Worked on enhancing distillation loss monitoring in the linkedin/Liger-Kernel repository by implementing a new parameter that returns soft and hard losses separately during model distillation. This addition improved the observability of training processes, making it easier to debug and analyze performance. The solution was developed using Python and PyTorch, with a focus on deep learning and machine learning workflows. Unit tests were added to validate the new monitoring functionality, specifically covering JSD loss and cosine loss scenarios. The work was completed collaboratively, with attention to code quality and documentation, resulting in more transparent and effective model training analysis.