Contributed to the huggingface/trl repository by delivering a targeted documentation update that clarifies the divergence in KL penalty formulas between RLOO and GRPO approaches within reinforcement learning workflows. The work involved adding explicit comments and rationale to the codebase, ensuring users understand the intentional design choices behind the reward penalty application. This update, implemented using Python and grounded in machine learning and reinforcement learning expertise, improves the clarity and maintainability of the RL training loop integration. By addressing edge-case formula differences, the contribution reduces onboarding time, minimizes potential misconfiguration, and supports faster, more accurate adoption of the repository’s RL features.