Training data for MJ1 (MultiModal Judge 1) - a multimodal reward model for evaluating vision-language model outputs.
Dataset Summary
MJ1 Training Data is a curated, multi-source preference dataset designed for training a multimodal judge capable of evaluating responses across text and image modalities. Every datapoint contains at least one image and covers three distinct evaluation scenarios:
