A benchmark for evaluating content safety detection in image editing. SafeEditBench tests whether vision-language models can correctly identify policy violations in edited images across diverse content safety policies.
SafeEditBench contains image pairs (original + edited) annotated with safety labels and policy violation types.
train
901… See the full description on the dataset page:
https://huggingface.co/datasets/tyodd/SafeEditBench.
