MFC-Bench
收藏资源简介:
MFC-Bench是由北京邮电大学、香港浸会大学和香港科技大学联合创建的综合性多模态事实核查基准数据集,包含33,000个样本,用于评估大型视觉-语言模型(LVLMs)的事实准确性。该数据集涵盖了三个主要任务:操纵分类、上下文外分类和真实性分类,旨在检测和纠正多模态内容中的事实错误。创建过程中,数据集从多个来源精心挑选视觉和文本查询,确保对LVLMs在多模态事实核查中的能力进行全面评估。MFC-Bench的应用领域主要集中在提高人工智能在处理复杂视觉和文本元素时的准确性和责任感,以确保在线信息的可信度。
MFC-Bench is a comprehensive multimodal fact-checking benchmark dataset jointly developed by Beijing University of Posts and Telecommunications, Hong Kong Baptist University, and The Hong Kong University of Science and Technology. It consists of 33,000 samples, specifically designed to assess the factual accuracy of Large Vision-Language Models (LVLMs). This dataset encompasses three core tasks: manipulation classification, out-of-context classification, and authenticity classification, with the goal of detecting and rectifying factual errors in multimodal content. During its development, visual and textual queries were meticulously selected from multiple sources to ensure a comprehensive evaluation of LVLMs' capabilities in multimodal fact-checking. The primary application domains of MFC-Bench focus on enhancing the accuracy and accountability of artificial intelligence systems when processing complex visual and textual elements, thereby safeguarding the credibility of online information.




