MM-Hallu/MMHalSnowball
收藏官方服务:
资源简介:
MMHalSnowball数据集用于评估大型视觉语言模型中的多模态幻觉滚雪球现象。它研究先前生成的幻觉是否会误导模型在后续查询中做出错误声明,即使有可用的视觉基础信息。该基准使用GQA/Visual Genome图像,并包含跨多种对话设置的精心策划的幻觉对话。
MMHalSnowball evaluates multimodal hallucination snowballing in Large Vision-Language Models (LVLMs). It investigates whether previously generated hallucinations can mislead LVLMs into making incorrect claims in subsequent queries, even when ground visual information is available. The benchmark uses GQA/Visual Genome images with curated hallucinatory conversations across multiple conversation settings.
提供机构:
MM-Hallu


