TruthfulQA

TruthfulQA 是衡量语言模型在生成问题答案时是否真实的基准。该基准包括 817 个问题，涵盖 38 个类别，包括健康、法律、金融和政治。作者精心设计了一些人会因为错误的信念或误解而错误地回答的问题。

TruthfulQA is a benchmark for assessing whether language models generate truthful answers to questions. It comprises 817 questions across 38 categories, including health, law, finance, and politics. The authors have carefully designed questions that people are prone to answer incorrectly due to false beliefs or misunderstandings.

搜集汇总

背景与挑战

背景概述

TruthfulQA是一个公开的多模态预训练数据集，专注于评测视觉问答任务，采用Apache 2.0许可证。数据集规模较小，包含8个文件，总大小为16.4MB，在社区中较受欢迎，拥有3.6k星和126个复刻。