BEAF is a benchmark for evaluating object hallucination in vision-language models using before-after image manipulation pairs. 26,064 QA pairs over 2,223 images (500 original COCO images + 1,723 manipulated images) with POPE-style yes/no questions.

	Fields

Field Description

image The image (original COCO or manipulated)

question POPE-style question: "Is there a/an {object} in the image?"

gt Ground truth… See the full description on the dataset page: https://huggingface.co/datasets/MM-Hallu/BEAF.