Developed a robust cross-entropy loss function with padding handling for the facebookresearch/fairseq2 repository, targeting improved training stability and accuracy on variable-length, padded input sequences. The solution focused on precise treatment of padding indices and enhanced numerical stability, addressing issues such as NaNs and inconsistent loss during model training. Leveraging Python, PyTorch, and deep learning techniques, the implementation included careful design, thorough testing, and clear commit traceability. This work enabled more reliable batching pipelines and reduced debugging time for models handling padded data, demonstrating a strong end-to-end engineering approach and attention to code quality within a collaborative open-source environment.