Saint-lsy/EndoBench-Extended

EndoBench是一个全面的MLLM评估框架，涵盖4个胃镜场景和12个临床任务，以及12个二级子任务，这些任务反映了胃镜检查工作流程的进展。该数据集具有五个级别的视觉提示粒度，用于评估区域特定的理解能力。EndoBench包含6,832个经过临床验证的VQA对，这些对是从22个胃镜数据集中派生出来的。数据集的构建涉及收集20个公共和1个私有胃镜数据集，并标准化QA对，形成了包含446,535个VQA对的EndoVQA-Instruct数据集，这是目前最大的胃镜指导调整集合。此外，还提供了EndoBench-extended数据集，这是EndoBench的扩展版本，包括48个图像和91个开放集VQA对，由医学专家精心挑选，以确保临床相关性和诊断难度。

EndoBench is a comprehensive MLLM evaluation framework that spans 4 endoscopy scenarios and 12 clinical tasks, along with 12 secondary subtasks that mirror the progression of endoscopic examination workflows. The dataset features five levels of visual prompting granularity to assess region-specific understanding. EndoBench contains 6,832 clinically validated VQA pairs derived from 22 endoscopy datasets. The dataset construction involves collecting 20 public and 1 private endoscopy datasets and standardizing QA pairs, resulting in the EndoVQA-Instruct dataset with 446,535 VQA pairs, the largest endoscopic instruction-tuning collection currently available. Additionally, the EndoBench-extended dataset is provided, which is an extended version of EndoBench including 48 images and 91 open-set VQA pairs, carefully selected by medical experts to ensure clinical relevance and diagnostic difficulty.