Upload results.json
Browse files- results.json +588 -0
results.json
ADDED
|
@@ -0,0 +1,588 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"timestamp": "2024-12-17T07:14:59.243305",
|
| 4 |
+
"vidore_benchmark_version": ""
|
| 5 |
+
},
|
| 6 |
+
"metrics": {
|
| 7 |
+
"syntheticDocQA_artificial_intelligence_test": {
|
| 8 |
+
"ndcg_at_1": 0.96,
|
| 9 |
+
"ndcg_at_3": 0.98524,
|
| 10 |
+
"ndcg_at_5": 0.98524,
|
| 11 |
+
"ndcg_at_10": 0.98524,
|
| 12 |
+
"ndcg_at_20": 0.98524,
|
| 13 |
+
"ndcg_at_50": 0.98524,
|
| 14 |
+
"ndcg_at_100": 0.98524,
|
| 15 |
+
"map_at_1": 0.96,
|
| 16 |
+
"map_at_3": 0.98,
|
| 17 |
+
"map_at_5": 0.98,
|
| 18 |
+
"map_at_10": 0.98,
|
| 19 |
+
"map_at_20": 0.98,
|
| 20 |
+
"map_at_50": 0.98,
|
| 21 |
+
"map_at_100": 0.98,
|
| 22 |
+
"recall_at_1": 0.96,
|
| 23 |
+
"recall_at_3": 1.0,
|
| 24 |
+
"recall_at_5": 1.0,
|
| 25 |
+
"recall_at_10": 1.0,
|
| 26 |
+
"recall_at_20": 1.0,
|
| 27 |
+
"recall_at_50": 1.0,
|
| 28 |
+
"recall_at_100": 1.0,
|
| 29 |
+
"precision_at_1": 0.96,
|
| 30 |
+
"precision_at_3": 0.33333,
|
| 31 |
+
"precision_at_5": 0.2,
|
| 32 |
+
"precision_at_10": 0.1,
|
| 33 |
+
"precision_at_20": 0.05,
|
| 34 |
+
"precision_at_50": 0.02,
|
| 35 |
+
"precision_at_100": 0.01,
|
| 36 |
+
"mrr_at_1": 0.96,
|
| 37 |
+
"mrr_at_3": 0.98,
|
| 38 |
+
"mrr_at_5": 0.98,
|
| 39 |
+
"mrr_at_10": 0.98,
|
| 40 |
+
"mrr_at_20": 0.98,
|
| 41 |
+
"mrr_at_50": 0.98,
|
| 42 |
+
"mrr_at_100": 0.98,
|
| 43 |
+
"naucs_at_1_max": 0.06302521008403594,
|
| 44 |
+
"naucs_at_1_std": -0.04843604108309577,
|
| 45 |
+
"naucs_at_1_diff1": 0.9305555555555578,
|
| 46 |
+
"naucs_at_3_max": 1.0,
|
| 47 |
+
"naucs_at_3_std": 1.0,
|
| 48 |
+
"naucs_at_3_diff1": 1.0,
|
| 49 |
+
"naucs_at_5_max": 1.0,
|
| 50 |
+
"naucs_at_5_std": 1.0,
|
| 51 |
+
"naucs_at_5_diff1": 1.0,
|
| 52 |
+
"naucs_at_10_max": 1.0,
|
| 53 |
+
"naucs_at_10_std": 1.0,
|
| 54 |
+
"naucs_at_10_diff1": 1.0,
|
| 55 |
+
"naucs_at_20_max": 1.0,
|
| 56 |
+
"naucs_at_20_std": 1.0,
|
| 57 |
+
"naucs_at_20_diff1": 1.0,
|
| 58 |
+
"naucs_at_50_max": null,
|
| 59 |
+
"naucs_at_50_std": null,
|
| 60 |
+
"naucs_at_50_diff1": null,
|
| 61 |
+
"naucs_at_100_max": null,
|
| 62 |
+
"naucs_at_100_std": null,
|
| 63 |
+
"naucs_at_100_diff1": null
|
| 64 |
+
},
|
| 65 |
+
"tabfquad_test_subsampled": {
|
| 66 |
+
"ndcg_at_1": 0.76429,
|
| 67 |
+
"ndcg_at_3": 0.81548,
|
| 68 |
+
"ndcg_at_5": 0.83054,
|
| 69 |
+
"ndcg_at_10": 0.84162,
|
| 70 |
+
"ndcg_at_20": 0.85069,
|
| 71 |
+
"ndcg_at_50": 0.85663,
|
| 72 |
+
"ndcg_at_100": 0.85846,
|
| 73 |
+
"map_at_1": 0.76429,
|
| 74 |
+
"map_at_3": 0.80238,
|
| 75 |
+
"map_at_5": 0.81095,
|
| 76 |
+
"map_at_10": 0.81526,
|
| 77 |
+
"map_at_20": 0.81777,
|
| 78 |
+
"map_at_50": 0.81884,
|
| 79 |
+
"map_at_100": 0.81903,
|
| 80 |
+
"recall_at_1": 0.76429,
|
| 81 |
+
"recall_at_3": 0.85357,
|
| 82 |
+
"recall_at_5": 0.88929,
|
| 83 |
+
"recall_at_10": 0.925,
|
| 84 |
+
"recall_at_20": 0.96071,
|
| 85 |
+
"recall_at_50": 0.98929,
|
| 86 |
+
"recall_at_100": 1.0,
|
| 87 |
+
"precision_at_1": 0.76429,
|
| 88 |
+
"precision_at_3": 0.28452,
|
| 89 |
+
"precision_at_5": 0.17786,
|
| 90 |
+
"precision_at_10": 0.0925,
|
| 91 |
+
"precision_at_20": 0.04804,
|
| 92 |
+
"precision_at_50": 0.01979,
|
| 93 |
+
"precision_at_100": 0.01,
|
| 94 |
+
"mrr_at_1": 0.7642857142857142,
|
| 95 |
+
"mrr_at_3": 0.8047619047619048,
|
| 96 |
+
"mrr_at_5": 0.812261904761905,
|
| 97 |
+
"mrr_at_10": 0.8157568027210885,
|
| 98 |
+
"mrr_at_20": 0.8182429277072134,
|
| 99 |
+
"mrr_at_50": 0.8196521812519686,
|
| 100 |
+
"mrr_at_100": 0.819844389714301,
|
| 101 |
+
"naucs_at_1_max": 0.3930648524086476,
|
| 102 |
+
"naucs_at_1_std": 0.13014375068583456,
|
| 103 |
+
"naucs_at_1_diff1": 0.8538351805113578,
|
| 104 |
+
"naucs_at_3_max": 0.3984224656102249,
|
| 105 |
+
"naucs_at_3_std": 0.1907296096552706,
|
| 106 |
+
"naucs_at_3_diff1": 0.7447435318714221,
|
| 107 |
+
"naucs_at_5_max": 0.3885787498856053,
|
| 108 |
+
"naucs_at_5_std": 0.16155699948140706,
|
| 109 |
+
"naucs_at_5_diff1": 0.795170983191481,
|
| 110 |
+
"naucs_at_10_max": 0.3904450669156529,
|
| 111 |
+
"naucs_at_10_std": 0.15479525143390607,
|
| 112 |
+
"naucs_at_10_diff1": 0.7791561068871982,
|
| 113 |
+
"naucs_at_20_max": 0.28736949325184535,
|
| 114 |
+
"naucs_at_20_std": -0.07932263814616634,
|
| 115 |
+
"naucs_at_20_diff1": 0.7797725150666338,
|
| 116 |
+
"naucs_at_50_max": 0.34267040149392364,
|
| 117 |
+
"naucs_at_50_std": -0.1545284780578993,
|
| 118 |
+
"naucs_at_50_diff1": 0.8078120136943696,
|
| 119 |
+
"naucs_at_100_max": 1.0,
|
| 120 |
+
"naucs_at_100_std": 1.0,
|
| 121 |
+
"naucs_at_100_diff1": 1.0
|
| 122 |
+
},
|
| 123 |
+
"shiftproject_test": {
|
| 124 |
+
"ndcg_at_1": 0.53,
|
| 125 |
+
"ndcg_at_3": 0.6444,
|
| 126 |
+
"ndcg_at_5": 0.68185,
|
| 127 |
+
"ndcg_at_10": 0.70901,
|
| 128 |
+
"ndcg_at_20": 0.71713,
|
| 129 |
+
"ndcg_at_50": 0.72325,
|
| 130 |
+
"ndcg_at_100": 0.7248,
|
| 131 |
+
"map_at_1": 0.53,
|
| 132 |
+
"map_at_3": 0.615,
|
| 133 |
+
"map_at_5": 0.636,
|
| 134 |
+
"map_at_10": 0.64798,
|
| 135 |
+
"map_at_20": 0.6505,
|
| 136 |
+
"map_at_50": 0.65154,
|
| 137 |
+
"map_at_100": 0.65165,
|
| 138 |
+
"recall_at_1": 0.53,
|
| 139 |
+
"recall_at_3": 0.73,
|
| 140 |
+
"recall_at_5": 0.82,
|
| 141 |
+
"recall_at_10": 0.9,
|
| 142 |
+
"recall_at_20": 0.93,
|
| 143 |
+
"recall_at_50": 0.96,
|
| 144 |
+
"recall_at_100": 0.97,
|
| 145 |
+
"precision_at_1": 0.53,
|
| 146 |
+
"precision_at_3": 0.24333,
|
| 147 |
+
"precision_at_5": 0.164,
|
| 148 |
+
"precision_at_10": 0.09,
|
| 149 |
+
"precision_at_20": 0.0465,
|
| 150 |
+
"precision_at_50": 0.0192,
|
| 151 |
+
"precision_at_100": 0.0097,
|
| 152 |
+
"mrr_at_1": 0.55,
|
| 153 |
+
"mrr_at_3": 0.6300000000000001,
|
| 154 |
+
"mrr_at_5": 0.655,
|
| 155 |
+
"mrr_at_10": 0.6632341269841271,
|
| 156 |
+
"mrr_at_20": 0.6648575036075036,
|
| 157 |
+
"mrr_at_50": 0.6659570406445406,
|
| 158 |
+
"mrr_at_100": 0.6660746877033642,
|
| 159 |
+
"naucs_at_1_max": -0.058021416101939505,
|
| 160 |
+
"naucs_at_1_std": -0.29796973154270173,
|
| 161 |
+
"naucs_at_1_diff1": 0.632737219755598,
|
| 162 |
+
"naucs_at_3_max": 0.0032239674457922728,
|
| 163 |
+
"naucs_at_3_std": -0.39310778666771473,
|
| 164 |
+
"naucs_at_3_diff1": 0.5492736244077923,
|
| 165 |
+
"naucs_at_5_max": 0.18798876528252045,
|
| 166 |
+
"naucs_at_5_std": -0.11273267981054919,
|
| 167 |
+
"naucs_at_5_diff1": 0.6309890957153872,
|
| 168 |
+
"naucs_at_10_max": 0.11223155929038518,
|
| 169 |
+
"naucs_at_10_std": -0.22661064425770017,
|
| 170 |
+
"naucs_at_10_diff1": 0.565966386554623,
|
| 171 |
+
"naucs_at_20_max": -0.13685474189675803,
|
| 172 |
+
"naucs_at_20_std": -0.5446178471388509,
|
| 173 |
+
"naucs_at_20_diff1": 0.6702681072429001,
|
| 174 |
+
"naucs_at_50_max": -0.27065826330532117,
|
| 175 |
+
"naucs_at_50_std": -0.5157563025210059,
|
| 176 |
+
"naucs_at_50_diff1": 0.6458916900093365,
|
| 177 |
+
"naucs_at_100_max": -0.30376595082477015,
|
| 178 |
+
"naucs_at_100_std": -0.6305633364456832,
|
| 179 |
+
"naucs_at_100_diff1": 0.5714285714285773
|
| 180 |
+
},
|
| 181 |
+
"syntheticDocQA_government_reports_test": {
|
| 182 |
+
"ndcg_at_1": 0.9,
|
| 183 |
+
"ndcg_at_3": 0.93524,
|
| 184 |
+
"ndcg_at_5": 0.95203,
|
| 185 |
+
"ndcg_at_10": 0.95203,
|
| 186 |
+
"ndcg_at_20": 0.95203,
|
| 187 |
+
"ndcg_at_50": 0.95203,
|
| 188 |
+
"ndcg_at_100": 0.95203,
|
| 189 |
+
"map_at_1": 0.9,
|
| 190 |
+
"map_at_3": 0.92667,
|
| 191 |
+
"map_at_5": 0.93617,
|
| 192 |
+
"map_at_10": 0.93617,
|
| 193 |
+
"map_at_20": 0.93617,
|
| 194 |
+
"map_at_50": 0.93617,
|
| 195 |
+
"map_at_100": 0.93617,
|
| 196 |
+
"recall_at_1": 0.9,
|
| 197 |
+
"recall_at_3": 0.96,
|
| 198 |
+
"recall_at_5": 1.0,
|
| 199 |
+
"recall_at_10": 1.0,
|
| 200 |
+
"recall_at_20": 1.0,
|
| 201 |
+
"recall_at_50": 1.0,
|
| 202 |
+
"recall_at_100": 1.0,
|
| 203 |
+
"precision_at_1": 0.9,
|
| 204 |
+
"precision_at_3": 0.32,
|
| 205 |
+
"precision_at_5": 0.2,
|
| 206 |
+
"precision_at_10": 0.1,
|
| 207 |
+
"precision_at_20": 0.05,
|
| 208 |
+
"precision_at_50": 0.02,
|
| 209 |
+
"precision_at_100": 0.01,
|
| 210 |
+
"mrr_at_1": 0.9,
|
| 211 |
+
"mrr_at_3": 0.9333333333333332,
|
| 212 |
+
"mrr_at_5": 0.9403333333333334,
|
| 213 |
+
"mrr_at_10": 0.9403333333333334,
|
| 214 |
+
"mrr_at_20": 0.9403333333333334,
|
| 215 |
+
"mrr_at_50": 0.9403333333333334,
|
| 216 |
+
"mrr_at_100": 0.9403333333333334,
|
| 217 |
+
"naucs_at_1_max": 0.1526143790849667,
|
| 218 |
+
"naucs_at_1_std": -0.28823529411764676,
|
| 219 |
+
"naucs_at_1_diff1": 0.8867880485527548,
|
| 220 |
+
"naucs_at_3_max": 0.07119514472455671,
|
| 221 |
+
"naucs_at_3_std": -0.29236694677871167,
|
| 222 |
+
"naucs_at_3_diff1": 0.8190943043884199,
|
| 223 |
+
"naucs_at_5_max": 1.0,
|
| 224 |
+
"naucs_at_5_std": 1.0,
|
| 225 |
+
"naucs_at_5_diff1": 1.0,
|
| 226 |
+
"naucs_at_10_max": 1.0,
|
| 227 |
+
"naucs_at_10_std": 1.0,
|
| 228 |
+
"naucs_at_10_diff1": 1.0,
|
| 229 |
+
"naucs_at_20_max": 1.0,
|
| 230 |
+
"naucs_at_20_std": 1.0,
|
| 231 |
+
"naucs_at_20_diff1": 1.0,
|
| 232 |
+
"naucs_at_50_max": null,
|
| 233 |
+
"naucs_at_50_std": null,
|
| 234 |
+
"naucs_at_50_diff1": null,
|
| 235 |
+
"naucs_at_100_max": null,
|
| 236 |
+
"naucs_at_100_std": null,
|
| 237 |
+
"naucs_at_100_diff1": null
|
| 238 |
+
},
|
| 239 |
+
"infovqa_test_subsampled": {
|
| 240 |
+
"ndcg_at_1": 0.76316,
|
| 241 |
+
"ndcg_at_3": 0.82921,
|
| 242 |
+
"ndcg_at_5": 0.84289,
|
| 243 |
+
"ndcg_at_10": 0.85188,
|
| 244 |
+
"ndcg_at_20": 0.85644,
|
| 245 |
+
"ndcg_at_50": 0.86119,
|
| 246 |
+
"ndcg_at_100": 0.86283,
|
| 247 |
+
"map_at_1": 0.76316,
|
| 248 |
+
"map_at_3": 0.8141,
|
| 249 |
+
"map_at_5": 0.8219,
|
| 250 |
+
"map_at_10": 0.8255,
|
| 251 |
+
"map_at_20": 0.82673,
|
| 252 |
+
"map_at_50": 0.82746,
|
| 253 |
+
"map_at_100": 0.8276,
|
| 254 |
+
"recall_at_1": 0.76316,
|
| 255 |
+
"recall_at_3": 0.87247,
|
| 256 |
+
"recall_at_5": 0.90486,
|
| 257 |
+
"recall_at_10": 0.9332,
|
| 258 |
+
"recall_at_20": 0.95142,
|
| 259 |
+
"recall_at_50": 0.97571,
|
| 260 |
+
"recall_at_100": 0.98583,
|
| 261 |
+
"precision_at_1": 0.76316,
|
| 262 |
+
"precision_at_3": 0.29082,
|
| 263 |
+
"precision_at_5": 0.18097,
|
| 264 |
+
"precision_at_10": 0.09332,
|
| 265 |
+
"precision_at_20": 0.04757,
|
| 266 |
+
"precision_at_50": 0.01951,
|
| 267 |
+
"precision_at_100": 0.00986,
|
| 268 |
+
"mrr_at_1": 0.7611336032388664,
|
| 269 |
+
"mrr_at_3": 0.813090418353576,
|
| 270 |
+
"mrr_at_5": 0.8216936572199727,
|
| 271 |
+
"mrr_at_10": 0.8248160465265728,
|
| 272 |
+
"mrr_at_20": 0.825886692856507,
|
| 273 |
+
"mrr_at_50": 0.8265042447979247,
|
| 274 |
+
"mrr_at_100": 0.8266668975545796,
|
| 275 |
+
"naucs_at_1_max": 0.5439433197090425,
|
| 276 |
+
"naucs_at_1_std": 0.24776250817754733,
|
| 277 |
+
"naucs_at_1_diff1": 0.8926114371628262,
|
| 278 |
+
"naucs_at_3_max": 0.5680546695594344,
|
| 279 |
+
"naucs_at_3_std": 0.2541748591933084,
|
| 280 |
+
"naucs_at_3_diff1": 0.853534819493221,
|
| 281 |
+
"naucs_at_5_max": 0.705693682065501,
|
| 282 |
+
"naucs_at_5_std": 0.4124626625681076,
|
| 283 |
+
"naucs_at_5_diff1": 0.8406271757500486,
|
| 284 |
+
"naucs_at_10_max": 0.7825666550711661,
|
| 285 |
+
"naucs_at_10_std": 0.5614849682029542,
|
| 286 |
+
"naucs_at_10_diff1": 0.842869767889761,
|
| 287 |
+
"naucs_at_20_max": 0.7543170269410486,
|
| 288 |
+
"naucs_at_20_std": 0.5443151217010491,
|
| 289 |
+
"naucs_at_20_diff1": 0.8462829819306663,
|
| 290 |
+
"naucs_at_50_max": 0.8236318160505198,
|
| 291 |
+
"naucs_at_50_std": 0.7478265745755217,
|
| 292 |
+
"naucs_at_50_diff1": 0.8852960979962946,
|
| 293 |
+
"naucs_at_100_max": 0.7986413595283582,
|
| 294 |
+
"naucs_at_100_std": 0.6710512069467076,
|
| 295 |
+
"naucs_at_100_diff1": 0.8616984811111583
|
| 296 |
+
},
|
| 297 |
+
"syntheticDocQA_healthcare_industry_test": {
|
| 298 |
+
"ndcg_at_1": 0.92,
|
| 299 |
+
"ndcg_at_3": 0.95024,
|
| 300 |
+
"ndcg_at_5": 0.95885,
|
| 301 |
+
"ndcg_at_10": 0.95885,
|
| 302 |
+
"ndcg_at_20": 0.95885,
|
| 303 |
+
"ndcg_at_50": 0.96077,
|
| 304 |
+
"ndcg_at_100": 0.96077,
|
| 305 |
+
"map_at_1": 0.92,
|
| 306 |
+
"map_at_3": 0.94333,
|
| 307 |
+
"map_at_5": 0.94833,
|
| 308 |
+
"map_at_10": 0.94833,
|
| 309 |
+
"map_at_20": 0.94833,
|
| 310 |
+
"map_at_50": 0.94861,
|
| 311 |
+
"map_at_100": 0.94861,
|
| 312 |
+
"recall_at_1": 0.92,
|
| 313 |
+
"recall_at_3": 0.97,
|
| 314 |
+
"recall_at_5": 0.99,
|
| 315 |
+
"recall_at_10": 0.99,
|
| 316 |
+
"recall_at_20": 0.99,
|
| 317 |
+
"recall_at_50": 1.0,
|
| 318 |
+
"recall_at_100": 1.0,
|
| 319 |
+
"precision_at_1": 0.92,
|
| 320 |
+
"precision_at_3": 0.32333,
|
| 321 |
+
"precision_at_5": 0.198,
|
| 322 |
+
"precision_at_10": 0.099,
|
| 323 |
+
"precision_at_20": 0.0495,
|
| 324 |
+
"precision_at_50": 0.02,
|
| 325 |
+
"precision_at_100": 0.01,
|
| 326 |
+
"mrr_at_1": 0.95,
|
| 327 |
+
"mrr_at_3": 0.9633333333333333,
|
| 328 |
+
"mrr_at_5": 0.9658333333333333,
|
| 329 |
+
"mrr_at_10": 0.9658333333333333,
|
| 330 |
+
"mrr_at_20": 0.9658333333333333,
|
| 331 |
+
"mrr_at_50": 0.96625,
|
| 332 |
+
"mrr_at_100": 0.96625,
|
| 333 |
+
"naucs_at_1_max": 0.6046918767507006,
|
| 334 |
+
"naucs_at_1_std": 0.07207049486461263,
|
| 335 |
+
"naucs_at_1_diff1": 0.9279295051353874,
|
| 336 |
+
"naucs_at_3_max": 0.460939931528168,
|
| 337 |
+
"naucs_at_3_std": 0.08839091192032704,
|
| 338 |
+
"naucs_at_3_diff1": 0.8513849984438244,
|
| 339 |
+
"naucs_at_5_max": 1.0,
|
| 340 |
+
"naucs_at_5_std": 1.0,
|
| 341 |
+
"naucs_at_5_diff1": 0.5541549953314738,
|
| 342 |
+
"naucs_at_10_max": 1.0,
|
| 343 |
+
"naucs_at_10_std": 1.0,
|
| 344 |
+
"naucs_at_10_diff1": 0.5541549953314738,
|
| 345 |
+
"naucs_at_20_max": 1.0,
|
| 346 |
+
"naucs_at_20_std": 1.0,
|
| 347 |
+
"naucs_at_20_diff1": 0.5541549953314738,
|
| 348 |
+
"naucs_at_50_max": null,
|
| 349 |
+
"naucs_at_50_std": null,
|
| 350 |
+
"naucs_at_50_diff1": null,
|
| 351 |
+
"naucs_at_100_max": null,
|
| 352 |
+
"naucs_at_100_std": null,
|
| 353 |
+
"naucs_at_100_diff1": null
|
| 354 |
+
},
|
| 355 |
+
"syntheticDocQA_energy_test": {
|
| 356 |
+
"ndcg_at_1": 0.91,
|
| 357 |
+
"ndcg_at_3": 0.94786,
|
| 358 |
+
"ndcg_at_5": 0.94786,
|
| 359 |
+
"ndcg_at_10": 0.95142,
|
| 360 |
+
"ndcg_at_20": 0.95142,
|
| 361 |
+
"ndcg_at_50": 0.9536,
|
| 362 |
+
"ndcg_at_100": 0.95525,
|
| 363 |
+
"map_at_1": 0.91,
|
| 364 |
+
"map_at_3": 0.94,
|
| 365 |
+
"map_at_5": 0.94,
|
| 366 |
+
"map_at_10": 0.94167,
|
| 367 |
+
"map_at_20": 0.94167,
|
| 368 |
+
"map_at_50": 0.9421,
|
| 369 |
+
"map_at_100": 0.94226,
|
| 370 |
+
"recall_at_1": 0.91,
|
| 371 |
+
"recall_at_3": 0.97,
|
| 372 |
+
"recall_at_5": 0.97,
|
| 373 |
+
"recall_at_10": 0.98,
|
| 374 |
+
"recall_at_20": 0.98,
|
| 375 |
+
"recall_at_50": 0.99,
|
| 376 |
+
"recall_at_100": 1.0,
|
| 377 |
+
"precision_at_1": 0.91,
|
| 378 |
+
"precision_at_3": 0.32333,
|
| 379 |
+
"precision_at_5": 0.194,
|
| 380 |
+
"precision_at_10": 0.098,
|
| 381 |
+
"precision_at_20": 0.049,
|
| 382 |
+
"precision_at_50": 0.0198,
|
| 383 |
+
"precision_at_100": 0.01,
|
| 384 |
+
"mrr_at_1": 0.95,
|
| 385 |
+
"mrr_at_3": 0.96,
|
| 386 |
+
"mrr_at_5": 0.9625,
|
| 387 |
+
"mrr_at_10": 0.9625,
|
| 388 |
+
"mrr_at_20": 0.9631666666666666,
|
| 389 |
+
"mrr_at_50": 0.9631666666666666,
|
| 390 |
+
"mrr_at_100": 0.9633279569892474,
|
| 391 |
+
"naucs_at_1_max": 0.6794791990870418,
|
| 392 |
+
"naucs_at_1_std": -0.20370370370370447,
|
| 393 |
+
"naucs_at_1_diff1": 0.9854756717501807,
|
| 394 |
+
"naucs_at_3_max": 0.807812013694371,
|
| 395 |
+
"naucs_at_3_std": -0.45238095238094883,
|
| 396 |
+
"naucs_at_3_diff1": 0.9564270152505466,
|
| 397 |
+
"naucs_at_5_max": 0.807812013694364,
|
| 398 |
+
"naucs_at_5_std": -0.45238095238095277,
|
| 399 |
+
"naucs_at_5_diff1": 0.9564270152505424,
|
| 400 |
+
"naucs_at_10_max": 0.9346405228758136,
|
| 401 |
+
"naucs_at_10_std": 0.1914098972922579,
|
| 402 |
+
"naucs_at_10_diff1": 0.9346405228758136,
|
| 403 |
+
"naucs_at_20_max": 0.9346405228758136,
|
| 404 |
+
"naucs_at_20_std": 0.1914098972922579,
|
| 405 |
+
"naucs_at_20_diff1": 0.9346405228758136,
|
| 406 |
+
"naucs_at_50_max": 1.0,
|
| 407 |
+
"naucs_at_50_std": 0.554154995331464,
|
| 408 |
+
"naucs_at_50_diff1": 1.0,
|
| 409 |
+
"naucs_at_100_max": null,
|
| 410 |
+
"naucs_at_100_std": null,
|
| 411 |
+
"naucs_at_100_diff1": null
|
| 412 |
+
},
|
| 413 |
+
"arxivqa_test_subsampled": {
|
| 414 |
+
"ndcg_at_1": 0.822,
|
| 415 |
+
"ndcg_at_3": 0.86124,
|
| 416 |
+
"ndcg_at_5": 0.87217,
|
| 417 |
+
"ndcg_at_10": 0.87958,
|
| 418 |
+
"ndcg_at_20": 0.88759,
|
| 419 |
+
"ndcg_at_50": 0.8913,
|
| 420 |
+
"ndcg_at_100": 0.89258,
|
| 421 |
+
"map_at_1": 0.822,
|
| 422 |
+
"map_at_3": 0.85133,
|
| 423 |
+
"map_at_5": 0.85753,
|
| 424 |
+
"map_at_10": 0.86076,
|
| 425 |
+
"map_at_20": 0.86318,
|
| 426 |
+
"map_at_50": 0.86383,
|
| 427 |
+
"map_at_100": 0.86394,
|
| 428 |
+
"recall_at_1": 0.822,
|
| 429 |
+
"recall_at_3": 0.89,
|
| 430 |
+
"recall_at_5": 0.916,
|
| 431 |
+
"recall_at_10": 0.938,
|
| 432 |
+
"recall_at_20": 0.968,
|
| 433 |
+
"recall_at_50": 0.986,
|
| 434 |
+
"recall_at_100": 0.994,
|
| 435 |
+
"precision_at_1": 0.822,
|
| 436 |
+
"precision_at_3": 0.29667,
|
| 437 |
+
"precision_at_5": 0.1832,
|
| 438 |
+
"precision_at_10": 0.0938,
|
| 439 |
+
"precision_at_20": 0.0484,
|
| 440 |
+
"precision_at_50": 0.01972,
|
| 441 |
+
"precision_at_100": 0.00994,
|
| 442 |
+
"mrr_at_1": 0.81,
|
| 443 |
+
"mrr_at_3": 0.8446666666666667,
|
| 444 |
+
"mrr_at_5": 0.8526666666666665,
|
| 445 |
+
"mrr_at_10": 0.8558063492063492,
|
| 446 |
+
"mrr_at_20": 0.8576622827722674,
|
| 447 |
+
"mrr_at_50": 0.8582444004699996,
|
| 448 |
+
"mrr_at_100": 0.858351807213441,
|
| 449 |
+
"naucs_at_1_max": 0.6618424420480973,
|
| 450 |
+
"naucs_at_1_std": 0.1726204386872768,
|
| 451 |
+
"naucs_at_1_diff1": 0.8807854527648874,
|
| 452 |
+
"naucs_at_3_max": 0.6544884460097947,
|
| 453 |
+
"naucs_at_3_std": 0.2072674168885846,
|
| 454 |
+
"naucs_at_3_diff1": 0.8465338029378915,
|
| 455 |
+
"naucs_at_5_max": 0.6638877773331578,
|
| 456 |
+
"naucs_at_5_std": 0.23538304210572925,
|
| 457 |
+
"naucs_at_5_diff1": 0.8338224178560337,
|
| 458 |
+
"naucs_at_10_max": 0.6184000481913182,
|
| 459 |
+
"naucs_at_10_std": 0.1655371826149804,
|
| 460 |
+
"naucs_at_10_diff1": 0.8101713803801093,
|
| 461 |
+
"naucs_at_20_max": 0.7978232959850622,
|
| 462 |
+
"naucs_at_20_std": 0.47426470588234737,
|
| 463 |
+
"naucs_at_20_diff1": 0.8186566293183944,
|
| 464 |
+
"naucs_at_50_max": 0.9019607843137241,
|
| 465 |
+
"naucs_at_50_std": 0.7225556889422341,
|
| 466 |
+
"naucs_at_50_diff1": 0.8832866479925231,
|
| 467 |
+
"naucs_at_100_max": 1.0,
|
| 468 |
+
"naucs_at_100_std": 0.8078120136943184,
|
| 469 |
+
"naucs_at_100_diff1": 0.8202614379084721
|
| 470 |
+
},
|
| 471 |
+
"tatdqa_test": {
|
| 472 |
+
"ndcg_at_1": 0.60571,
|
| 473 |
+
"ndcg_at_3": 0.70518,
|
| 474 |
+
"ndcg_at_5": 0.73148,
|
| 475 |
+
"ndcg_at_10": 0.75212,
|
| 476 |
+
"ndcg_at_20": 0.76107,
|
| 477 |
+
"ndcg_at_50": 0.76775,
|
| 478 |
+
"ndcg_at_100": 0.7707,
|
| 479 |
+
"map_at_1": 0.60571,
|
| 480 |
+
"map_at_3": 0.68094,
|
| 481 |
+
"map_at_5": 0.69555,
|
| 482 |
+
"map_at_10": 0.70408,
|
| 483 |
+
"map_at_20": 0.70656,
|
| 484 |
+
"map_at_50": 0.7076,
|
| 485 |
+
"map_at_100": 0.70786,
|
| 486 |
+
"recall_at_1": 0.60571,
|
| 487 |
+
"recall_at_3": 0.77521,
|
| 488 |
+
"recall_at_5": 0.839,
|
| 489 |
+
"recall_at_10": 0.90279,
|
| 490 |
+
"recall_at_20": 0.93803,
|
| 491 |
+
"recall_at_50": 0.97205,
|
| 492 |
+
"recall_at_100": 0.99028,
|
| 493 |
+
"precision_at_1": 0.60571,
|
| 494 |
+
"precision_at_3": 0.2584,
|
| 495 |
+
"precision_at_5": 0.1678,
|
| 496 |
+
"precision_at_10": 0.09028,
|
| 497 |
+
"precision_at_20": 0.0469,
|
| 498 |
+
"precision_at_50": 0.01944,
|
| 499 |
+
"precision_at_100": 0.0099,
|
| 500 |
+
"mrr_at_1": 0.6057108140947752,
|
| 501 |
+
"mrr_at_3": 0.6830700688537875,
|
| 502 |
+
"mrr_at_5": 0.6966788173349543,
|
| 503 |
+
"mrr_at_10": 0.7058545005689602,
|
| 504 |
+
"mrr_at_20": 0.707882381086114,
|
| 505 |
+
"mrr_at_50": 0.7091241073972927,
|
| 506 |
+
"mrr_at_100": 0.7093644660054671,
|
| 507 |
+
"naucs_at_1_max": 0.3183541397363705,
|
| 508 |
+
"naucs_at_1_std": 0.09826700774797242,
|
| 509 |
+
"naucs_at_1_diff1": 0.7187936758795999,
|
| 510 |
+
"naucs_at_3_max": 0.37504605152871257,
|
| 511 |
+
"naucs_at_3_std": 0.14677854908146837,
|
| 512 |
+
"naucs_at_3_diff1": 0.6383523296619431,
|
| 513 |
+
"naucs_at_5_max": 0.42709986452798276,
|
| 514 |
+
"naucs_at_5_std": 0.22990979779837553,
|
| 515 |
+
"naucs_at_5_diff1": 0.6357454223598389,
|
| 516 |
+
"naucs_at_10_max": 0.48845166873972423,
|
| 517 |
+
"naucs_at_10_std": 0.33255307682895896,
|
| 518 |
+
"naucs_at_10_diff1": 0.6184766146789907,
|
| 519 |
+
"naucs_at_20_max": 0.45722581804053714,
|
| 520 |
+
"naucs_at_20_std": 0.3272040057998003,
|
| 521 |
+
"naucs_at_20_diff1": 0.6092689407956671,
|
| 522 |
+
"naucs_at_50_max": 0.4613805353039302,
|
| 523 |
+
"naucs_at_50_std": 0.35387081782577695,
|
| 524 |
+
"naucs_at_50_diff1": 0.5797820548654825,
|
| 525 |
+
"naucs_at_100_max": 0.5928666738326348,
|
| 526 |
+
"naucs_at_100_std": 0.40800919961891313,
|
| 527 |
+
"naucs_at_100_diff1": 0.5890653536441407
|
| 528 |
+
},
|
| 529 |
+
"docvqa_test_subsampled": {
|
| 530 |
+
"ndcg_at_1": 0.47672,
|
| 531 |
+
"ndcg_at_3": 0.55229,
|
| 532 |
+
"ndcg_at_5": 0.56851,
|
| 533 |
+
"ndcg_at_10": 0.58892,
|
| 534 |
+
"ndcg_at_20": 0.60374,
|
| 535 |
+
"ndcg_at_50": 0.6134,
|
| 536 |
+
"ndcg_at_100": 0.62175,
|
| 537 |
+
"map_at_1": 0.47672,
|
| 538 |
+
"map_at_3": 0.53326,
|
| 539 |
+
"map_at_5": 0.54213,
|
| 540 |
+
"map_at_10": 0.55075,
|
| 541 |
+
"map_at_20": 0.55494,
|
| 542 |
+
"map_at_50": 0.55648,
|
| 543 |
+
"map_at_100": 0.55724,
|
| 544 |
+
"recall_at_1": 0.47672,
|
| 545 |
+
"recall_at_3": 0.60754,
|
| 546 |
+
"recall_at_5": 0.64745,
|
| 547 |
+
"recall_at_10": 0.70953,
|
| 548 |
+
"recall_at_20": 0.76718,
|
| 549 |
+
"recall_at_50": 0.81596,
|
| 550 |
+
"recall_at_100": 0.86696,
|
| 551 |
+
"precision_at_1": 0.47672,
|
| 552 |
+
"precision_at_3": 0.20251,
|
| 553 |
+
"precision_at_5": 0.12949,
|
| 554 |
+
"precision_at_10": 0.07095,
|
| 555 |
+
"precision_at_20": 0.03836,
|
| 556 |
+
"precision_at_50": 0.01632,
|
| 557 |
+
"precision_at_100": 0.00867,
|
| 558 |
+
"mrr_at_1": 0.48337028824833705,
|
| 559 |
+
"mrr_at_3": 0.5354767184035477,
|
| 560 |
+
"mrr_at_5": 0.5487804878048779,
|
| 561 |
+
"mrr_at_10": 0.5557869637137927,
|
| 562 |
+
"mrr_at_20": 0.5593928559116855,
|
| 563 |
+
"mrr_at_50": 0.5615025624726138,
|
| 564 |
+
"mrr_at_100": 0.5621791399632005,
|
| 565 |
+
"naucs_at_1_max": 0.3785349172346823,
|
| 566 |
+
"naucs_at_1_std": 0.6557287667252649,
|
| 567 |
+
"naucs_at_1_diff1": 0.8526440408297176,
|
| 568 |
+
"naucs_at_3_max": 0.28374137300417,
|
| 569 |
+
"naucs_at_3_std": 0.7269399252140738,
|
| 570 |
+
"naucs_at_3_diff1": 0.7922662049837966,
|
| 571 |
+
"naucs_at_5_max": 0.23868786072943135,
|
| 572 |
+
"naucs_at_5_std": 0.7281901708135905,
|
| 573 |
+
"naucs_at_5_diff1": 0.7793870480543929,
|
| 574 |
+
"naucs_at_10_max": 0.1650957871644876,
|
| 575 |
+
"naucs_at_10_std": 0.7712043441704308,
|
| 576 |
+
"naucs_at_10_diff1": 0.7842842146130585,
|
| 577 |
+
"naucs_at_20_max": 0.0797800670615453,
|
| 578 |
+
"naucs_at_20_std": 0.8135232985740711,
|
| 579 |
+
"naucs_at_20_diff1": 0.7706236528786001,
|
| 580 |
+
"naucs_at_50_max": 0.07226615905116034,
|
| 581 |
+
"naucs_at_50_std": 0.8406489117674839,
|
| 582 |
+
"naucs_at_50_diff1": 0.7290414525233613,
|
| 583 |
+
"naucs_at_100_max": -0.023763600742901323,
|
| 584 |
+
"naucs_at_100_std": 0.8665455621668288,
|
| 585 |
+
"naucs_at_100_diff1": 0.6958809819906159
|
| 586 |
+
}
|
| 587 |
+
}
|
| 588 |
+
}
|