zhichen commited on
Commit
55acd3a
1 Parent(s): b583870

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +189 -3
README.md CHANGED
@@ -1,3 +1,189 @@
1
- ---
2
- license: apache-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <p align="center">
2
+ <a href='https://huggingface.co/spaces/zhichen'>
3
+ <img src='./images/logo.png'>
4
+ </a>
5
+ </p>
6
+
7
+ <div align="center">
8
+ <p align="center">
9
+ <h3> Qwen-WisdomVast (千问-智瀚)</h3>
10
+
11
+ <p align="center">
12
+ <a href='https://huggingface.co/zhichen'>
13
+ <img src='https://img.shields.io/badge/%F0%9F%A4%97%20HuggingFace-Qwen%20WisdomVast-yellow'>
14
+ </a>
15
+ <a href='https://modelscope.cn/profile/seanzhang'>
16
+ <img src='https://img.shields.io/badge/🤖 ModelScope-Qwen%20WisdomVast-blue'>
17
+ </a>
18
+ <br>
19
+ <a href=href="https://github.com/seanzhang-zhichen/Qwen-WisdomVast/stargazers">
20
+ <img src="https://img.shields.io/github/stars/seanzhang-zhichen/Qwen-WisdomVast?color=ccf">
21
+ </a>
22
+ <a href="https://github.com/seanzhang-zhichen/Qwen-WisdomVast/blob/main/LICENSE">
23
+ <img alt="GitHub Contributors" src="https://img.shields.io/badge/license-Apache%202.0-blue.svg" />
24
+ </a>
25
+ </p>
26
+ </div>
27
+
28
+
29
+ ## 介绍
30
+
31
+ **Qwen-WisdomVast**是**以Qwen1.5-7B为底座**,使用 [DORA](https://arxiv.org/pdf/2402.09353.pdf) + [LORA+](https://arxiv.org/pdf/2402.12354.pdf) 的训练方法,在100w高质量中文多轮SFT数据 + 20w英文多轮SFT数据 + 2000单轮自我认知数据训练而来的大模型,**数学能力**相比Qwen1.5-7B-Chat**提升了5.16%**,在**HumanEval**数据集上相比Qwen1.5-7B-Chat**提升了12.8**,在**MBPP**数据集上**提升了11.6%**,在**BBH**数据集上**提升了12.44%**,全部评测表现见下表。
32
+
33
+ ![DEMO](./images/image.png)
34
+
35
+ ## 评测表现
36
+
37
+ | Model | MMLU | C-Eval | GSM8K | MATH | HumanEval | MBPP | BBH |
38
+ |-------------------|-------|--------|-------|-------|-----------|-------|-------|
39
+ | **Qwen1.5-7B-Chat** | 60.88 | 70.18 | 54.13 | 7.96 | 31.10 | 15.00 | 31.67 |
40
+ | **Qwen-WisdomVast** | 57.09 | **70.82** | 51.93 | **13.12** | **43.90** | **26.60** | **44.11** |
41
+
42
+ 说明:
43
+
44
+ 由于官方并未公布Qwen1.5-7B-Chat的评测表现,所以我们自己使用[opencompass](https://github.com/open-compass/opencompass)测试得到以上结果
45
+
46
+ Qwen-WisdomVast使用和Qwen1.5-7B-Chat一样的参数进行测试
47
+
48
+ ## 模型下载
49
+
50
+ | Model | Download |
51
+ |:-------------------:|:-----------:|
52
+ | Qwen1.5-7B |[ 🤗 HuggingFace](https://huggingface.co/Qwen/Qwen1.5-7B) [ 🤖 ModelScope](https://modelscope.cn/models/qwen/Qwen1.5-7B)|
53
+ | Qwen-WisdomVast-Lora |[ 🤗 HuggingFace](https://huggingface.co/zhichen/Qwen-WisdomVast-Lora) [ 🤖 ModelScope](https://modelscope.cn/models/seanzhang/Qwen-WisdomVast-Lora)|
54
+ | Qwen-WisdomVast (合并好的模型) |[ 🤗 HuggingFace](https://huggingface.co/zhichen/Qwen-WisdomVast) [ 🤖 ModelScope](https://modelscope.cn/models/seanzhang/Qwen-WisdomVast)|
55
+
56
+
57
+
58
+ ## 合并LORA模型(可跳过)
59
+
60
+ 1、下载 [Qwen1.5-7B](https://modelscope.cn/models/qwen/Qwen1.5-7B)
61
+
62
+ ```bash
63
+ git clone https://www.modelscope.cn/qwen/Qwen1.5-7B.git
64
+ ```
65
+
66
+ 2、下载[Qwen-WisdomVast-Lora](https://www.modelscope.cn/models/seanzhang/Qwen-WisdomVast-Lora)
67
+
68
+ **From ModelScope**
69
+ ```bash
70
+ git lfs install
71
+ git clone https://www.modelscope.cn/seanzhang/Qwen-WisdomVast-Lora.git
72
+
73
+ ```
74
+
75
+ **From HuggingFace**
76
+ ```bash
77
+ git lfs install
78
+ git clone https://huggingface.co/zhichen/Qwen-WisdomVast-Lora
79
+ ```
80
+
81
+ 3、合并模型
82
+
83
+ ```bash
84
+ python merge_lora.py \
85
+ --base_model path/to/qwen/Qwen1.5-7B \
86
+ --lora_model path/to/lora/Qwen-WisdomVast-Lora \
87
+ --output_dir ./Qwen-WisdomVast
88
+ ```
89
+
90
+ ## 下载 Qwen-WisdomVast(合并好的模型)
91
+
92
+ **From ModelScope**
93
+ ```bash
94
+ git lfs install
95
+ git clone https://www.modelscope.cn/seanzhang/Qwen-WisdomVast.git
96
+
97
+ ```
98
+
99
+ **From HuggingFace**
100
+ ```bash
101
+ git lfs install
102
+ git clone https://huggingface.co/zhichen/Qwen-WisdomVast
103
+ ```
104
+
105
+
106
+ ## 命令行推理
107
+
108
+ ```bash
109
+ python cli_demo.py --model_path ./Qwen-WisdomVast(换成你自己的合并后的模型路径)
110
+ ```
111
+
112
+ ## web 推理
113
+
114
+ ```bash
115
+ python web_demo.py --model_path ./Qwen-WisdomVast(换成你自己的合并后的模型路径)
116
+ ```
117
+
118
+
119
+ ## vllm web 推理
120
+
121
+ 1、使用[vllm](https://github.com/vllm-project/vllm)部署模型
122
+
123
+ ```bash
124
+ python -m vllm.entrypoints.openai.api_server --served-model-name Qwen-WisdomVast --model ./Qwen-WisdomVast(换成你自己的合并后的模型路径)
125
+ ```
126
+
127
+ 2、在命令行执行
128
+
129
+ ```bash
130
+ python vllm_web_demo.py --model Qwen-WisdomVast
131
+ ```
132
+
133
+
134
+ ## 复现测试结果
135
+
136
+ 1、使用[vllm](https://github.com/vllm-project/vllm)部署`openai api server`
137
+
138
+ 部署命令:
139
+
140
+ ```bash
141
+ python -m vllm.entrypoints.openai.api_server --served-model-name Qwen-WisdomVast --model ./Qwen-WisdomVast(换成你自己的合并后的模型路径)
142
+ ```
143
+
144
+ 2、使用[opencompass](https://github.com/open-compass/opencompass)框架进行测试
145
+
146
+ 参考:[使用opencompass验证模型效果](https://blog.csdn.net/qq_44193969/article/details/134979054)
147
+
148
+ 按照以上文章修改好后,将`eval_qwen_wisdomvast.py`文件到 `opencompass/configs`文件���下
149
+
150
+
151
+ 3、执行测试脚本
152
+
153
+ ```bash
154
+ python run.py configs/eval_qwen_wisdomvast.py -w outputs/Qwen-WisdomVast
155
+ ```
156
+
157
+ ## LICENSE
158
+
159
+ 本项目仅可应用于研究目的,项目开发者不承担任何因使用本项目(包含但不限于数据、模型、代码等)导致的危害或损失。详细请参考[免责声明](https://github.com/seanzhang-zhichen/Qwen-WisdomVast/blob/main/DISCLAIMER)。
160
+
161
+ Qwen-WisdomVast项目代码的授权协议为 [The Apache License 2.0](.//LICENSE),代码可免费用做商业用途,模型权重和数据只能用于研究目的。请在产品说明中附加Qwen-WisdomVast的链接和授权协议。
162
+
163
+ ## Citation
164
+
165
+ 如果你在研究中使用了Qwen-WisdomVast,请按如下格式引用:
166
+
167
+ ```latex
168
+ @misc{Qwen-WisdomVast,
169
+ title={Qwen-WisdomVast},
170
+ author={Zhichen Zhang, Weihan Huang},
171
+ year={2024},
172
+ howpublished={\url{https://github.com/seanzhang-zhichen/Qwen-WisdomVast}},
173
+ }
174
+ ```
175
+
176
+
177
+ ## Acknowledgement
178
+
179
+ [QwenLM/Qwen1.5](https://github.com/QwenLM/Qwen1.5)
180
+ <br>
181
+ [hiyouga/LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory)
182
+ <br>
183
+ [shibing624/MedicalGPT](https://github.com/shibing624/MedicalGPT)
184
+ <br>
185
+ [modelscope/swift](https://github.com/modelscope/swift)
186
+
187
+ ## Star History
188
+
189
+ [![Star History Chart](https://api.star-history.com/svg?repos=seanzhang-zhichen/Qwen-WisdomVast&type=Date)](https://star-history.com/#seanzhang-zhichen/Qwen-WisdomVast&Date)