Worked on the stanford-crfm/helm repository to deliver a contextual input enhancement for the ArabicMMLU scenario, enriching input data by incorporating context passages when available. This feature improved the quality and reliability of datasets used in model evaluation, addressing a critical missing-context issue that previously limited experimental accuracy. The technical approach involved Python-based data processing to check for and integrate contextual fields, ensuring consistent input structure for downstream machine learning tasks. Collaboration with team members and adherence to repository standards were demonstrated throughout the process, resulting in cleaner code and more robust experimental setups for ArabicMMLU model performance assessment.