PKU-Alignment/TruthfulVQA

TruthfulVQA是一个用于评估视觉问答模型真实性和可靠性的数据集。它包含多个子集，每个子集都专注于视觉问答中真实性的不同方面，如识别伪造特征、检测模仿性错误、识别伪造信息、发现隐藏信息、理解视角限制以及处理视觉混淆场景。数据集以多种难度级别的问题构成，每个问题都附带一个相关图像、多个选择答案和一个正确答案。

TruthfulVQA is a visual question answering dataset designed to evaluate the truthfulness and reliability of vision-language models. It consists of multiple subsets, each focusing on different aspects of truthfulness in visual question answering such as identifying fabricated features, detecting imitative falsehoods, recognizing forged information, uncovering hidden information, understanding perspective limitations, and handling visually confusing scenarios. The dataset is composed of questions at various difficulty levels, each associated with an image, multiple choice options, and a correct answer.