lmm-r1

Why we included this project

OpenRLHF does not natively cover multimodal RL training, so teams wanting DeepSeek-R1-style rule-based RL on vision-language models usually have to assemble their own pipeline. LMM-R1 closes that gap by adapting OpenRLHF so 3B LMMs can be trained with a two-stage scheme, text-only reasoning first and multimodal generalization second, and ships the VerMulti dataset with training scripts to reproduce the reported geometry and perception-reasoning gains on local hardware. Much of the code has been upstreamed into OpenRLHF-M, so groups already on that stack can adopt the approach without maintaining a fork. For researchers and engineers applying rule-based RL to custom VLMs, it provides working data and configs, not just a paper.