PhD
收藏资源简介:
PhD数据集是由中国人民大学DEKE重点实验室和腾讯机器学习平台部联合开发的,专注于评估大型视觉-语言模型(LVLMs)中的内在视觉-语言幻觉(IVL-Hallu)问题。该数据集包含超过7000张图像和53976个问题,涵盖四种IVL-Hallu类型:物体幻觉、属性幻觉、多模态冲突幻觉和反常识幻觉。数据集通过自动化流程生成,确保数据的高质量和多样性,旨在帮助研究人员更好地理解和解决LVLMs中的幻觉问题,推动未来在IVL-Hallu和LVLMs领域的研究。
PhD Dataset is co-developed by the DEKE Key Laboratory of Renmin University of China and Tencent Machine Learning Platform Department, focusing on evaluating the intrinsic vision-language hallucination (IVL-Hallu) issues in large vision-language models (LVLMs). This dataset contains over 7,000 images and 53,976 questions, covering four types of IVL-Hallu: object hallucination, attribute hallucination, multimodal conflict hallucination, and counter-intuitive hallucination. Generated via automated pipelines, the dataset ensures high data quality and diversity, aiming to help researchers better understand and address the hallucination problems in LVLMs, and promote future research in the fields of IVL-Hallu and LVLMs.



