Developed compression helpers for logit capture and logprob distillation in the allenai/open-instruct repository, focusing on improving efficiency in offline machine learning workflows. The work introduced configurable compression settings, such as vocabulary size and precision, and implemented packing and unpacking functionality for logprob data to streamline data handling and reduce compute and transfer overhead. Leveraging Python and PyTorch, the developer contributed both independently and collaboratively, refining the approach through co-authored pull requests. The solution supports offline distillation via DistillKit and emphasizes robust unit testing to ensure reliability, demonstrating depth in data compression and practical application within machine learning pipelines.