update
Browse files- README.md +42 -0
- resources/img.png +0 -0
- resources/test1.png +0 -0
- resources/test2.png +0 -0
- resources/test3.png +0 -0
- resources/test4.png +0 -0
- resources/test5.png +0 -0
- resources/test6.png +0 -0
README.md
CHANGED
@@ -1,3 +1,45 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
+
datasets:
|
4 |
+
- BelleGroup/train_2M_CN
|
5 |
+
- BelleGroup/train_3.5M_CN
|
6 |
+
- BelleGroup/train_1M_CN
|
7 |
+
- BelleGroup/train_0.5M_CN
|
8 |
+
- BelleGroup/school_math_0.25M
|
9 |
+
language:
|
10 |
+
- zh
|
11 |
---
|
12 |
+
|
13 |
+
## GoGPT
|
14 |
+
|
15 |
+
基于中文指令数据微调BLOOM
|
16 |
+
![img.png](resources/img.png)
|
17 |
+
> 训练第一轮足够了,后续第二轮和第三轮提升不大
|
18 |
+
|
19 |
+
- 🚀多样性指令数据
|
20 |
+
- 🚀筛选高质量中文数据
|
21 |
+
|
22 |
+
| 模型名字 | 参数量 | 模型地址 |
|
23 |
+
|------------|--------|------|
|
24 |
+
| gogpt-560m | 5.6亿参数 | 🤗[golaxy/gogpt-560m](https://huggingface.co/golaxy/gogpt-560m) |
|
25 |
+
| gogpt-3b | 30亿参数 | 🤗[golaxy/gogpt-3b](https://huggingface.co/golaxy/gogpt-3b) |
|
26 |
+
|
27 |
+
|
28 |
+
## 测试效果
|
29 |
+
![img.png](resources/test1.png)
|
30 |
+
![img.png](resources/test2.png)
|
31 |
+
![img.png](resources/test3.png)
|
32 |
+
![img.png](resources/test4.png)
|
33 |
+
![img.png](resources/test5.png)
|
34 |
+
![img.png](resources/test6.png)
|
35 |
+
|
36 |
+
|
37 |
+
## TODO
|
38 |
+
- 进行RLFH训练
|
39 |
+
- 后续加入中英平行语料
|
40 |
+
|
41 |
+
## 感谢
|
42 |
+
|
43 |
+
- [@hz大佬-zero_nlp](https://github.com/yuanzhoulvpi2017/zero_nlp)
|
44 |
+
- [stanford_alpaca](https://github.com/tatsu-lab/stanford_alpaca)
|
45 |
+
- [Belle数据](https://huggingface.co/BelleGroup)
|
resources/img.png
ADDED
resources/test1.png
ADDED
resources/test2.png
ADDED
resources/test3.png
ADDED
resources/test4.png
ADDED
resources/test5.png
ADDED
resources/test6.png
ADDED