2026-10-10T17:11:31Z https://www.scidb.cn/oai

10.57760/sciencedb.41072 2026-06-25T11:06:47Z

2026-06-25 WushuQA: A Chinese Martial Arts Question-Answering Dataset for Large Language Model Fine-Tuning doi:10.57760/sciencedb.41072 en WushuQA is a question-answering dataset designed for instruction fine-tuning of large language models (LLMs) in the domain of Chinese martial arts (武术/Wushu).Background & Motivation: Chinese martial arts feature a vast and intricate knowledge system with highly specialized terminology and complex lineage structures. General-purpose LLMs consistently perform poorly in this domain — confusing styles, misattributing techniques, and giving vague or incorrect explanations of core concepts. The root cause is a near-total absence of high-quality martial arts text in standard training corpora.Scale: Built upon 1,056 source texts (~2.8 million Chinese characters) drawn from six authoritative sources — including the Wushu Management Center of the General Administration of Sport of China, the Chinese Wushu Association, the International Wushu Federation, Wikipedia, and Baidu Baike — plus one specialized martial arts website, the dataset yields 14,380 high-quality QA pairs in JSONL format. Each record retains the original source text for full traceability.Construction Pipeline: The dataset was built through an eight-stage pipeline: rule-based cleaning → structured segmentation → multi-level LLM question generation → semantic deduplication (embedding similarity threshold 0.92) → multi-model answer generation → multi-model cross-review → full-scale faithfulness scanning → question context completion. Starting from 15,565 initially generated questions, 14,380 passed all quality filters.Coverage: Five knowledge subdomains — martial arts styles, techniques & theory, historical figures, competition rules, and cultural heritage — across six question types: factual, explanatory, comparative, enumerative, judgmental, and comprehensive. WushuQA; Chinese martial arts; large language models wu cheng xue PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank