LLaMA-Factory微调DeepSeek-R1-Distill-Qwen-7B

1.数据准备

为了对比原生模型效果与微调后的效果，这里选择医疗诊断数据medical-o1-reasoning-SFT来进行微调实验，首先将数据转化为LLaMA-Factory支持的Alpaca数据格式,并划分数据集

{
  "instruction": "医疗问题示例",
  "input": "上下文信息",
  "output": "预期回答"
}

相关脚本如下：

from datasets import load_dataset
import json, os

dataset = load_dataset('/workspace/luoshiyong/deepseek-ai/medical_sft/medical-o1-reasoning-SFT', 'zh')
dataset = {"train": dataset['train'][:2000], 'val': dataset['train'][-200:]}
root = '/workspace/luoshiyong/deepseek-ai/medical_sft/convert'
for key in ['train', 'val']:
    list_r = []
    for q, r