grt-30m-embedding / eval_results.json
davidhouseholter's picture
Release PairAdam-trained GRT-30M-Embedding checkpoint with AdamW ablation note
23e04b5 verified
Raw History Blame Contribute Delete
462 Bytes
{
"optimizer": "PairAdam",
"stsb_dev_spearman_before": 0.4551,
"stsb_dev_spearman_after": 0.7179,
"stsb_test_spearman_before": 0.2849,
"stsb_test_spearman_after": 0.6242,
"stsb_test_pearson_before": 0.2981,
"stsb_test_pearson_after": 0.6330,
"adamw_ablation": {
"stsb_dev_spearman_after": 0.7289,
"stsb_test_spearman_after": 0.6317,
"stsb_test_pearson_after": 0.6372
},
"parameters": 29305497,
"trainable_parameters": 9891865
}