Datasets:
license: cc-by-nc-4.0
task_categories:
- visual-question-answering
- question-answering
language:
- en
- zh
tags:
- science
- multimodal
- benchmark
- evaluation
- astronomy
- earth-science
- biology
- materials-science
pretty_name: SFE (Scientists First Exam) Multimodal Science Benchmark
size_categories:
- n<1K
SFE (Scientists First Exam) Multimodal Science Benchmark
SFE 科学家第一考试多模态评测数据集
Dataset Description
- Homepage: obaydata.com
- Organization: New Oriental Bay Limited (obaydata)
- Contact: simon.su@obaydata.com
Description
SFE (Scientists First Exam) is a multimodal scientific LLM evaluation benchmark. It uses original research data combined with expert annotations to construct high-difficulty scientific reasoning datasets, evaluating models' full-chain research capabilities from signal perception and attribute understanding to complex reasoning.
Domains: Astronomy, Earth Science, Life Science, Materials Science, and more.
Data Production: Domain experts combine the latest research results to annotate high-difficulty multimodal image-text questions with answer explanations.
Difficulty Standard: Data is produced to the pass@4 standard where the model answers correctly at most twice out of four attempts.
中文描述
SFE（Scientists First Exam）是多模态科学大模型评测基准，核心是用原始科研数据 + 专家标注构建高难度科学推理数据集，评估模型从信号感知、属性理解到复杂推理的全链条科研能力。
由专家结合最新研究成果在天文学、地球科学、生命科学、材料科学等领域进行标注，生产出具备高难度的多模态图文题目并出具答案解析。生产数据按照pass@4标准模型至多答对两次。
Dataset Details
| Property | Value |
|---|---|
| Scale | Weekly production capacity: 300 entries |
| Format | JSON with images |
| License | CC-BY-NC-4.0 |
| Copyright | Full copyright provided |
| Use Case | Evaluating multimodal LLMs on expert-level scientific reasoning across multiple disciplines |
Citation
If you use this dataset, please cite:
@dataset{obaydata2026,
author = {OBayData Team},
title = {SFE (Scientists First Exam) Multimodal Science Benchmark},
year = {2026},
publisher = {HuggingFace},
url = {https://huggingface.co/datasets/obaydata/sfe-scientists-first-exam}
}