# OpenRLHF/OpenRLHF

An Easy-to-use, Scalable and High-performance Agentic RL Framework based on Ray (PPO & DAPO & REINFORCE++ &  VLM & TIS & vLLM & Ray & Async  RL)

Repository: https://github.com/OpenRLHF/OpenRLHF
Canonical: https://ross.abutalabs.com/products/openrlhf
Homepage: https://openrlhf.readthedocs.io/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: transformers, vllm, large-language-models, raylib, reinforcement-learning-from-human-feedback, reinforcement-learning, proximal-policy-optimization, visual-language-models
Last push: 2026-08-13T11:26:24+00:00

## Health v2 (maintenance only)
Score: 89/100 (v2, computed 2026-09-03T02:39:23.370411+00:00)
- activity 97, release rhythm 85, longevity 80
- inputs: {"age_days": 1131, "days_push": 20, "days_rel": 20, "gap_med": 5.0, "n_releases_24m": 73}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 9956, forks 1005 (observed 2026-08-28T04:10:38.955510+00:00)

## What it is
OpenRLHF is a high-performance, production-ready open-source RLHF framework built on Ray + vLLM + DeepSpeed for scalable reinforcement learning from human feedback. It supports state-of-the-art RL algorithms (PPO, GRPO, REINFORCE++, RLOO), SFT, reward modeling, DPO, and vision-language model training through a unified agent-based pipeline.

## Use cases
- train llm with rlhf ppo
- run grpo reinforcement learning fine-tuning
- fine-tune a 70b model with reinforcement learning from human feedback
- train vision-language model with rl
- sft and reward model training pipeline
- async rl training with vllm rollout
- multi-turn agent reinforcement learning training

## When to choose
- you need scalable RLHF/RL post-training for large language models on multi-GPU clusters
- you want to switch between PPO, GRPO, REINFORCE++, and RLOO with a single flag
- you need vLLM-accelerated generation and DeepSpeed ZeRO-3 training from HuggingFace checkpoints
- you want to train VLMs or multi-turn agents end-to-end with RL

## When to avoid
- you only need simple supervised fine-tuning without RL
- you lack multi-GPU infrastructure or distributed training experience
- you need a lightweight single-GPU hobbyist trainer
- you prefer a non-Python training stack

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, reinforcement-learning, machine-learning, gpu-computing, cli
- domain: reinforcement-learning, large-language-models, machine-learning, gpu-computing
- platform: python
- tags: rlhf, ppo, grpo, vllm, ray, deepspeed, dpo, vlm-training, distributed-training, agent-rl, linux, gpu, docker

## Member repositories
- OpenRLHF/OpenRLHF (main) score 89

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:38.955510+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:20:12.522710+00:00, confidence not recorded.
  - readme: https://github.com/OpenRLHF/OpenRLHF (fetched 2026-08-28T04:10:38.955510+00:00, sha dcee470df895)
  - homepage: https://openrlhf.readthedocs.io/ (fetched 2026-08-29T08:19:56.914305+00:00, sha d555b38ec2b2)
- Data as of 2026-08-30T08:39:29.467469+00:00.