Worked on the huggingface/trl repository to improve the consistency and reliability of KL divergence metric reporting during model training and evaluation. Addressed a bug by implementing global normalization for KL divergence logging, ensuring that the calculation sums over all tokens and applies a global mean, which aligns with the approach used for loss normalization. This adjustment reduced variance in metric reporting between training and evaluation modes, enhancing the reproducibility and observability of results. The work was carried out using Python and leveraged expertise in deep learning, machine learning, and reinforcement learning to improve trust in model evaluation and CI regression tests.