54 lines
1.5 KiB
Markdown
54 lines
1.5 KiB
Markdown
---
|
||
frameworks:
|
||
- Pytorch
|
||
license: Apache License 2.0
|
||
tasks:
|
||
- chatbot
|
||
|
||
language:
|
||
- cn
|
||
|
||
tags:
|
||
- RL-tuned
|
||
|
||
tools:
|
||
- vllm
|
||
---
|
||
# llama3.1-8b-instruct 中文DPO版
|
||
### 模型介绍
|
||
Github：[llama3中文仓库](https://github.com/CrazyBoyM/llama3-Chinese-chat)
|
||
像原版instruct一样，喜欢用有趣中文和表情符号回答问题。
|
||
|
||
|
||
特点：偏好中文和emoji表情，且不损伤原instruct版模型能力。实测中文DPO版问答性能体验超过现在市面上任何llama3.1中文微调版 （微调会大面积破坏llama3.1原版能力，导致遗忘）
|
||
|
||
|
||
### 训练细节
|
||
DPO(beta 0.5) + lora rank128, alpha256 + 打开"lm_head", "input_layernorm", "post_attention_layernorm", "norm"层训练.
|
||
|
||
### 模型部署
|
||
网页脚本文件：https://github.com/CrazyBoyM/llama3-Chinese-chat/blob/main/deploy/web_streamlit_for_instruct_v2.py
|
||
已经在模型文件夹里内置了一份web.py，可以直接使用。
|
||
```
|
||
pip install streamlit
|
||
pip install transformers==4.40.1
|
||
streamlit run web.py ./llama3.1-8b-instruct-dpo-zh
|
||
```
|
||
|
||
### 模型下载
|
||
SDK下载
|
||
```bash
|
||
#安装ModelScope
|
||
pip install modelscope
|
||
```
|
||
```python
|
||
#SDK模型下载
|
||
from modelscope import snapshot_download
|
||
model_dir = snapshot_download('shareAI/llama3.1-8b-instruct-dpo-zh')
|
||
```
|
||
Git下载
|
||
```
|
||
#Git模型下载
|
||
git clone https://www.modelscope.cn/shareAI/llama3.1-8b-instruct-dpo-zh.git
|
||
```
|