RefoMB

The RefoMB dataset is part of a project called RLAIF-V, which stands for "Aligning MLLMs through Open-Source AI Feedback for Super GPT-4V Trustworthiness." It's an open-source multimodal preference dataset that contains more than 30,000 high-quality comparison pairs. The dataset is designed to reduce hallucination in different Multimodal Large Language Models (MLLMs) and improve their trustworthiness by providing high-quality feedback data and an online feedback learning algorithm.

The RefoMB itself is a comprehensive multimodal evaluation set that covers 8 sub-abilities of multimodal model perception and reasoning. It includes a variety of image types such as cartoons, text-rich images, and photos to assess the credibility of responses and general performance of current multimodal models in open-ended generation tasks.