| set -x | |
| export PYTHONUNBUFFERED=1 | |
| MODEL_PATH=/mnt/moonfs/kimiv-ksyun/xulin/abs/LLaMA-Factory/saves/qwen2_5vl-7b/full/ocr_sft_llm | |
| python3 -m verl.trainer.main \ | |
| config=examples/ocr_llm_llm.yaml \ | |
| data.train_files=/mnt/moonfs/kimiv-ksyun/xulin/datasets/OCR-VQA/ocr_vqa_clean_dataset@train \ | |
| # data.val_files=/mnt/moonfs/kimiv-ksyun/xulin/datasets/OCR-VQA/ocr_vqa_clean_dataset/validation_subset_1000.parquet \ | |
| worker.actor.model.model_path=${MODEL_PATH} \ | |
| trainer.experiment_name=ocr_llm_llm \ | |
| trainer.n_gpus_per_node=8 | |