How AI describes MM-RLHF

MM RLHF is a comprehensive project for aligning Multimodal Large Language Models (MLLMs) with human preferences, including a high-quality 120K sample dataset, a state-of-the-art critique-based reward model, and a novel alignment algorithm called MM DPO. The project provides two new benchmarks for reward modeling and multimodal safety, and demonstrates broad performance gains across 10 dimensions and 27 benchmarks for open-source MLLMs.