OpenRLHF — RLHF-обучение больших языковых моделей

★ 7.7 · ai-tooling

openrlhf is a Claude Code skill that enables high-performance RLHF post-training of large language models (7B–70B+) using Ray, vLLM, and ZeRO-3. It covers the full training pipeline: reward model training followed by fine-tuning via PPO, GRPO, RLOO, or DPO — with GRPO offering a memory-efficient path that requires no critic model. The distributed architecture with GPU resource sharing delivers 2× faster throughput than DeepSpeedChat. It's designed for ML teams and researchers who need a unified alignment training framework across multi-GPU or multi-node clusters, without stitching together separate tools for each algorithm.

• #reinforcement-learning
• #rlhf-training
• #ppo-algorithm
• #grpo-algorithm
• #ray-framework
• #vllm
• #distributed-training