File size: 1,638 Bytes
f95efe0
 
 
 
 
 
 
 
 
 
 
 
 
 
b730de1
f95efe0
 
 
b730de1
 
63250c7
f95efe0
b730de1
 
 
f95efe0
b730de1
 
f95efe0
 
b730de1
 
f95efe0
 
 
 
 
 
 
 
 
 
 
b730de1
f95efe0
 
 
 
b730de1
 
48f4786
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
---
license: apache-2.0
datasets:
- shareAI/DPO-zh-en-emoji
language:
- zh
- en
pipeline_tag: question-answering
tags:
- dpo
- llama3.1
- llama3
- chat
---
# llama3.1-8b-instruct 中文DPO版
### 模型介绍
像原版instruct一样,喜欢用有趣中文和表情符号回答问题。  
Github:https://github.com/CrazyBoyM/llama3-Chinese-chat  

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)  
![Alt text](https://modelscope.cn/api/v1/models/baicai003/Llama3-Chinese-instruct-DPO-beta0.5/repo?Revision=master&FilePath=image.png&View=true)

### 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
```

目前已经上传 Q4 K_M 的GGUF量化模型,后续将进一步进行手机端或边缘算力端展示