Worked on the huggingface/trl repository to address a critical issue in the computation of the Generalized Jensen-Shannon Divergence (JSD) loss within the GKDTrainer. Focused on correcting the order of distributions in the KL divergence and ensuring the beta parameter was properly applied to mixture probabilities, which improved the accuracy of loss signals during training. Updated and expanded test coverage to prevent regressions and validate the fix. Utilized Python and PyTorch, applying deep learning and loss function expertise to enhance training reliability and reproducibility for TRL models. The work demonstrated strong debugging skills in math-heavy machine learning codebases.