[jev] device=cuda base=/tmp/base tokenizers=0.22.2 Some weights of ModernBertForSequenceClassification were not initialized from the model checkpoint at /tmp/base and are newly initialized: ['classifier.bias', 'classifier.weight'] You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. [jev] LoRA targets: ['Wqkv', 'Wo'] trainable params: 2,298,625 || all params: 309,829,634 || trainable%: 0.7419 [jev] train=27954 calib=5291 [jev] canaries: ['en', 'zh'] [jev] ep0 step100/1746 loss=0.2554 [jev] ep0 step200/1746 loss=0.4286 [jev] ep0 step300/1746 loss=0.4042 [jev] ep0 step400/1746 loss=0.3410 [jev] ep0 step500/1746 loss=0.3232 [jev] ep0 step600/1746 loss=0.2050 [jev] ep0 step700/1746 loss=0.2220 [jev] ep0 step800/1746 loss=0.2669 /usr/local/lib/python3.10/dist-packages/torch/_inductor/compile_fx.py:282: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance. warnings.warn( [jev] epoch 0 calibration: {'auc': 0.9936, 'ece': 0.0129, 'n': 5291} [jev] epoch 0 canary_en: {'auc': 0.8138, 'ece': 0.075, 'n': 300} [jev] epoch 0 canary_zh: {'auc': 1.0, 'ece': 0.0038, 'n': 300} [jev] live-checkpoint 已写 /tmp/jd/model_live [jev] ep1 step900/1746 loss=0.2454 [jev] ep1 step1000/1746 loss=0.2127 [jev] ep1 step1100/1746 loss=0.1612 [jev] ep1 step1200/1746 loss=0.1135 [jev] ep1 step1300/1746 loss=0.1401 [jev] ep1 step1400/1746 loss=0.1381 [jev] ep1 step1500/1746 loss=0.3087 [jev] ep1 step1600/1746 loss=0.1147 [jev] ep1 step1700/1746 loss=0.1655