Contributed to the stanford-crfm/helm repository by expanding the SHC Benchmark Dataset and refining prompt instructions to improve biomedical NLP evaluation. Focused on adding privacy- and proxy-oriented SHC datasets, the work enabled more comprehensive benchmarking for privacy-sensitive biomedical text understanding. Applied data engineering and prompt engineering skills to curate datasets and standardize 'A'/'B' response formats, supporting reproducible benchmarking workflows. Leveraged Python and version-controlled collaboration to ensure consistent integration and maintainability. The feature enhanced HELM’s ability to evaluate machine learning models in biomedical contexts, facilitating faster iteration on prompts and providing clearer signals for production readiness without introducing major bug fixes.