veRL (HybridFlow)

The dominant open RL post-training framework for LLMs (22.5K★), born as the HybridFlow paper (EuroSys 2025): a hybrid programming model that decouples single-controller orchestration from multi-controller computation, making PPO/GRPO-style RLHF pipelines flexible and efficient at scale. It became the default infrastructure for the open RL-for-reasoning wave (used by SimpleRL, DAPO, and many R1-replication efforts).

Joint HKU / ByteDance Seed co-development that ships under ByteDance branding (originally volcengine/verl, now verl-project/verl) — filed here per the co-development attribution, not a sole-HKU artifact.