# THUDM/slime

slime is an LLM post-training framework for RL Scaling.

Repository: https://github.com/THUDM/slime
Canonical: https://ross.abutalabs.com/products/slime
Homepage: https://thudm.github.io/slime
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-08-26T08:14:40+00:00

## Health v2 (maintenance only)
Score: 80/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 84, longevity 31
- inputs: {"age_days": 441, "days_push": 7, "days_rel": 27, "gap_med": 44.0, "n_releases_24m": 9}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 8261, forks 1191 (observed 2026-08-28T04:10:19.906584+00:00)

## What it is
slime is an open-source LLM post-training framework for reinforcement learning scaling, connecting Megatron-based training with SGLang-based rollout through a unified training/rollout/data-buffer pipeline. It supports flexible custom data generation workflows including tool use, verifiers, sandboxes, and multi-agent environments, and has been battle-tested in training frontier GLM models.

## Use cases
- run RL post-training on large language models
- train LLMs with reinforcement learning at scale
- generate rollout data with custom tools and verifiers
- fine-tune models with agentic workflows and environment feedback
- connect Megatron training with SGLang inference for RL loops
- reproduce frontier-model RL training pipelines

## When to choose
- you need a battle-tested RL post-training framework for large LLMs
- your stack already uses Megatron for training and SGLang for serving
- you need flexible custom data generation, reward, and environment workflows
- you want upstream Megatron/SGLang optimizations without extra abstraction layers

## When to avoid
- you need a simple supervised fine-tuning pipeline without RL
- you require multiple rollout backends beyond SGLang
- you lack multi-GPU infrastructure for large-scale training
- you want a lightweight single-node fine-tuning tool

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, reinforcement-learning, machine-learning, gpu-computing
- domain: large-language-models, reinforcement-learning, machine-learning, gpu-computing
- platform: python
- tags: rlhf, post-training, megatron, sglang, rollout, reward-modeling, agentic-training, gpu, linux, docker

## Member repositories
- THUDM/slime (main) score 80

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:19.906584+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 2026-08-29T17:28:11.074142+00:00, confidence not recorded.
  - readme: https://github.com/THUDM/slime (fetched 2026-08-28T04:10:19.906584+00:00, sha a091eb092b3d)
  - homepage: https://thudm.github.io/slime (fetched 2026-08-29T08:27:52.910444+00:00, sha ddeaafeb4c86)
- Data as of 2026-08-30T08:39:29.467469+00:00.