File size: 697 Bytes
f1c3aed
 
 
 
 
 
 
 
7bb7590
 
d38b71a
 
f1c3aed
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
language:
- zh
library_name: transformers
tags:
- llama
- llama3
- dpo-zh
- emoji
datasets:
- shareAI/DPO-zh-en-emoji
---

Github:https://github.com/CrazyBoyM/llama3-Chinese-chat  
放出训练配方细节供网友参考分享: 
DPO(beta 0.5) + lora rank128, alpha256 + 打开"lm_head", "input_layernorm", "post_attention_layernorm", "norm"层训练。  
偏好中文和emoji表情,且不损伤原instruct版模型能力。  

![image/png](https://cdn-uploads.huggingface.co/production/uploads/631f5b422225f12fc0f2c838/2xlWxZvN0gahckA2EPmlE.png)

Git下载
```
#Git模型下载
git clone https://www.modelscope.cn/baicai003/Llama3-Chinese-instruct-DPO-beta0.5.git
```