During December 2024, contributed to the pytorch/torchtune repository by implementing a Torchdata-based integration for multi-dataset and streaming training data. This feature enabled the simultaneous use of multiple datasets and streaming inputs within the training pipeline, addressing challenges in data handling efficiency and scalability. The solution leveraged Python, PyTorch, and distributed computing techniques to streamline data processing and support heterogeneous data sources. By engineering a more flexible data pipeline, the work laid the groundwork for faster experimentation cycles and more robust machine learning workflows, enhancing both throughput and data utilization without introducing major bug fixes during the development period.