GRPO & RLVR Training
Guides reinforcement learning with verifiable rewards (GRPO) to improve tasks with checkable right answers
- Repository
- wshobson/agents
- Stars
- 42,095
- Licence
- MIT
- Files
- 3 files
What it does
When installed, the agent follows a structured recipe for reinforcement-learning fine-tuning of language models on tasks where success can be checked automatically, such as math answers, code passing tests, or valid structured output. It guides choosing this method over simpler fine-tuning, a GRPO training configuration using TRL with vLLM, a required manual inspection of reward-function scores before training, and when to switch to variants like DAPO, Dr.GRPO, or GSPO. It also points to reference files with reward-function code and memory sizing.
Great for
- ML engineers fine-tuning a model for reasoning, math, code, or tool-call tasks with automatic checks
- Teams that already have a test suite, parser, or answer checker to use as a reward signal
- Developers debugging a GRPO run that diverges or games its reward
Skip it if
- Tasks needing subjective judgment or human taste, since the skill says those need evaluation and preference methods instead (reward hacking risk rises with fuzzy rewards)
- Models that never succeed on the task; the skill says to do supervised fine-tuning first
- Vision-language model RL, which the skill explicitly says is not supported as a runnable path in v1
Try asking
- “Set up a GRPO training run for a math model that already solves some problems some of the time”
- “Write a composite reward function that scores both output format and answer correctness”
- “My GRPO run's reward keeps climbing but output quality dropped; help me diagnose reward hacking”
- “Should I use DAPO or Dr.GRPO for my long chain-of-thought model that shows entropy collapse?”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 3 original files, the MIT licence and an attribution note.
grpo-rlvr-training.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip grpo-rlvr-training.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip grpo-rlvr-training.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Inspect 50-100 reward outputs by hand before any full training run; the skill treats this as a mandatory gate
- Start with plain GRPO and only add a variant after you observe its specific failure symptom
- Confirm the base model sometimes succeeds on the task before starting RL, since RL refines rather than teaches from zero
What's inside
- references/grpo-memory.md
- references/reward-functions.md
- SKILL.md
Redistributed unmodified from wshobson/agents under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.