TIGER-Lab/MEGA-Bench

MEGA-Bench是一个大规模的多模态评估数据集，包含超过500个现实世界任务，覆盖了丰富的多模态任务类型，支持多种输出格式，并提供了40多种评估指标。数据集分为Core和Open两个主要子集，以及针对没有多图像支持模型的Core Single-image和Open Single-image两个单图像子集。

MEGA-Bench is a large-scale multimodal evaluation dataset containing over 500 real-world tasks, covering a rich set of multimodal task types, supporting various output formats, and providing over 40 evaluation metrics. The dataset is divided into two main subsets, Core and Open, as well as two single-image subsets, Core Single-image and Open Single-image, for models without multi-image support.