88 lines
1.5 KiB
Markdown
88 lines
1.5 KiB
Markdown
|
# 微调指南
|
||
|
|
|
||
|
|
- 本项目不仅在心理健康数据集上进行了微调，同时也对模型进行了自我认知微调，下面是微调的详细指南。
|
||
|
|
|
||
|
|
## 一、基于xtuner的微调🎉🎉🎉🎉🎉
|
||
|
|
|
||
|
|
### 环境准备
|
||
|
|
|
||
|
|
```markdown
|
||
|
|
datasets==2.16.1
|
||
|
|
deepspeed==0.13.1
|
||
|
|
einops==0.7.0
|
||
|
|
flash_attn==2.5.0
|
||
|
|
mmengine==0.10.2
|
||
|
|
openxlab==0.0.34
|
||
|
|
peft==0.7.1
|
||
|
|
sentencepiece==0.1.99
|
||
|
|
torch==2.1.2
|
||
|
|
transformers==4.36.2
|
||
|
|
xtuner==0.1.11
|
||
|
|
```
|
||
|
|
|
||
|
|
也可以一键安装
|
||
|
|
|
||
|
|
```bash
|
||
|
|
cd xtuner_config/
|
||
|
|
pip3 install -r requirements.txt
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
### 微调
|
||
|
|
|
||
|
|
```bash
|
||
|
|
cd xtuner_config/
|
||
|
|
xtuner train internlm2_7b_chat_qlora_e3.py --deepspeed deepspeed_zero2
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
### 将得到的 PTH 模型转换为 HuggingFace 模型
|
||
|
|
|
||
|
|
**即：生成 Adapter 文件夹**
|
||
|
|
|
||
|
|
```bash
|
||
|
|
cd xtuner_config/
|
||
|
|
mkdir hf
|
||
|
|
export MKL_SERVICE_FORCE_INTEL=1
|
||
|
|
|
||
|
|
xtuner convert pth_to_hf internlm2_7b_chat_qlora_e3.py ./work_dirs/internlm_chat_7b_qlora_oasst1_e3_copy/epoch_3.pth ./hf
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
### 将 HuggingFace adapter 合并到大语言模型
|
||
|
|
|
||
|
|
```bash
|
||
|
|
xtuner convert merge ./internlm2-chat-7b ./hf ./merged --max-shard-size 2GB
|
||
|
|
# xtuner convert merge \
|
||
|
|
# ${NAME_OR_PATH_TO_LLM} \
|
||
|
|
# ${NAME_OR_PATH_TO_ADAPTER} \
|
||
|
|
# ${SAVE_PATH} \
|
||
|
|
# --max-shard-size 2GB
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
### 测试
|
||
|
|
|
||
|
|
```
|
||
|
|
cd demo/
|
||
|
|
python cli_internlm2.py
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 二、基于Transformers的微调🎉🎉🎉🎉🎉
|
||
|
|
|
||
|
|
- 请查看[ChatGLM3-6b lora微调指南](ChatGLM3-6b-ft.md)
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## 其他
|
||
|
|
|
||
|
|
欢迎大家给[xtuner](https://github.com/InternLM/xtuner)和[EmoLLM](https://github.com/aJupyter/EmoLLM)点点star~
|
||
|
|
|
||
|
|
🎉🎉🎉🎉🎉
|