MM-Hallu/VALOR-EVAL

VALOR-EVAL是一个用于评估大型视觉语言模型全面覆盖性和忠实性的多维基准数据集。它涵盖了对象、属性和关系等多个维度，并基于关联偏差选择了具有挑战性的图像。数据集包含多种任务类型，如对象存在性、属性对象、属性人物、关系比较和关系位置等。每个任务都有对应的图像来源（GQA或Pexels）和注释格式。数据集的字段包括图像、图像ID、任务类型和注释。

VALOR-EVAL is a multi-dimensional benchmark for holistic coverage and faithfulness evaluation of large vision-language models. It covers objects, attributes, and relations with challenging images selected based on associative bias. The dataset includes various task types such as object_existence, attribute_object, attribute_people, relation_comparative, and relation_positional. Each task has corresponding image sources (GQA or Pexels) and annotation formats. The dataset fields include image, image_id, task_type, and annotation.