Bong

I faced same issue.

### Training config ```json { "stage": "dpo", "do_train": true, "model_name_or_path": "(removed)", "dataset": "(removed)", "dataset_dir": "../data", "template": "qwen", "finetuning_type": "full", "output_dir": "(removed)", "overwrite_cache": true, "overwrite_output_dir": true, "cutoff_len": 4096, "val_size": 0, "evaluation_strategy":...