Aigaolc developed a per-token loss calculation feature with granular logging for model training in the alibaba/ROLL repository. Using Python and leveraging data analysis and machine learning skills, Aigaolc engineered a mechanism that computes loss at the token level, updating the logging system to capture this detailed feedback throughout training runs. This approach enhanced the observability and traceability of model performance, allowing for more targeted optimization and efficient debugging. By linking changes directly to specific commits, Aigaolc improved the review and rollback process. The work demonstrated a focused, in-depth contribution to model training infrastructure within a short project period.