DisentQA, TriviaQA

本研究构建了DisentQA和TriviaQA两个数据集，用于检测和干预大型语言模型中的幻觉问题。DisentQA用于开放式问答场景，TriviaQA用于封闭式问答场景。数据集包含至少1000个标记为‘幻觉’和1000个标记为‘真实’的示例。这些数据集通过自动方法区分不同类型的知识情况，从而创建训练、验证和测试集，以帮助理解和减少语言模型中的幻觉现象。

This study constructs two datasets, DisentQA and TriviaQA, for detecting and mitigating hallucinations in large language models (LLMs). DisentQA is designed for open-domain question answering scenarios, while TriviaQA targets closed-domain question answering scenarios. The datasets contain at least 1,000 examples labeled as 'hallucinatory' and 1,000 labeled as 'factual'. These datasets utilize automated methods to distinguish different types of knowledge contexts, thereby creating training, validation, and test sets to facilitate the understanding and mitigation of hallucinations in language models.

搜集汇总

背景与挑战

背景概述

DisentQA和TriviaQA是两个专为检测和干预大型语言模型幻觉问题而构建的数据集，分别针对开放式和封闭式问答场景。每个数据集包含至少1000个标记为'幻觉'和1000个标记为'真实'的示例，通过自动方法构建训练、验证和测试集，以支持减少模型幻觉的研究。