TopoMAS Evaluation Datasets and Results for Materials Science Question-Answering System

作者：Li, Xin · 发表于：Zenodo (CERN European Organization for Nuclear Research) · 年份：2026 · DOI：10.5281/zenodo.17657657 · 研究领域：Computer science、Data science、Data mining、Machine learning、Information retrieval、Management science、Artificial intelligence

This repository contains the evaluation datasets and experimental results for TopoMAS, a knowledge-enhanced question-answering framework for topological materials science. The specific contents include: 📊 Evaluation Datasets: - LLM4Mat-Bench Benchmark Dataset: Contains 85790 questions with standard answers - TopoQA Specialized Dataset: Covers 110 domain-specific questions and answers in materials science (developed by our research team and continuously updated) - TopoOQ Specialized Dataset: Includes 95 open-ended questions in materials science (developed by our research team and continuously updated) 📈 Experimental Results: - Performance metrics on LLM4Mat-Bench (Wtd. Avg. (MAD:MAE) = 14.421, Wtd. Avg. AUC = 0.891) - Performance metrics on TopoQA (Accuracy: 91.21%±5.55%) - Performance metrics on TopoOQ (Composite Score: 8.75±0.01/10) - Comparative results with multiple models in the framework (Qwen3, Qwen2.5, DeepSeek-V3) This dataset supports the research presented in the paper "TopoMAS: Large Language Model Driven Topological Materials Multi-Agent System"and can be used to reproduce experimental findings or serve as benchmark data for materials science QA systems. We sincerely welcome contributions and suggestions from the academic community to further improve dataset quality.