Worked on expanding evaluation resources for visual-language tasks in the upstash/FlagEmbedding repository by integrating the Circo and FashionIQ datasets into the BGE-VL suite. Leveraged data engineering and dataset management skills to package artifacts in JSON format, ensuring reproducible benchmarking and centralized evaluation assets within the repository. This approach increased evaluation coverage and streamlined the benchmarking process, enabling faster iteration cycles for visual-language models. The work focused on enhancing the depth and breadth of available datasets, supporting more robust model evaluation. No bugs were addressed during this period, with efforts concentrated on feature development and improving dataset accessibility for the team.