A closed-book benchmark probing large language models for historical-fact hallucination and source attribution over the pre-Qin → Han–Wei–Three-Kingdoms Chinese classics. 400 questions, answer keys 100% derived from the primary sources, adversarially validated. Reputation asset, not a commercial product.

一句话发现：前沿大模型仍会把史事安到纪年上根本不可能的典籍里（例如把春秋事件说成见于《资治通鉴》——通鉴起于公元前 403 年）。横评报告见 REPORT.md。

数据集内容

400 道闭卷题（题面不含原文，测模型的参数化历史知识 /… See the full description on the dataset page: https://huggingface.co/datasets/lizhuojun/chhallu-src-v1.