OpenMOSS/DiRL: Post-training framework for Diffusion Language Models
OpenMOSS researchers have introduced DiRL, an open-source post-training framework for Diffusion Language Models that integrates Supervised Fine-Tuning and Reinforcement Learning stages. Using this framework, the team trained DiRL-8B-Instruct, an 8B-scale model featuring a novel DiPO algorithm and Flex-Attention support that achieves state-of-the-art results on mathematical reasoning benchmarks, outperforming several larger 32B models. Open article →