MM-Hallu/MMMC

MMMC（多模态模态冲突基准）是一个包含40,000个示例的数据集，旨在测试视觉语言模型处理图像和文本模态之间冲突信息的能力。数据集特征包括：图像（输入图像）、图像ID（图像标识符）、问题（关于图像的问题）、答案（真实答案）、冲突类型（模态冲突的类型）、关键组件（JSON编码的关键组件）、关键组件关系（JSON编码的组件关系）和关键组件属性（JSON编码的组件属性）。该数据集用于多模态任务，特别是视觉问答，涉及冲突、幻觉和基准测试，规模在10K到100K之间，语言为英语。MMMC (Multimodal Modal Conflict Benchmark) is a dataset consisting of 40,000 examples, developed to evaluate the capability of vision-language models to handle conflicting information across image and text modalities. The dataset comprises the following attributes: input images, image IDs (image identifiers), questions (queries about the input images), ground-truth answers, conflict types (types of modal conflicts), key components (JSON-encoded key components), key component relationships (JSON-encoded component relationships), and key component attributes (JSON-encoded component attributes). This dataset is intended for multimodal tasks, especially visual question answering, covering scenarios involving conflict, hallucination and benchmark testing, with a scale falling within the range of 10K to 100K, and all textual content is in English.