datawhalechina / agentic-rl-lab
git clone https://github.com/datawhalechina/agentic-rl-lab.git
Reproduces and dissects frontier LLM agent RL algorithms (GRPO, GSPO, DAPO, OPSD, Search-R1, ReTool, ALFWorld and more) with simplified code.
Why you might care
Aimed at learners and researchers without multi-GPU clusters: each method is explained from its paper, then reproduced with runnable code covering data, reward, loss, training loop, and SwanLab logging. Built on the PyTrio Train API, which the author says keeps the setup far lighter than verl, Slime, or AReaL.
+304
stars, 24h · #121
+321
stars, 7 days · #913
+~908
stars, 30 days · #1,193
840
★ total · +62% this week
54
forks · 2 open issues
Star history (as recorded here)
LanguagePython
LicenseApache-2.0
Created2026-07-06 (4mo ago)
Last push1d ago
Watchers1
OwnerOrganization
GitHub topics
Links
- GitHub: github.com/datawhalechina/agentic-rl-lab
- Homepage: https://datawhalechina.github.io/agentic-rl-lab/
- Releases · Issues · Hacker News mentions
README
Loading README…
Blurb generated by claude-haiku-5-5 today. Star deltas marked ~ are extrapolated from partial history.