We trained a Chinese version of Shepherd based on Chinese-LLaMA-2-7B, and we used 2 V100 GPUs with 32G for supervised fine-tuning based on LoRA.
We designed the appropriate prompt template, and the dataset we used has been published in this HuggingFace repository: frankminors123/chinese-shepherd-critic-dataset, please go to the data page to view details.
1PROMPT_TEMPLATE = (
2 "请试着评论下面问题的答案.\n"
3 "### 问题:\n{question}\n### 答案:\n{answer}\n### 评论:\n"
4)
