EasyR1

Why we included this project

Training a model with reinforcement learning rather than plain supervised fine-tuning usually means gluing together a rollout engine, a trainer, and a custom data loop. EasyR1 bundles those pieces on top of the veRL engine, supporting GRPO plus newer methods like DAPO and Reinforce++, and it handles vision-language models such as Qwen2.5-VL alongside text-only Llama and Qwen checkpoints. Hybrid-engine scheduling and vLLM's SPMD mode keep multi-GPU runs fast, which matters for long RL rollouts. For teams trying to reproduce DeepSeek-R1-style reasoning training, the included algorithms and a clear custom dataset format remove much of the hand-wiring. Prebuilt Docker images and LoRA support make it usable on anything from a single research box to a larger cluster.