Update README.md
Browse files
README.md
CHANGED
|
@@ -1,11 +1,38 @@
|
|
|
|
|
| 1 |
使用[Firefly](https://github.com/yangjianxin1/Firefly)项目微调baichuan-13b-base。训练数据约为一百万多轮对话数据,包括项目分享的moss数据+2万条school math数据。
|
| 2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
训练loss:
|
|
|
|
|
|
|
| 4 |

|
| 5 |
|
| 6 |
-
更多详情见项目:[Firefly](https://github.com/yangjianxin1/Firefly)
|
| 7 |
|
| 8 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
| 10 |
单轮对话:
|
| 11 |
```python
|
|
@@ -148,4 +175,4 @@ def main():
|
|
| 148 |
|
| 149 |
if __name__ == '__main__':
|
| 150 |
main()
|
| 151 |
-
```
|
|
|
|
| 1 |
+
|
| 2 |
使用[Firefly](https://github.com/yangjianxin1/Firefly)项目微调baichuan-13b-base。训练数据约为一百万多轮对话数据,包括项目分享的moss数据+2万条school math数据。
|
| 3 |
|
| 4 |
+
更多详情见项目:[Firefly](https://github.com/yangjianxin1/Firefly)
|
| 5 |
+
|
| 6 |
+
技术细节分享:[Firefly增强Baichuan-13B的多轮对话能力](https://mp.weixin.qq.com/s/djO8Tg3emmy6wzw_rTUlcw)
|
| 7 |
+
|
| 8 |
训练loss:
|
| 9 |
+
|
| 10 |
+
[//]: # (<img src="https://huggingface.co/YeungNLP/firefly-baichuan-13b/resolve/main/firefly-baichuan-13b-loss.jpg" width="450">)
|
| 11 |

|
| 12 |
|
|
|
|
| 13 |
|
| 14 |
+
C-Eval榜单:
|
| 15 |
+
|
| 16 |
+
| Model | C-Eval | STEM | Social Science | Humanities | Other |
|
| 17 |
+
|----------------------------------|--------|-------|----------------|------------|-------|
|
| 18 |
+
| Baichuan-13B-Chat(官方) | 52.05 | 42.23 | 65.27 | 58.61 | 51.32 |
|
| 19 |
+
| **firefly-baichuan-13b** | 51.36 | 44.24 | 61.65 | 54.63 | 51.68 |
|
| 20 |
+
| chatglm2-6b(官方) | 50.45 | 41.91 | 60.73 | 59.24 | 47.82 |
|
| 21 |
+
| **firefly-chatglm2-6b** | 49.13 | 43.6 | 58.83 | 54.48 | 45.03 |
|
| 22 |
+
| openbuddy-llama2-13b-v11.1-bf16 | 43.36 | 39.79 | 50.28 | 44.78 | 42.13 |
|
| 23 |
+
| chinese-alpaca-2-13b(哈工大) | 41.86 | 36.52 | 49.7 | 47.97 | 38.33 |
|
| 24 |
+
| openbuddy-llama2-13b-v8.1-fp16 | 41.62 | 38.82 | 44.66 | 40.28 | 45.32 |
|
| 25 |
+
| chinese-alpaca-2-7b(哈工大) | 41.48 | 35.01 | 50.08 | 43.02 | 43.87 |
|
| 26 |
+
| belle-llama2-13B-chat-0.4M | 41.11 | 40.04 | 44.71 | 42.09 | 38.82 |
|
| 27 |
+
| ziya-llama-13b | 39.1 | - | - | - | - |
|
| 28 |
+
| llama-2-13b-chat(官方) | 36.38 | 33.68 | 46.38 | 34.47 | 34.1 |
|
| 29 |
+
| lama-2-7b-chat(官方) | 35.86 | 32.85 | 40.04 | 37.37 | 36.01 |
|
| 30 |
+
| flagalpha/Llama2-Chinese-7b-Chat | 34.54 | 35.21 | 37.9 | 33.11 | 31.7 |
|
| 31 |
+
| yayi-13b-llama2 | 34.15 | 36.48 | 30.64 | 32.67 | 34.6 |
|
| 32 |
+
| yayi-7b-llama2 | 30.18 | 25.88 | 38.23 | 34.56 | 26.31 |
|
| 33 |
+
| linly-llama2-7b | 28.35 | 26.06 | 33.47 | 29.71 | 26.53 |
|
| 34 |
+
| linly-llama2-13b | 27.86 | 27.67 | 26.95 | 27.93 | 28.95 |
|
| 35 |
+
|
| 36 |
|
| 37 |
单轮对话:
|
| 38 |
```python
|
|
|
|
| 175 |
|
| 176 |
if __name__ == '__main__':
|
| 177 |
main()
|
| 178 |
+
```
|