grpo-rl-training

Quick Review

Exceptional skill for GRPO/RL training. The description accurately captures the expertise provided (alignment, reasoning, structured output training). Task knowledge is comprehensive with battle-tested patterns, complete code examples, hyperparameter guidance, debugging workflows, and critical insights (e.g., loss increases during training). Structure is excellent with clear sections, tables, and progressive disclosure from concepts to implementation to troubleshooting. Novelty is strong: GRPO is complex, requires orchestrating multiple reward functions, understanding RL dynamics, and avoiding pitfalls that would consume many agent tokens to discover. The skill provides non-obvious insights (reward scaling, multi-stage training, adaptive weights) that meaningfully reduce implementation cost. Minor opportunity: could slightly expand the description to mention debugging/troubleshooting capabilities, though current description is already strong and invokable.