GRPO — обучение языковых моделей с подкреплением

★ 7.7 · ai-tooling

grpo-rl-training is a Claude Code skill that delivers expert-level guidance for implementing Group Relative Policy Optimization (GRPO) with the TRL library, enabling fine-tuning of language models using custom reward functions. It covers dataset preparation in chat format, reward function design across four types — correctness (weight 2.0), format (0.5–1.0), length (0.1–0.5), and style (±0.5) — plus production-ready training workflows and debugging patterns. Required dependencies are transformers ≥ 4.47.0, trl ≥ 0.14.0, datasets ≥ 3.2.0, peft ≥ 0.14.0, and torch. The skill is the right choice when you need to enforce structured outputs like XML or JSON, train on verifiable tasks such as math or coding with objective correctness metrics, or improve chain-of-thought reasoning — specifically in scenarios where labeled preference pairs are unavailable and DPO or PPO would be overkill.