Overview FusionBench is the first comprehensive benchmark dataset designed to evaluate the knowledge mastery and reasoning capabilities of Large Language Models (LLMs) in the domain of nuclear fusion science and engineering. To address the critical lack of standard evaluation resources in this specialized field, FusionBench provides a rigorous, expert-validated testbed covering core scientific concepts, experimental phenomena, and engineering parameters. Dataset Characteristics Scale: The dataset comprises 10,605 question-answer (QA) pairs constructed from a vast corpus of over 140,000 scholarly documents (including arXiv papers, IAEA reports, and specialized textbooks). Coverage: It spans 13 core sub-disciplines of fusion research disciplines and 1 uncategorized subset, ranging from Plasma Instabilities and Transport Processes to Plasma-Facing Materials and Diagnostic Systems. Quality Assurance: Each item has undergone a rigorous multi-stage hybrid validation protocol, combining AI-powered confidence screening with meticulous adjudication by human domain experts (M.S. and Ph.D. candidates in nuclear fusion) to ensure scientific accuracy. Question Types: The benchmark includes multiple-choice, single-choice, and true/false questions to assess different cognitive levels, from knowledge retrieval to complex reasoning. Data Structure The dataset is distributed in a standard JSON format. Each entry contains the following core fields: question_type: Categorical label (multiple_choice, single_choice, or true_false). question: The stem of the question. options: A dictionary of options (keys: "A", "B", "C"..., values: option text). correct_answer: An array containing the keys of the correct options (e.g., ["A"] or ["A", "C"]). category: The specific thematic sub-domain (e.g., "Plasma Instabilities"). (Optional) explanation & source: Contextual information and source snippets for reference. Usage This dataset allows researchers to systematically track LLMs progress, diagnose domain-specific weaknesses, and guide the development of specialized AI tools for fusion energy research. Baseline evaluations of prominent models (including GPT-4 series, Claude, Gemini, and DeepSeek) are provided in the associated research paper. Code Availability The evaluation framework and scripts for reproducing the baselines are available in the attached code archive or on GitHub at: https://github.com/PKU-Xlab/FusionBench.