A visual multiple-choice benchmark for evaluating social bias and reasoning in vision-language models.
Melange Visual Bias Benchmark is a multimodal extension of the BBQ (Bias Benchmark for Question Answering) dataset, designed to probe social bias and fairness in VLMs (Vision-Language Models). Instead of relying on textual context, this dataset grounds each multiple-choice question in one or more scene images that depict the… See the full description on the dataset page:
https://huggingface.co/datasets/IDfree/melange_visual_bbq_new.
