Worked on expanding dataset integrations and evaluation workflows for the stanford-crfm/levanter and marin-community/marin repositories, focusing on machine learning dataset coverage and internal evaluation capabilities. Leveraged Python and YAML to implement supervised evaluation support, integrate benchmarks such as ARC, Hellaswag, PiQA, OpenQA, and Winogrande, and streamline data ingestion using Hugging Face utilities. Enhanced code quality through rigorous linting, documentation updates, and configuration management, ensuring reproducibility and maintainability. Addressed multiple dataset-related bugs and improved data pipeline reliability, enabling faster experimentation and more robust model evaluation. The work supported broader benchmark coverage and accelerated model development within production environments.