Grpo Rlvr Training

If your model already gets a task right sometimes but not reliably, and you can check the answer with code (a unit test, a parser, a math ground truth), this skill walks you through GRPO with verifiable rewards. It gives you a working TRL GRPOTrainer setup with vLLM generation, a composite format-plus-correctness reward, and the settled learning rate and KL starting values. The part I'd pay the most attention to is the inspection gate: it insists you read 50 to 100 reward-scored samples by hand before training, because a reward that scores the wrong thing produces a run that looks healthy while optimizing the wrong target. Variants like DAPO, Dr.GRPO, and GSPO are held back until a specific failure shows up, which I think is the right call. The VLM section is explicitly reference-only, so don't expect a runnable path there.

skills.sh installs: 2.9K
GitHub stars: 42.4K
last updated: Jul 22, 2026

View on GitHub

Install

$ npx -y skills add wshobson/agents --skill grpo-rlvr-training --agent claude-code

Installs into .claude/skills of the current project.

Files

GRPO & RLVR Training

This skill assumes finetuning-method-selection already routed here because the target behavior has a verifiable pass/fail signal — not demonstrations (lora-qlora-recipes) or preference pairs (preference-optimization). What follows is when RL is the right tool, the reference recipe, the mandatory reward-inspection gate, and how to pick a GRPO variant when the base recipe misbehaves.

Input: a routing decision (RLVR via GRPO) plus a verifier (code executor, test suite, schema checker, or grader) for the target task. Output format: a validated GRPO config — the kwarg values in references/grpo-memory.md and the reward functions in references/reward-functions.md, not free-form advice — that llm-finetuning-training-engineer consumes directly.

When RL Applies

SKILL.md on GitHub

Details

Repository: github.com/wshobson/agents
License: MIT
First seen: Aug 2, 2026