A Hallucination Detection Dataset for Large Language Models Based on Four Chinese Classical Novels

It is constructed based on four representative Chinese classical novels: Dream of the Red Chamber, Journey to the West, Romance of the Three Kingdoms, and Water Margin.The dataset is derived from the original texts of the four classical novels and constructed through a series of preprocessing procedures, including data cleaning, text filtering, and hallucination sample generation.

It consists of both non-hallucinated samples and hallucinated samples. Specifically, non-hallucinated samples preserve factual information consistent with the original literary texts, representing correct factual expressions in model-generated content. In contrast, hallucinated samples are constructed by introducing information inconsistent with the original texts, factual errors, or content that conflicts with the original narratives, thereby simulating factual errors that may occur during the generation process of large language models.