geolip-captionbert-8192 / benchmarks /captionbert_8192_post_upgrade_60epoch_5test.json
AbstractPhil's picture
Create benchmarks/captionbert_8192_post_upgrade_60epoch_5test.json
d0976e9 verified
{
"captionbert": {
"stsb": {
"spearman": 0.5032419166436953,
"pearson": 0.5099728825250903,
"cos_mean": 0.8032620549201965
},
"sick": {
"spearman": 0.6138000143572749,
"pearson": 0.6645180256078914,
"cos_mean": 0.8160619139671326
},
"mrpc": {
"f1": 0.8067528735632183,
"accuracy": 0.6881159420289855,
"threshold": 0.7100000000000002
},
"retrieval": {
"mean_top1_cos": 0.5476557016372681,
"mean_top5_cos": 0.4852602183818817,
"mean_top10_cos": 0.449970543384552,
"self_cos_mean": 0.003989221062511206,
"self_cos_max": 0.7181165218353271
},
"params": 25958016
},
"bert-base": {
"stsb": {
"spearman": 0.472910817217408,
"pearson": 0.47908503084020204,
"cos_mean": 0.8478000164031982
},
"sick": {
"spearman": 0.5864509004196323,
"pearson": 0.6101682762599208,
"cos_mean": 0.8693334460258484
},
"mrpc": {
"f1": 0.8061626657112146,
"accuracy": 0.6863768115942029,
"threshold": 0.8000000000000003
},
"retrieval": {
"mean_top1_cos": 0.9359426498413086,
"mean_top5_cos": 0.9275503158569336,
"mean_top10_cos": 0.9225468635559082,
"self_cos_mean": 0.8072085976600647,
"self_cos_max": 0.9807626008987427
},
"params": 109482240
},
"modern-bert": {
"stsb": {
"spearman": 0.42145967861852557,
"pearson": 0.3676700049123431,
"cos_mean": 0.962863564491272
},
"sick": {
"spearman": 0.5479310614625831,
"pearson": 0.5677992239217774,
"cos_mean": 0.9647596478462219
},
"mrpc": {
"f1": 0.8038418913926856,
"accuracy": 0.6921739130434783,
"threshold": 0.9600000000000004
},
"retrieval": {
"mean_top1_cos": 0.98164963722229,
"mean_top5_cos": 0.9794968962669373,
"mean_top10_cos": 0.9782372117042542,
"self_cos_mean": 0.948133647441864,
"self_cos_max": 0.991678774356842
},
"params": 149014272
},
"roberta": {
"stsb": {
"spearman": 0.5435540620132823,
"pearson": 0.5237119708674596,
"cos_mean": 0.9867845177650452
},
"sick": {
"spearman": 0.629613720814436,
"pearson": 0.6470603199376301,
"cos_mean": 0.987252950668335
},
"mrpc": {
"f1": 0.8122199592668025,
"accuracy": 0.7327536231884058,
"threshold": 0.9900000000000004
},
"retrieval": {
"mean_top1_cos": 0.9934692978858948,
"mean_top5_cos": 0.9925827383995056,
"mean_top10_cos": 0.9920614957809448,
"self_cos_mean": 0.9759101271629333,
"self_cos_max": 0.9974486827850342
},
"params": 124645632
},
"albert": {
"stsb": {
"spearman": 0.4783865016260685,
"pearson": 0.4657200438279754,
"cos_mean": 0.919990062713623
},
"sick": {
"spearman": 0.5364140177707308,
"pearson": 0.5652283882710937,
"cos_mean": 0.9238592386245728
},
"mrpc": {
"f1": 0.8067047075606276,
"accuracy": 0.6857971014492754,
"threshold": 0.9000000000000004
},
"retrieval": {
"mean_top1_cos": 0.9703265428543091,
"mean_top5_cos": 0.9666717052459717,
"mean_top10_cos": 0.9645306468009949,
"self_cos_mean": 0.9050750732421875,
"self_cos_max": 0.9897439479827881
},
"params": 11683584
},
"distilbert": {
"stsb": {
"spearman": 0.5716834757596537,
"pearson": 0.5676559551414473,
"cos_mean": 0.9003205299377441
},
"sick": {
"spearman": 0.6423566433306357,
"pearson": 0.6965453520447837,
"cos_mean": 0.9173732995986938
},
"mrpc": {
"f1": 0.8054755043227666,
"accuracy": 0.6869565217391305,
"threshold": 0.8700000000000003
},
"retrieval": {
"mean_top1_cos": 0.9534780979156494,
"mean_top5_cos": 0.9466020464897156,
"mean_top10_cos": 0.9423779249191284,
"self_cos_mean": 0.840302050113678,
"self_cos_max": 0.9858899116516113
},
"params": 66362880
}
}