SHALE
收藏资源简介:
SHALE是一个针对大型视觉语言模型(LVLMs)细粒度幻觉评估的可扩展基准数据集。该数据集由中国科学院计算技术研究所的研究团队构建,旨在评估LVLMs在忠实度和事实性幻觉方面的表现。数据集包含超过3万个图像-指令对,涵盖了12个代表性的视觉感知方面和6个主要的知识领域,同时考虑了干净和噪声场景。该数据集是通过自动化的数据构建流程生成的,具有可扩展性、可控性和多样性。数据集的构建过程包括提示设计、图像生成和指令生成三个阶段,并采用了层次化的幻觉诱导框架,通过图像级别、指令级别和组合级别的输入扰动来模拟现实中的噪声场景。SHALE数据集的构建旨在解决现有基准数据集在规模、数据泄露和细粒度评估方面的局限性,为LVLMs的稳定性和幻觉抵抗能力提供全面和精细的评估。
SHALE is a scalable benchmark dataset for fine-grained hallucination evaluation of Large Vision-Language Models (LVLMs). The dataset was constructed by the research team from the Institute of Computing Technology, Chinese Academy of Sciences, aiming to evaluate the performance of LVLMs in terms of faithfulness and factual hallucinations. It contains over 30,000 image-instruction pairs, covering 12 representative visual perception aspects and 6 major knowledge domains, while considering both clean and noisy scenarios. Generated through an automated data construction pipeline, the dataset features scalability, controllability, and diversity. Its construction process consists of three stages: prompt design, image generation, and instruction generation, and it adopts a hierarchical hallucination induction framework that simulates real-world noisy scenarios via input perturbations at the image-level, instruction-level, and compositional-level. The development of the SHALE dataset aims to address the limitations of existing benchmark datasets in terms of scale, data leakage, and fine-grained evaluation, providing comprehensive and meticulous assessments for the stability and hallucination resistance of LVLMs.




