Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/llmeval/llmeval-2
中文大语言模型评测第二期
https://github.com/llmeval/llmeval-2
Last synced: about 1 month ago
JSON representation
中文大语言模型评测第二期
- Host: GitHub
- URL: https://github.com/llmeval/llmeval-2
- Owner: llmeval
- Created: 2023-07-06T09:55:55.000Z (over 1 year ago)
- Default Branch: master
- Last Pushed: 2023-10-23T07:57:31.000Z (about 1 year ago)
- Last Synced: 2024-08-03T09:06:48.261Z (5 months ago)
- Homepage:
- Size: 8.78 MB
- Stars: 68
- Watchers: 4
- Forks: 3
- Open Issues: 0
-
Metadata Files:
- Readme: README-en.md
Awesome Lists containing this project
- StarryDivineSky - llmeval/llmeval-2
README
# LLMEVAL-2 Chinese large Language Model evaluation II
(TODO: abstract and Axive link)
## Introduction
LLMEVAL-1 has come to an end, with a large number of public users participating, providing us with detailed data for our evaluation. We also provided a detailed evaluation report after the data collection phase ended (https://github.com/llmeval/llmeval-1).
However, we also discovered some problems. Because this evaluation hopes to involve more of the public, we chose to design questions that the vast majority of users can evaluate. These kinds of questions do not fully reflect the knowledge coverage of the model.
Therefore, we launched LLMEVAL-2, a professional field evaluation. LLMEVAL-2 evaluation is mainly based on users' daily use, combined with the distribution of online user questions, focusing on the tasks that different undergraduate and graduate students hope to get help from the large model in their daily study and life.
## Dataset
We have constructed domain knowledge test sets for 12 subjects respectively. The question setters, based on external databases, construct corresponding test question sets for each subject area. The types of questions are multiple-choice and Q&A. Among them, about 25-30 objective questions and about 10-15 subjective questions are designed for each subject, totaling 12*40=480 questions.
Example of objective question and answer content:
```text
"question content": "请回答工业制取氯气的原理方程式"
"answer content": "工业上通常采用氯化氢气体溶于水的方法制备氯气。具体反应式为:HCl + NaOH → NaCl + H2O + Cl-,其中NaOH为碱溶液，H+和Cl-离子在溶液中发生反应生成氯气和水。"
```
Example of subjective question and answer content:
```
"question content": "我是一名大学生，请问我该如何学好有机化学？请为我给出建议。"
"answer content": "作为一名大学生学习有机化学，以下是一些建议可以帮助你更好地掌握这门学科: 1. 建立坚实的基础:在学习有机化学之前，你需要先打好化学基础知识，如物理、数学等。这样才能更好地理解有机化学中的各种概念和原理。 2. 实践操作:通过实验来加深对有机化学的理解，提高实验技能。可以选择一些简单的实验项目开始，逐渐提高自己的难度和复杂程度。 3. 阅读相关书籍和论文:阅读有机化学领域的经典著作和相关文献，了解最新的研究进展和技术发展趋势。 4. 参加课外课程和活动:参加学校组织的有机化学相关的课外活动，如讲座、研讨会等，与其他学生交流经验，扩展自己的知识面。 5. 寻求专业指导:向老师和同学请教问题，寻求帮助解决问题，及时反馈学习进度，并制定合理的学习计划和目标。"
```
## **Criteria**
| Question type | Criteria | Points | Description |
| :----------------------------------------------------------- | :---------------------- | :----- | :---------------------------------------------------- |
| Objective questions
(such as multiple-choice or fill-in-the-blank questions) | Correctness | 3' | Whether the answer is correct or not |
| | Explanation Correctness | 2' | Whether the explanation contains any errors |
| Subjective questions (such as essay questions) | Accuracy | 5' | Whether the answer is correct or not |
| | Informativeness | 3' | Whether the answer information provided is sufficient |
| | Fluency | 3' | Whether the syntax of the answer is correct |
| | Logical Coherence | 3' | Whether the answer is logically rigorous |
## Ranking
A total of 20 large language models were evaluated in this review and the results are as follows:
| | Objective questions | | Subjective questions | | | | Ranking | Scores |
| ----------------- | ------------------- | --------------------------- | -------------------- | ------------- | -------------- | ------------------- | ------- | -------------- |
| **LLM** | **Correctness** | **Explanation Correctness** | **Fluency** | **Accuracy** | **Logicality** | **Informativeness** | | |
| GPT4 | 2.378 (2.395) | 1.670 (1.595) | 2.895 (2.989) | 4.260 (4.545) | 2.779 (2.903) | 2.691 (2.886) | 1(1) | 86.72 (89.54) |
| GPT3.5 | 2.160 (2.138) | 1.542 (1.503) | 2.861 (3.000) | 3.822 (4.295) | 2.694 (2.818) | 2.489 (2.750) | 2(2) | 80.71 (84.69) |
| Xunfei-Xinghuo | 2.114 (2.243) | 1.557 (1.632) | 2.815 (2.977) | 3.750 (4.193) | 2.560 (2.739) | 2.196 (2.716) | 3(5) | 78.05 (82.26) |
| Baichuan-13B-Chat | 2.003 (2.013) | 1.428 (1.441) | 2.847 (2.949) | 3.727 (4.102) | 2.631 (2.778) | 2.472 (2.756) | 4(6) | 77.51 (81.82) |
| MiniMax-abab5 | 1.922 (1.928) | 1.443 (1.493) | 2.878 (2.989) | 3.800 (3.977) | 2.656 (2.722) | 2.478 (2.699) | 5(7) | 77.47 (80.64) |
| NewBing | 2.197 (2.211) | 1.583 (1.615) | 2.796 (2.989) | 3.608 (3.875) | 2.558 (2.773) | 2.061 (2.511) | 6(4) | 77.28 (82.63) |
| Claude | 1.923 (2.066) | 1.463 (1.576) | 2.680 (2.977) | 3.597 (4.125) | 2.613 (2.801) | 2.414 (2.710) | 7(3) | 75.57 (83.49) |
| moss-mars | 1.961 (1.967) | 1.465 (1.470) | 2.737 (3.000) | 3.480 (3.807) | 2.508 (2.648) | 2.229 (2.534) | 8(9) | 74.41 (79.21) |
| Kunlun-Tiangong | 1.933 (1.961) | 1.354 (1.500) | 2.774 (2.983) | 3.520 (3.807) | 2.576 (2.682) | 2.339 (2.523) | 9(8) | 74.36 (79.31) |
| Ziya-LLaMA-13B-v1 | 1.681 (1.592) | 1.306 (1.201) | 2.804 (3.000) | 3.207 (3.364) | 2.473 (2.585) | 2.120 (2.278) | 10(13) | 69.48 (70.92) |
| Ali-Tongyiqianwen | 1.638 (1.618) | 1.275 (1.280) | 2.776 (3.000) | 3.098 (3.239) | 2.443 (2.511) | 2.126 (2.335) | 11(12) | 68.01 (71.02) |
| 360 | 1.720 (1.678) | 1.322 (1.352) | 2.700 (2.989) | 3.022 (3.352) | 2.394 (2.608) | 2.056 (2.313) | 12(10) | 67.97 (72.86) |
| CIIC-Zhigong | 1.680 (2.072) | 1.297 (1.516) | 2.764 (2.983) | 3.067 (4.080) | 2.427 (2.744) | 1.916 (2.631) | 13(14) | 67.27 (70.53) |
| ChatGLM2-6B | 1.690 (1.671) | 1.345 (1.306) | 2.758 (2.920) | 2.934 (3.011) | 2.401 (2.386) | 1.956 (2.210) | 14(17) | 67.07 (69.06) |
| Vicuna-33B | 1.567 (1.684) | 1.277 (1.270) | 2.599 (2.943) | 3.033 (3.080) | 2.440 (2.398) | 2.143 (2.199) | 15(16) | 66.53 (69.16) |
| InternLM-7B | 1.655 (1.658) | 1.355 (1.174) | 2.636 (2.847) | 3.091 (3.330) | 2.295 (2.392) | 1.938 (2.233) | 16(18) | 66.52 (69.00) |
| ChatGLM-130B | 1.602 (1.638) | 1.239 (1.280) | 2.670 (2.926) | 3.022 (3.114) | 2.374 (2.443) | 2.084 (2.278) | 17(15) | 66.05 (69.48) |
| TigerBot-180B | 1.604 (1.592) | 1.294 (1.220) | 2.573 (2.926) | 3.079 (3.557) | 2.489 (2.602) | 1.882 (2.352) | 18(11) | 65.90 (71.77) |
| AquilaChat-7b | 1.548 (1.553) | 1.239 (1.207) | 2.710 (2.932) | 2.945 (3.136) | 2.383 (2.443) | 1.918 (2.244) | 19(19) | 64.82 (68.19) |
| BELLE-7B-2M | 1.484 (1.461) | 1.224 (1.164) | 2.685 (2.824) | 2.695 (3.000) | 2.347 (2.335) | 1.880 (2.131) | 20(20) | 62.98 (65.27) |
\* Note: Values in brackets are GPT-4 automatic evaluation scores and rankings.
Total score calculation formula is
$$
\text{score}=\frac 1 n \sum_{i=1}^n score_i,
$$
where
$$
score_i=\frac{score_{ai}}{score_{fi}}\times 100.
$$
$score_{ai}$ is the **actual score** of the model in the $i$-th subject, and $score_{fi}$ is the **total score** of the $i$-th subject.
## Report
link:
https://github.com/llmeval/llmeval-2/blob/master/LLMEVAL-2.pdf
## Citation
(TODO)