688,
|
| 112 |
+
"tpp_threshold_20_intended_diff_only": 0.42280001640319825,
|
| 113 |
+
"tpp_threshold_20_unintended_diff_only": 0.30620001554489135,
|
| 114 |
+
"tpp_threshold_50_total_metric": 0.09555000364780426,
|
| 115 |
+
"tpp_threshold_50_intended_diff_only": 0.45420002937316895,
|
| 116 |
+
"tpp_threshold_50_unintended_diff_only": 0.3586500257253647,
|
| 117 |
+
"tpp_threshold_100_total_metric": 0.0835000067949295,
|
| 118 |
+
"tpp_threshold_100_intended_diff_only": 0.4646000266075134,
|
| 119 |
+
"tpp_threshold_100_unintended_diff_only": 0.3811000198125839,
|
| 120 |
+
"tpp_threshold_500_total_metric": 0.06215002834796905,
|
| 121 |
+
"tpp_threshold_500_intended_diff_only": 0.46920005083084104,
|
| 122 |
+
"tpp_threshold_500_unintended_diff_only": 0.40705002248287203
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"dataset_name": "canrager/amazon_reviews_mcauley_1and5_tpp_results",
|
| 126 |
+
"tpp_threshold_2_total_metric": 0.00509999692440033,
|
| 127 |
+
"tpp_threshold_2_intended_diff_only": 0.00979999303817749,
|
| 128 |
+
"tpp_threshold_2_unintended_diff_only": 0.004699996113777161,
|
| 129 |
+
"tpp_threshold_5_total_metric": 0.024600011110305787,
|
| 130 |
+
"tpp_threshold_5_intended_diff_only": 0.03380000591278076,
|
| 131 |
+
"tpp_threshold_5_unintended_diff_only": 0.009199994802474975,
|
| 132 |
+
"tpp_threshold_10_total_metric": 0.07005000710487366,
|
| 133 |
+
"tpp_threshold_10_intended_diff_only": 0.08359999656677246,
|
| 134 |
+
"tpp_threshold_10_unintended_diff_only": 0.013549989461898804,
|
| 135 |
+
"tpp_threshold_20_total_metric": 0.12339999377727509,
|
| 136 |
+
"tpp_threshold_20_intended_diff_only": 0.14459999799728393,
|
| 137 |
+
"tpp_threshold_20_unintended_diff_only": 0.02120000422000885,
|
| 138 |
+
"tpp_threshold_50_total_metric": 0.2517500311136246,
|
| 139 |
+
"tpp_threshold_50_intended_diff_only": 0.28580002784729003,
|
| 140 |
+
"tpp_threshold_50_unintended_diff_only": 0.03404999673366547,
|
| 141 |
+
"tpp_threshold_100_total_metric": 0.3203500032424927,
|
| 142 |
+
"tpp_threshold_100_intended_diff_only": 0.3646000027656555,
|
| 143 |
+
"tpp_threshold_100_unintended_diff_only": 0.044249999523162845,
|
| 144 |
+
"tpp_threshold_500_total_metric": 0.3372500389814377,
|
| 145 |
+
"tpp_threshold_500_intended_diff_only": 0.4210000395774841,
|
| 146 |
+
"tpp_threshold_500_unintended_diff_only": 0.08375000059604645
|
| 147 |
+
}
|
| 148 |
+
],
|
| 149 |
+
"sae_bench_commit_hash": "Unknown",
|
| 150 |
+
"sae_lens_id": "blocks.12.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.12.hook_resid_post",
|
| 163 |
+
"hook_layer": 12,
|
| 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": "gemma-2-2b/12-res-matryoshka-dc",
|
| 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": 0.22425004839897156,
|
| 187 |
+
"tpp_threshold_2_intended_diff_only": 0.2640000581741333,
|
| 188 |
+
"tpp_threshold_2_unintended_diff_only": 0.03975000977516174,
|
| 189 |
+
"tpp_threshold_5_total_metric": 0.20450004935264587,
|
| 190 |
+
"tpp_threshold_5_intended_diff_only": 0.35200005769729614,
|
| 191 |
+
"tpp_threshold_5_unintended_diff_only": 0.14750000834465027,
|
| 192 |
+
"tpp_threshold_10_total_metric": 0.1757500022649765,
|
| 193 |
+
"tpp_threshold_10_intended_diff_only": 0.37400001287460327,
|
| 194 |
+
"tpp_threshold_10_unintended_diff_only": 0.19825001060962677,
|
| 195 |
+
"tpp_threshold_20_total_metric": 0.11675001680850983,
|
| 196 |
+
"tpp_threshold_20_intended_diff_only": 0.4150000214576721,
|
| 197 |
+
"tpp_threshold_20_unintended_diff_only": 0.2982500046491623,
|
| 198 |
+
"tpp_threshold_50_total_metric": 0.08200004696846008,
|
| 199 |
+
"tpp_threshold_50_intended_diff_only": 0.4440000653266907,
|
| 200 |
+
"tpp_threshold_50_unintended_diff_only": 0.3620000183582306,
|
| 201 |
+
"tpp_threshold_100_total_metric": 0.07000002264976501,
|
| 202 |
+
"tpp_threshold_100_intended_diff_only": 0.4500000476837158,
|
| 203 |
+
"tpp_threshold_100_unintended_diff_only": 0.3800000250339508,
|
| 204 |
+
"tpp_threshold_500_total_metric": 0.048500046133995056,
|
| 205 |
+
"tpp_threshold_500_intended_diff_only": 0.4580000638961792,
|
| 206 |
+
"tpp_threshold_500_unintended_diff_only": 0.40950001776218414
|
| 207 |
+
},
|
| 208 |
+
"1": {
|
| 209 |
+
"tpp_threshold_2_total_metric": 0.14475004374980927,
|
| 210 |
+
"tpp_threshold_2_intended_diff_only": 0.23200005292892456,
|
| 211 |
+
"tpp_threshold_2_unintended_diff_only": 0.0872500091791153,
|
| 212 |
+
"tpp_threshold_5_total_metric": 0.15125001966953278,
|
| 213 |
+
"tpp_threshold_5_intended_diff_only": 0.3020000457763672,
|
| 214 |
+
"tpp_threshold_5_unintended_diff_only": 0.1507500261068344,
|
| 215 |
+
"tpp_threshold_10_total_metric": 0.11549998819828033,
|
| 216 |
+
"tpp_threshold_10_intended_diff_only": 0.34700000286102295,
|
| 217 |
+
"tpp_threshold_10_unintended_diff_only": 0.23150001466274261,
|
| 218 |
+
"tpp_threshold_20_total_metric": 0.07825002074241638,
|
| 219 |
+
"tpp_threshold_20_intended_diff_only": 0.38600003719329834,
|
| 220 |
+
"tpp_threshold_20_unintended_diff_only": 0.30775001645088196,
|
| 221 |
+
"tpp_threshold_50_total_metric": 0.07075002789497375,
|
| 222 |
+
"tpp_threshold_50_intended_diff_only": 0.4310000538825989,
|
| 223 |
+
"tpp_threshold_50_unintended_diff_only": 0.3602500259876251,
|
| 224 |
+
"tpp_threshold_100_total_metric": 0.07225000858306885,
|
| 225 |
+
"tpp_threshold_100_intended_diff_only": 0.4570000171661377,
|
| 226 |
+
"tpp_threshold_100_unintended_diff_only": 0.38475000858306885,
|
| 227 |
+
"tpp_threshold_500_total_metric": 0.05825003981590271,
|
| 228 |
+
"tpp_threshold_500_intended_diff_only": 0.4630000591278076,
|
| 229 |
+
"tpp_threshold_500_unintended_diff_only": 0.4047500193119049
|
| 230 |
+
},
|
| 231 |
+
"2": {
|
| 232 |
+
"tpp_threshold_2_total_metric": 0.132999986410141,
|
| 233 |
+
"tpp_threshold_2_intended_diff_only": 0.24199998378753662,
|
| 234 |
+
"tpp_threshold_2_unintended_diff_only": 0.10899999737739563,
|
| 235 |
+
"tpp_threshold_5_total_metric": 0.1315000057220459,
|
| 236 |
+
"tpp_threshold_5_intended_diff_only": 0.32200002670288086,
|
| 237 |
+
"tpp_threshold_5_unintended_diff_only": 0.19050002098083496,
|
| 238 |
+
"tpp_threshold_10_total_metric": 0.06574997305870056,
|
| 239 |
+
"tpp_threshold_10_intended_diff_only": 0.3619999885559082,
|
| 240 |
+
"tpp_threshold_10_unintended_diff_only": 0.29625001549720764,
|
| 241 |
+
"tpp_threshold_20_total_metric": 0.0627499669790268,
|
| 242 |
+
"tpp_threshold_20_intended_diff_only": 0.38499999046325684,
|
| 243 |
+
"tpp_threshold_20_unintended_diff_only": 0.32225002348423004,
|
| 244 |
+
"tpp_threshold_50_total_metric": 0.08000001311302185,
|
| 245 |
+
"tpp_threshold_50_intended_diff_only": 0.437000036239624,
|
| 246 |
+
"tpp_threshold_50_unintended_diff_only": 0.3570000231266022,
|
| 247 |
+
"tpp_threshold_100_total_metric": 0.07574999332427979,
|
| 248 |
+
"tpp_threshold_100_intended_diff_only": 0.4520000219345093,
|
| 249 |
+
"tpp_threshold_100_unintended_diff_only": 0.3762500286102295,
|
| 250 |
+
"tpp_threshold_500_total_metric": 0.04125000536441803,
|
| 251 |
+
"tpp_threshold_500_intended_diff_only": 0.45500004291534424,
|
| 252 |
+
"tpp_threshold_500_unintended_diff_only": 0.4137500375509262
|
| 253 |
+
},
|
| 254 |
+
"6": {
|
| 255 |
+
"tpp_threshold_2_total_metric": 0.2780000567436218,
|
| 256 |
+
"tpp_threshold_2_intended_diff_only": 0.3250000476837158,
|
| 257 |
+
"tpp_threshold_2_unintended_diff_only": 0.046999990940093994,
|
| 258 |
+
"tpp_threshold_5_total_metric": 0.33150002360343933,
|
| 259 |
+
"tpp_threshold_5_intended_diff_only": 0.44600003957748413,
|
| 260 |
+
"tpp_threshold_5_unintended_diff_only": 0.1145000159740448,
|
| 261 |
+
"tpp_threshold_10_total_metric": 0.26775000989437103,
|
| 262 |
+
"tpp_threshold_10_intended_diff_only": 0.4710000157356262,
|
| 263 |
+
"tpp_threshold_10_unintended_diff_only": 0.2032500058412552,
|
| 264 |
+
"tpp_threshold_20_total_metric": 0.22550003230571747,
|
| 265 |
+
"tpp_threshold_20_intended_diff_only": 0.47800004482269287,
|
| 266 |
+
"tpp_threshold_20_unintended_diff_only": 0.2525000125169754,
|
| 267 |
+
"tpp_threshold_50_total_metric": 0.13999997079372406,
|
| 268 |
+
"tpp_threshold_50_intended_diff_only": 0.48100000619888306,
|
| 269 |
+
"tpp_threshold_50_unintended_diff_only": 0.341000035405159,
|
| 270 |
+
"tpp_threshold_100_total_metric": 0.11675000190734863,
|
| 271 |
+
"tpp_threshold_100_intended_diff_only": 0.48500001430511475,
|
| 272 |
+
"tpp_threshold_100_unintended_diff_only": 0.3682500123977661,
|
| 273 |
+
"tpp_threshold_500_total_metric": 0.09075003862380981,
|
| 274 |
+
"tpp_threshold_500_intended_diff_only": 0.49100005626678467,
|
| 275 |
+
"tpp_threshold_500_unintended_diff_only": 0.40025001764297485
|
| 276 |
+
},
|
| 277 |
+
"9": {
|
| 278 |
+
"tpp_threshold_2_total_metric": 0.1522500067949295,
|
| 279 |
+
"tpp_threshold_2_intended_diff_only": 0.2250000238418579,
|
| 280 |
+
"tpp_threshold_2_unintended_diff_only": 0.0727500170469284,
|
| 281 |
+
"tpp_threshold_5_total_metric": 0.12449997663497925,
|
| 282 |
+
"tpp_threshold_5_intended_diff_only": 0.3790000081062317,
|
| 283 |
+
"tpp_threshold_5_unintended_diff_only": 0.25450003147125244,
|
| 284 |
+
"tpp_threshold_10_total_metric": 0.09700000286102295,
|
| 285 |
+
"tpp_threshold_10_intended_diff_only": 0.4140000343322754,
|
| 286 |
+
"tpp_threshold_10_unintended_diff_only": 0.31700003147125244,
|
| 287 |
+
"tpp_threshold_20_total_metric": 0.09974996745586395,
|
| 288 |
+
"tpp_threshold_20_intended_diff_only": 0.44999998807907104,
|
| 289 |
+
"tpp_threshold_20_unintended_diff_only": 0.3502500206232071,
|
| 290 |
+
"tpp_threshold_50_total_metric": 0.10499995946884155,
|
| 291 |
+
"tpp_threshold_50_intended_diff_only": 0.4779999852180481,
|
| 292 |
+
"tpp_threshold_50_unintended_diff_only": 0.37300002574920654,
|
| 293 |
+
"tpp_threshold_100_total_metric": 0.08275000751018524,
|
| 294 |
+
"tpp_threshold_100_intended_diff_only": 0.4790000319480896,
|
| 295 |
+
"tpp_threshold_100_unintended_diff_only": 0.39625002443790436,
|
| 296 |
+
"tpp_threshold_500_total_metric": 0.07200001180171967,
|
| 297 |
+
"tpp_threshold_500_intended_diff_only": 0.4790000319480896,
|
| 298 |
+
"tpp_threshold_500_unintended_diff_only": 0.40700002014636993
|
| 299 |
+
}
|
| 300 |
+
},
|
| 301 |
+
"canrager/amazon_reviews_mcauley_1and5": {
|
| 302 |
+
"1": {
|
| 303 |
+
"tpp_threshold_2_total_metric": 0.010500043630599976,
|
| 304 |
+
"tpp_threshold_2_intended_diff_only": 0.01500004529953003,
|
| 305 |
+
"tpp_threshold_2_unintended_diff_only": 0.004500001668930054,
|
| 306 |
+
"tpp_threshold_5_total_metric": 0.00600007176399231,
|
| 307 |
+
"tpp_threshold_5_intended_diff_only": 0.010000050067901611,
|
| 308 |
+
"tpp_threshold_5_unintended_diff_only": 0.003999978303909302,
|
| 309 |
+
"tpp_threshold_10_total_metric": 0.01400001347064972,
|
| 310 |
+
"tpp_threshold_10_intended_diff_only": 0.023000001907348633,
|
| 311 |
+
"tpp_threshold_10_unintended_diff_only": 0.008999988436698914,
|
| 312 |
+
"tpp_threshold_20_total_metric": 0.05600005388259888,
|
| 313 |
+
"tpp_threshold_20_intended_diff_only": 0.0700000524520874,
|
| 314 |
+
"tpp_threshold_20_unintended_diff_only": 0.013999998569488525,
|
| 315 |
+
"tpp_threshold_50_total_metric": 0.20850007236003876,
|
| 316 |
+
"tpp_threshold_50_intended_diff_only": 0.23600006103515625,
|
| 317 |
+
"tpp_threshold_50_unintended_diff_only": 0.027499988675117493,
|
| 318 |
+
"tpp_threshold_100_total_metric": 0.3085000365972519,
|
| 319 |
+
"tpp_threshold_100_intended_diff_only": 0.3450000286102295,
|
| 320 |
+
"tpp_threshold_100_unintended_diff_only": 0.0364999920129776,
|
| 321 |
+
"tpp_threshold_500_total_metric": 0.3837500810623169,
|
| 322 |
+
"tpp_threshold_500_intended_diff_only": 0.4530000686645508,
|
| 323 |
+
"tpp_threshold_500_unintended_diff_only": 0.06924998760223389
|
| 324 |
+
},
|
| 325 |
+
"2": {
|
| 326 |
+
"tpp_threshold_2_total_metric": 0.004999935626983643,
|
| 327 |
+
"tpp_threshold_2_intended_diff_only": 0.008999943733215332,
|
| 328 |
+
"tpp_threshold_2_unintended_diff_only": 0.0040000081062316895,
|
| 329 |
+
"tpp_threshold_5_total_metric": 0.0037500113248825073,
|
| 330 |
+
"tpp_threshold_5_intended_diff_only": 0.018999993801116943,
|
| 331 |
+
"tpp_threshold_5_unintended_diff_only": 0.015249982476234436,
|
| 332 |
+
"tpp_threshold_10_total_metric": 0.05750000476837158,
|
| 333 |
+
"tpp_threshold_10_intended_diff_only": 0.06999999284744263,
|
| 334 |
+
"tpp_threshold_10_unintended_diff_only": 0.012499988079071045,
|
| 335 |
+
"tpp_threshold_20_total_metric": 0.10599994659423828,
|
| 336 |
+
"tpp_threshold_20_intended_diff_only": 0.11899995803833008,
|
| 337 |
+
"tpp_threshold_20_unintended_diff_only": 0.013000011444091797,
|
| 338 |
+
"tpp_threshold_50_total_metric": 0.275749996304512,
|
| 339 |
+
"tpp_threshold_50_intended_diff_only": 0.3100000023841858,
|
| 340 |
+
"tpp_threshold_50_unintended_diff_only": 0.03425000607967377,
|
| 341 |
+
"tpp_threshold_100_total_metric": 0.35874998569488525,
|
| 342 |
+
"tpp_threshold_100_intended_diff_only": 0.3999999761581421,
|
| 343 |
+
"tpp_threshold_100_unintended_diff_only": 0.041249990463256836,
|
| 344 |
+
"tpp_threshold_500_total_metric": 0.341499999165535,
|
| 345 |
+
"tpp_threshold_500_intended_diff_only": 0.43400001525878906,
|
| 346 |
+
"tpp_threshold_500_unintended_diff_only": 0.09250001609325409
|
| 347 |
+
},
|
| 348 |
+
"3": {
|
| 349 |
+
"tpp_threshold_2_total_metric": -0.009750008583068848,
|
| 350 |
+
"tpp_threshold_2_intended_diff_only": -0.003000020980834961,
|
| 351 |
+
"tpp_threshold_2_unintended_diff_only": 0.006749987602233887,
|
| 352 |
+
"tpp_threshold_5_total_metric": 0.007749959826469421,
|
| 353 |
+
"tpp_threshold_5_intended_diff_only": 0.015999972820281982,
|
| 354 |
+
"tpp_threshold_5_unintended_diff_only": 0.008250012993812561,
|
| 355 |
+
"tpp_threshold_10_total_metric": 0.03950001299381256,
|
| 356 |
+
"tpp_threshold_10_intended_diff_only": 0.050999999046325684,
|
| 357 |
+
"tpp_threshold_10_unintended_diff_only": 0.011499986052513123,
|
| 358 |
+
"tpp_threshold_20_total_metric": 0.072999969124794,
|
| 359 |
+
"tpp_threshold_20_intended_diff_only": 0.09499996900558472,
|
| 360 |
+
"tpp_threshold_20_unintended_diff_only": 0.02199999988079071,
|
| 361 |
+
"tpp_threshold_50_total_metric": 0.1652500331401825,
|
| 362 |
+
"tpp_threshold_50_intended_diff_only": 0.19700002670288086,
|
| 363 |
+
"tpp_threshold_50_unintended_diff_only": 0.031749993562698364,
|
| 364 |
+
"tpp_threshold_100_total_metric": 0.2590000182390213,
|
| 365 |
+
"tpp_threshold_100_intended_diff_only": 0.30800002813339233,
|
| 366 |
+
"tpp_threshold_100_unintended_diff_only": 0.04900000989437103,
|
| 367 |
+
"tpp_threshold_500_total_metric": 0.33275003731250763,
|
| 368 |
+
"tpp_threshold_500_intended_diff_only": 0.4280000329017639,
|
| 369 |
+
"tpp_threshold_500_unintended_diff_only": 0.09524999558925629
|
| 370 |
+
},
|
| 371 |
+
"5": {
|
| 372 |
+
"tpp_threshold_2_total_metric": 0.004250004887580872,
|
| 373 |
+
"tpp_threshold_2_intended_diff_only": 0.009000003337860107,
|
| 374 |
+
"tpp_threshold_2_unintended_diff_only": 0.004749998450279236,
|
| 375 |
+
"tpp_threshold_5_total_metric": 0.045000046491622925,
|
| 376 |
+
"tpp_threshold_5_intended_diff_only": 0.05600005388259888,
|
| 377 |
+
"tpp_threshold_5_unintended_diff_only": 0.011000007390975952,
|
| 378 |
+
"tpp_threshold_10_total_metric": 0.10975003242492676,
|
| 379 |
+
"tpp_threshold_10_intended_diff_only": 0.12800002098083496,
|
| 380 |
+
"tpp_threshold_10_unintended_diff_only": 0.018249988555908203,
|
| 381 |
+
"tpp_threshold_20_total_metric": 0.15524999797344208,
|
| 382 |
+
"tpp_threshold_20_intended_diff_only": 0.18000000715255737,
|
| 383 |
+
"tpp_threshold_20_unintended_diff_only": 0.024750009179115295,
|
| 384 |
+
"tpp_threshold_50_total_metric": 0.30125007033348083,
|
| 385 |
+
"tpp_threshold_50_intended_diff_only": 0.3380000591278076,
|
| 386 |
+
"tpp_threshold_50_unintended_diff_only": 0.03674998879432678,
|
| 387 |
+
"tpp_threshold_100_total_metric": 0.3515000194311142,
|
| 388 |
+
"tpp_threshold_100_intended_diff_only": 0.3970000147819519,
|
| 389 |
+
"tpp_threshold_100_unintended_diff_only": 0.04549999535083771,
|
| 390 |
+
"tpp_threshold_500_total_metric": 0.32325008511543274,
|
| 391 |
+
"tpp_threshold_500_intended_diff_only": 0.4160000681877136,
|
| 392 |
+
"tpp_threshold_500_unintended_diff_only": 0.09274998307228088
|
| 393 |
+
},
|
| 394 |
+
"6": {
|
| 395 |
+
"tpp_threshold_2_total_metric": 0.015500009059906006,
|
| 396 |
+
"tpp_threshold_2_intended_diff_only": 0.018999993801116943,
|
| 397 |
+
"tpp_threshold_2_unintended_diff_only": 0.0034999847412109375,
|
| 398 |
+
"tpp_threshold_5_total_metric": 0.06049996614456177,
|
| 399 |
+
"tpp_threshold_5_intended_diff_only": 0.0679999589920044,
|
| 400 |
+
"tpp_threshold_5_unintended_diff_only": 0.007499992847442627,
|
| 401 |
+
"tpp_threshold_10_total_metric": 0.12949997186660767,
|
| 402 |
+
"tpp_threshold_10_intended_diff_only": 0.1459999680519104,
|
| 403 |
+
"tpp_threshold_10_unintended_diff_only": 0.016499996185302734,
|
| 404 |
+
"tpp_threshold_20_total_metric": 0.22675000131130219,
|
| 405 |
+
"tpp_threshold_20_intended_diff_only": 0.2590000033378601,
|
| 406 |
+
"tpp_threshold_20_unintended_diff_only": 0.03225000202655792,
|
| 407 |
+
"tpp_threshold_50_total_metric": 0.30799998342990875,
|
| 408 |
+
"tpp_threshold_50_intended_diff_only": 0.3479999899864197,
|
| 409 |
+
"tpp_threshold_50_unintended_diff_only": 0.040000006556510925,
|
| 410 |
+
"tpp_threshold_100_total_metric": 0.32399995625019073,
|
| 411 |
+
"tpp_threshold_100_intended_diff_only": 0.37299996614456177,
|
| 412 |
+
"tpp_threshold_100_unintended_diff_only": 0.04900000989437103,
|
| 413 |
+
"tpp_threshold_500_total_metric": 0.3049999922513962,
|
| 414 |
+
"tpp_threshold_500_intended_diff_only": 0.37400001287460327,
|
| 415 |
+
"tpp_threshold_500_unintended_diff_only": 0.06900002062320709
|
| 416 |
+
}
|
| 417 |
+
}
|
| 418 |
+
}
|
| 419 |
+
}
|