MSQA
收藏资源简介:
MSQA是一个用于评估大型语言模型(LLMs)在材料科学领域的知识推理和复杂推理能力的综合评估基准。该数据集包含1757个研究生级别的材料科学问题,分为详细解释性回答和二元真/假评估两种格式。MSQA通过要求LLMs在七个材料科学子领域中同时具备精确的事实知识和多步推理能力来挑战它们。通过使用10种最先进的LLMs进行实验,我们发现当前LLMs在性能上存在显著差距。MSQA代表了第一个联合评估LLMs的事实和推理能力的基准,这对于高级材料科学中的LLMs至关重要。
MSQA is a comprehensive evaluation benchmark for assessing the knowledge reasoning and complex reasoning capabilities of large language models (LLMs) in the field of materials science. This dataset contains 1757 graduate-level materials science questions, divided into two formats: detailed explanatory answers and binary true/false assessments. MSQA challenges LLMs by requiring them to possess both precise factual knowledge and multi-step reasoning capabilities across seven subfields of materials science. Through experiments conducted with 10 state-of-the-art LLMs, we found that there is a significant performance gap among current LLMs. MSQA represents the first benchmark that jointly evaluates the factual and reasoning capabilities of LLMs, which is critical for LLMs in advanced materials science.
MSQA数据集概述
数据集基本信息
- 名称:MSQA
- 数据量:1757个问答对
- 存储文件:
data/MSQA_Dataset.json - 关联论文:MSQA: Benchmarking LLMs on Graduate-Level Materials Science Reasoning and Knowledge
数据结构
每个条目包含以下字段:
question:长形式问题answer:长形式答案topic:GPT-4o生成的论文摘要总结,包含:- 论文目的
- 目的分类(方法或结果)
- 候选问题
source:原始论文DOIsource_section:生成答案的原始论文章节question_type:长形式问题的类型(共7种)true_false_question:二元形式问题true_false_question_answer:二元问题答案(YES/NO)
数据集特点
- 领域:材料科学
- 问题级别:研究生水平
- 问题类型:包含长形式问答和二元问答
- 数据来源:学术论文
处理流程
- 生成流程:
- 候选答案生成(使用GPT-4o、Gemini-2.0-pro、Deepseek v3)
- 候选答案合并
- 二元问题生成
- 评估流程:
- 基线模型推理
- 使用GPT-4o或基于规则的方法进行评估




