01-InternLM2-Math-Plus-1.8B 微调.md

01-InternLM2-Math-Plus-1.8B 微调

环境搭建

请参考README.md, 本教程重点为如何微调新模型InternLM2-Math-Plus-1.8B.

XTuner微调

1.准备配置文件

# 拉取github上的AMchat代码模型
git clone https://github.com/AXYZdong/AMchat.git
# 将配置文件复制到自己的模型目录下
cp /root/Amath_xtuner/AMchat/config/internlm2_chat_7b_qlora_oasst1_e3_copy.py  /root/Amath_xtuner/config/

2.模型下载

创建python文件/root/Amath_xtuner/data/data_download.py

import os
os.environ["HF_HOME"]='/group_share/hf'
os.environ["MODELSCOPE_CACHE"]='/group_share/modelscope_hub'
os.environ["MODELSCOPE_MODULES_CACHE"]='/group_share/modelscope_modules'

#模型下载
from modelscope import snapshot_download
model_dir = snapshot_download('Shanghai_AI_Laboratory/internlm2-math-plus-1_8b',
                               cache_dir='/group_share/models')

将模型下载到指定的cache_dir目录下

3.修改配置文件

修改/root/Amath_xtuner/config/internlm2_chat_7b_qlora_oasst1_e3_copy.py配置文件：

# 修改模型为本地路径
- pretrained_model_name_or_path = 'internlm/internlm-chat-7b'
+ pretrained_model_name_or_path = '/root/Amath_xtuner/model/internlm2-math-plus-1_8b'

# 修改训练数据集为本地路径
- data_path = 'timdettmers/openassistant-guanaco'
+ data_path = '/root/Amath_xtuner/data/math_xtuner.json'

4.开始微调

# 采用deepspeed_zero2节省更多显存, 降低微调硬件需求
xtuner train /root/Amath_xtuner/config/internlm2_chat_7b_qlora_oasst1_e3_copy.py --work-dir /root/Amath_xtuner/train_deepspeed --deepspeed deepspeed_zero2

5.PTH 模型转换为 HuggingFace 模型

xtuner convert pth_to_hf /root/Amath_xtuner/train_deepspeed/internlm2_chat_7b_qlora_oasst1_e3_copy.py /root/Amath_xtuner/train_deepspeed/epoch_3.pth /root/Amath_xtuner/huggingface

6.HuggingFace 模型合并到大语言模型

export MKL_SERVICE_FORCE_INTEL=1
export MKL_THREADING_LAYER='GNU'# 原始模型参数存放的位置export 
xtuner convert merge /root/Amath_xtuner/model/internlm2-math-plus-1_8b /root/Amath_xtuner/huggingface /root/Amath_xtuner/final_model

7.运行Demo

streamlit run /root/Amath_xtuner/web_demo/InternLM/chat/web_demo.py --server.address 127.0.0.1 --server.port 6006

8.微调后的模型下载地址

• openxlab: 高等数学解题小助手
• modelscope: 高等数学解题小助手