Worked on the allenai/rslearn repository to enhance data handling and metrics reliability for machine learning workflows. Refactored the data module to better manage disabled datasets, introducing configurable in-memory dataset usage and improved logging for greater flexibility and transparency. Addressed a metrics saving issue by aligning metric computation and storage with the testing phase, ensuring that reported results accurately reflect model performance after testing rather than during validation. Leveraged Python, PyTorch Lightning, and data engineering techniques to improve data integrity, reproducibility, and maintainability, enabling faster iteration cycles and more trustworthy performance metrics for stakeholders and downstream decision-making processes.