LMIS-ORG / slime-agentic

What is slime-agentic?

Slime-agentic is a Python project on GitHub that implements agentic RL pipelines using the Slime post-training framework. It lets you train LLMs on multi-turn agent behaviors—like Planner-Executor-Verifier loops for tool use and reasoning—without manual step annotations, applying RL signals like GRPO directly to trajectories. Developers get Docker-ready setups and scripts to run training on math datasets, boosting benchmarks like AIME 2024 from 10% to 26.7% in quick experiments.

Why is it gaining traction?

It hooks seamlessly into Slime's generate and reward functions, turning single-step RLHF into full agent rollouts with minimal boilerplate—ideal for @project github copilot-style tool integration. Scripts handle SGLang inference for planners/executors, low-precision quantization, and eval baselines, making it faster to prototype than from-scratch agentic RL repos. The project GitHub example for Qwen2.5-7B shows plug-and-play model swaps via GitHub Actions or API.

Who should use this?

RL engineers fine-tuning LLMs for math, coding, or tool-use tasks, especially those already on Slime. It's for teams implementing agentic workflows—like project implementing agency partners building reasoning agents—wanting quick wins on datasets like DAPO-Math without custom infra. Suited for project GitHub Python setups testing slime agentic RL before scaling.

Verdict

Grab it if you're in the Slime ecosystem—solid for agentic experiments despite 19 stars and 1.0% credibility score signaling early maturity. Docs cover Docker pulls and eval scripts, but expect tweaks for production; pair with project GitHub repo forks for custom tools. Promising starter for slime agentic RL, not yet battle-tested at scale. (198 words)