# alibaba/ROLL

An Efficient and User-Friendly Scaling Library for Reinforcement Learning with Large Language Models

Repository: https://github.com/alibaba/ROLL
Canonical: https://ross.abutalabs.com/products/roll
Homepage: https://alibaba.github.io/ROLL/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: agentic, rlhf, rlvr
Last push: 2026-08-26T01:34:48+00:00

## Health v2 (maintenance only)
Score: 78/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 77, longevity 33
- inputs: {"age_days": 462, "days_push": 8, "days_rel": 76, "gap_med": 58, "n_releases_24m": 4}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3374, forks 309 (observed 2026-08-28T04:07:58.441984+00:00)

## What it is
ROLL is an open-source reinforcement learning library from Alibaba for training large language models at scale, supporting algorithms like PPO and GRPO. It uses a Ray-based distributed architecture integrating Megatron-Core, SGLang, and vLLM for efficient training and inference across large GPU clusters.

## Use cases
- train an LLM with RLHF for human preference alignment
- run RLVR training to improve model reasoning
- fine-tune an LLM with PPO or GRPO on a GPU cluster
- train multi-turn agentic RL policies
- scale reinforcement learning post-training to thousands of GPUs
- on-policy distillation of large language models

## When to choose
- you need large-scale RL post-training for LLMs with flexible multi-node GPU allocation
- you want built-in support for PPO, GRPO, multi-task RL, and agentic interaction scenarios
- you want to combine Megatron-Core training with vLLM/SGLang inference engines

## When to avoid
- you only need simple supervised fine-tuning without reinforcement learning
- you lack access to substantial GPU resources
- you need a lightweight single-GPU RLHF tool with minimal setup

## Facets
- artifact type: library
- maturity: active
- function: llm-training, reinforcement-learning, llm-inference, agent-framework, gpu-computing
- domain: large-language-models, reinforcement-learning, machine-learning, gpu-computing
- platform: python, cloud
- tags: rlhf, rlvr, ppo, grpo, ray, vllm, sglang, megatron-core, distributed-training, agentic-rl, ai-agents, gpu, linux, docker

## Member repositories
- alibaba/ROLL (main) score 78

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:58.441984+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 2026-08-29T18:39:55.853826+00:00, confidence not recorded.
  - readme: https://github.com/alibaba/ROLL (fetched 2026-08-28T04:07:58.441984+00:00, sha be3c183a7a01)
  - homepage: https://alibaba.github.io/ROLL/ (fetched 2026-08-29T09:33:38.692482+00:00, sha db0bd3822ad2)
- Data as of 2026-08-30T08:39:29.467469+00:00.