PHD (Passage-level Hallucination Detection)

PHD数据集是由北京大学计算机技术研究所等机构创建，用于评估段落级幻觉检测方法的高质量基准。该数据集通过ChatGPT生成，包含300个样本，分为低、中、高三个数据量级，反映了模型在不同知识量级下的幻觉生成情况。数据集的创建过程涉及从维基百科提取实体，使用ChatGPT生成幻觉数据，并通过严格的人工标注流程确保数据质量。PHD数据集主要应用于自然语言处理领域，旨在解决大型语言模型在生成文本时可能产生的幻觉问题，提高模型在关键任务中的可靠性和准确性。

The PHD dataset is a high-quality benchmark developed by the Institute of Computer Technology at Peking University and other institutions, designed for evaluating paragraph-level hallucination detection methods. Generated using ChatGPT, this dataset comprises 300 samples categorized into three tiers: low, medium, and high, which reflect the hallucination generation performance of models across varying knowledge scales. The dataset's construction workflow involves extracting entities from Wikipedia, generating hallucinatory text data with ChatGPT, and employing a rigorous manual annotation process to ensure data quality. Primarily utilized in the field of natural language processing, the PHD dataset aims to address the hallucination issues that large language models (LLMs) may produce during text generation, thereby improving the reliability and accuracy of models in key tasks.