Create benchmarks/captionbert_8192_post_upgrade_60epoch_5test.json
Browse files
benchmarks/captionbert_8192_post_upgrade_60epoch_5test.json
ADDED
|
@@ -0,0 +1,152 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"captionbert": {
|
| 3 |
+
"stsb": {
|
| 4 |
+
"spearman": 0.5032419166436953,
|
| 5 |
+
"pearson": 0.5099728825250903,
|
| 6 |
+
"cos_mean": 0.8032620549201965
|
| 7 |
+
},
|
| 8 |
+
"sick": {
|
| 9 |
+
"spearman": 0.6138000143572749,
|
| 10 |
+
"pearson": 0.6645180256078914,
|
| 11 |
+
"cos_mean": 0.8160619139671326
|
| 12 |
+
},
|
| 13 |
+
"mrpc": {
|
| 14 |
+
"f1": 0.8067528735632183,
|
| 15 |
+
"accuracy": 0.6881159420289855,
|
| 16 |
+
"threshold": 0.7100000000000002
|
| 17 |
+
},
|
| 18 |
+
"retrieval": {
|
| 19 |
+
"mean_top1_cos": 0.5476557016372681,
|
| 20 |
+
"mean_top5_cos": 0.4852602183818817,
|
| 21 |
+
"mean_top10_cos": 0.449970543384552,
|
| 22 |
+
"self_cos_mean": 0.003989221062511206,
|
| 23 |
+
"self_cos_max": 0.7181165218353271
|
| 24 |
+
},
|
| 25 |
+
"params": 25958016
|
| 26 |
+
},
|
| 27 |
+
"bert-base": {
|
| 28 |
+
"stsb": {
|
| 29 |
+
"spearman": 0.472910817217408,
|
| 30 |
+
"pearson": 0.47908503084020204,
|
| 31 |
+
"cos_mean": 0.8478000164031982
|
| 32 |
+
},
|
| 33 |
+
"sick": {
|
| 34 |
+
"spearman": 0.5864509004196323,
|
| 35 |
+
"pearson": 0.6101682762599208,
|
| 36 |
+
"cos_mean": 0.8693334460258484
|
| 37 |
+
},
|
| 38 |
+
"mrpc": {
|
| 39 |
+
"f1": 0.8061626657112146,
|
| 40 |
+
"accuracy": 0.6863768115942029,
|
| 41 |
+
"threshold": 0.8000000000000003
|
| 42 |
+
},
|
| 43 |
+
"retrieval": {
|
| 44 |
+
"mean_top1_cos": 0.9359426498413086,
|
| 45 |
+
"mean_top5_cos": 0.9275503158569336,
|
| 46 |
+
"mean_top10_cos": 0.9225468635559082,
|
| 47 |
+
"self_cos_mean": 0.8072085976600647,
|
| 48 |
+
"self_cos_max": 0.9807626008987427
|
| 49 |
+
},
|
| 50 |
+
"params": 109482240
|
| 51 |
+
},
|
| 52 |
+
"modern-bert": {
|
| 53 |
+
"stsb": {
|
| 54 |
+
"spearman": 0.42145967861852557,
|
| 55 |
+
"pearson": 0.3676700049123431,
|
| 56 |
+
"cos_mean": 0.962863564491272
|
| 57 |
+
},
|
| 58 |
+
"sick": {
|
| 59 |
+
"spearman": 0.5479310614625831,
|
| 60 |
+
"pearson": 0.5677992239217774,
|
| 61 |
+
"cos_mean": 0.9647596478462219
|
| 62 |
+
},
|
| 63 |
+
"mrpc": {
|
| 64 |
+
"f1": 0.8038418913926856,
|
| 65 |
+
"accuracy": 0.6921739130434783,
|
| 66 |
+
"threshold": 0.9600000000000004
|
| 67 |
+
},
|
| 68 |
+
"retrieval": {
|
| 69 |
+
"mean_top1_cos": 0.98164963722229,
|
| 70 |
+
"mean_top5_cos": 0.9794968962669373,
|
| 71 |
+
"mean_top10_cos": 0.9782372117042542,
|
| 72 |
+
"self_cos_mean": 0.948133647441864,
|
| 73 |
+
"self_cos_max": 0.991678774356842
|
| 74 |
+
},
|
| 75 |
+
"params": 149014272
|
| 76 |
+
},
|
| 77 |
+
"roberta": {
|
| 78 |
+
"stsb": {
|
| 79 |
+
"spearman": 0.5435540620132823,
|
| 80 |
+
"pearson": 0.5237119708674596,
|
| 81 |
+
"cos_mean": 0.9867845177650452
|
| 82 |
+
},
|
| 83 |
+
"sick": {
|
| 84 |
+
"spearman": 0.629613720814436,
|
| 85 |
+
"pearson": 0.6470603199376301,
|
| 86 |
+
"cos_mean": 0.987252950668335
|
| 87 |
+
},
|
| 88 |
+
"mrpc": {
|
| 89 |
+
"f1": 0.8122199592668025,
|
| 90 |
+
"accuracy": 0.7327536231884058,
|
| 91 |
+
"threshold": 0.9900000000000004
|
| 92 |
+
},
|
| 93 |
+
"retrieval": {
|
| 94 |
+
"mean_top1_cos": 0.9934692978858948,
|
| 95 |
+
"mean_top5_cos": 0.9925827383995056,
|
| 96 |
+
"mean_top10_cos": 0.9920614957809448,
|
| 97 |
+
"self_cos_mean": 0.9759101271629333,
|
| 98 |
+
"self_cos_max": 0.9974486827850342
|
| 99 |
+
},
|
| 100 |
+
"params": 124645632
|
| 101 |
+
},
|
| 102 |
+
"albert": {
|
| 103 |
+
"stsb": {
|
| 104 |
+
"spearman": 0.4783865016260685,
|
| 105 |
+
"pearson": 0.4657200438279754,
|
| 106 |
+
"cos_mean": 0.919990062713623
|
| 107 |
+
},
|
| 108 |
+
"sick": {
|
| 109 |
+
"spearman": 0.5364140177707308,
|
| 110 |
+
"pearson": 0.5652283882710937,
|
| 111 |
+
"cos_mean": 0.9238592386245728
|
| 112 |
+
},
|
| 113 |
+
"mrpc": {
|
| 114 |
+
"f1": 0.8067047075606276,
|
| 115 |
+
"accuracy": 0.6857971014492754,
|
| 116 |
+
"threshold": 0.9000000000000004
|
| 117 |
+
},
|
| 118 |
+
"retrieval": {
|
| 119 |
+
"mean_top1_cos": 0.9703265428543091,
|
| 120 |
+
"mean_top5_cos": 0.9666717052459717,
|
| 121 |
+
"mean_top10_cos": 0.9645306468009949,
|
| 122 |
+
"self_cos_mean": 0.9050750732421875,
|
| 123 |
+
"self_cos_max": 0.9897439479827881
|
| 124 |
+
},
|
| 125 |
+
"params": 11683584
|
| 126 |
+
},
|
| 127 |
+
"distilbert": {
|
| 128 |
+
"stsb": {
|
| 129 |
+
"spearman": 0.5716834757596537,
|
| 130 |
+
"pearson": 0.5676559551414473,
|
| 131 |
+
"cos_mean": 0.9003205299377441
|
| 132 |
+
},
|
| 133 |
+
"sick": {
|
| 134 |
+
"spearman": 0.6423566433306357,
|
| 135 |
+
"pearson": 0.6965453520447837,
|
| 136 |
+
"cos_mean": 0.9173732995986938
|
| 137 |
+
},
|
| 138 |
+
"mrpc": {
|
| 139 |
+
"f1": 0.8054755043227666,
|
| 140 |
+
"accuracy": 0.6869565217391305,
|
| 141 |
+
"threshold": 0.8700000000000003
|
| 142 |
+
},
|
| 143 |
+
"retrieval": {
|
| 144 |
+
"mean_top1_cos": 0.9534780979156494,
|
| 145 |
+
"mean_top5_cos": 0.9466020464897156,
|
| 146 |
+
"mean_top10_cos": 0.9423779249191284,
|
| 147 |
+
"self_cos_mean": 0.840302050113678,
|
| 148 |
+
"self_cos_max": 0.9858899116516113
|
| 149 |
+
},
|
| 150 |
+
"params": 66362880
|
| 151 |
+
}
|
| 152 |
+
}
|