Description
SCIEVAL is an evaluation benchmark for large language models in the scientific domain. It consists of approximately 18,000 objective evaluation questions and few subjective questions, covering the fundamental scientific fields of chemistry, physics, and biology. This benchmark assesses the understanding and generation capabilities of large language models in scientific content from four aspects: basic knowledge, knowledge application, scientific calculation, and research ability.
Files Description
- scieval-dev.json is the dev set, containing 5 samples for each task name, each ability and each category, which is specially used for few shot.
- scieval-valid.json is the valid set, containing the answer for each question.
- scieval-test.json is the test set.
- make_few_shot.py is the code for generating the few shot data, you can modify it as you need.
- eval.py is the evaluation code for the valid set, which is the same as the one we used for the test set. Note the the prediction should follow the format:
[{
"id": "5534a4ef45aea8a6f1835750a54c01d0",
"pred": "C",
}]
- dynamic_chem.json and dynamic_phy.json is the dynamic data, which is a re-generated version and is different from the data we used in the leaderboard. We will update it regularly.
- eval_dynamic.py is the evalution code for the dynamic data. To use this script, you need to add "pred" key directly to the original dynamic data.
Reference
If you use any source codes or datasets included in this repository in your work, please cite the corresponding papers. The bibtex are listed below:
@article{sun2023scieval,
title={SciEval: A Multi-Level Large Language Model Evaluation Benchmark for Scientific Research},
author={Sun, Liangtai and Han, Yang and Zhao, Zihan and Ma, Da and Shen, Zhennan and Chen, Baocai and Chen, Lu and Yu, Kai},
journal={arXiv preprint arXiv:2308.13149},
year={2023}
}