license: apache-2.0
language:
- zh
Welcome to AcademicBrowse created by PKU-DS-LAB!
Dataset Description
AcademicBrowse is the first dataset specifically designed to evaluate the complex information retrieval capabilities of Large Language Models (LLMs) in academic research.
Key characteristics of AcademicBrowse include:
- Academic Practicality: Questions are based on real academic learning and research environments, avoiding misleading the models.
- High Difficulty: Answers often require at least three deep searches to derive, making them challenging for single models.
- Concise Evaluation: Answers are unique, with clear sources and brief explanations, facilitating audit and verification.
- Broad Coverage: The dataset spans at least 12 different academic disciplines, including Computer Science, Literature, Biology, Political Science, Economics, Mathematics, Demography, History of Science and Technology, Chemistry, Sociology, Public Health, and Physics.
The dataset consists of 223 meticulously curated questions in Chinese, each accompanied by an answer, explanation, and domain. It was created by a team of undergraduate and graduate students from various faculties at Peking University, ensuring the questions reflect genuine academic search scenarios.
Dataset Structure
Each entry in the dataset contains the following fields:
- question: The academic query or problem.
- answer: The correct answer to the question.
- explanation: A brief explanation or justification for the answer, including sources.
- domain: The academic discipline or field to which the question belongs.
The dataset is provided as a JSON file containing a list of entries.
Experiment Result
| Model | All (%) | Science & Engineering (%) | Social Sciences & Humanities (%) |
|---|---|---|---|
| gpt-4o-search-preview | 18.83 | 18.64 | 19.05 |
| gpt-4o-mini-search-preview | 10.31 | 10.17 | 10.48 |
| deepseek-r1-0528 | 8.52 | 5.08 | 12.38 |
| gpt-4.1 | 7.17 | 5.93 | 8.57 |
| gpt-4o-2024-11-20 | 3.59 | 1.69 | 5.71 |
| gpt-4o-mini | 2.24 | 0.85 | 3.81 |
The judge model for all experiments is GPT-4o-mini.
Citation Information
This paper will soon be published on arXiv for open access.
Additional Information
- This project was funded by Grant 624B2005.
- We would like to thank the following individuals for their contributions to problem-solving and evaluation: Xun Zhao, Zizhuo Fu, Yuqian Zhan, Xinhao Ji, Jiarui Sun, Junhao Zhang, Shengfan Wang, Ziteng Lu, Yumeng Song, Ziyan Yang, Hongjiao Wang, Shan Zhang, Huahui Lin, Junhong Liu, Zhengyang Wang, Yuchen Lu, Yanxi Xu.
Team Members:
Leading By:
Tong Yang; Yuhan Wu;
Core Contributors:
Junting Zhou; Wang Li; Yiyan Liao; Nengyuan Zhang; Tingjia Miao; Zhihui Qi
Dataset Card Contact
For more details, please contact: [email protected]