jev-training-data/logs/r41_train_self.log

23 lines
960 B
Plaintext

[jev] Phase 1: env setup
[jev] laya.common OK
[jev] data OK (cached)
[jev] Phase 2: download mmBERT (ModelScope)
[jev] config.json: 0MB
[jev] tokenizer.json: 16MB
[jev] tokenizer_config.json: 0MB
[jev] special_tokens_map.json: 0MB
[jev] pytorch_model.bin: 1174MB
[jev] device=cuda tiny=False max_len=1024 head_len=160 win=800 aug_p=0.2
[jev] Phase 3: build DecisionModel
[jev] encoder load: ckpt=138 matched=134 unexpected=4 core_missing=0
[jev] model on cuda, trainable 17,267,203
[jev] crosspair long negs: +2499 (total 23125)
[jev] train=23125 calib=5291
[jev] ep-1 pretrain calib {'auc': 0.3696, 'ece': 0.0436, 'n': 5291}
[jev] saved best (auc=0.9832)
[jev] TRAINING-DONE best_auc=0.9832
[jev] CALIBRATION T=1.35 raw(ece=0.0206) calibrated(ece=0.012) auc=0.9941
[jev] AGG-CALIB {'n': 300, 'auc_max_raw': 0.6027, 'ece_max_raw': 0.1124, 'T_agg': 1.3, 'ece_max_calibrated': 0.1122}
[jev] Phase 5: push back
[jev] adapter params=17,267,206