Dataset for the paper: "Evaluating Software Architecture Diagram Connectivity with Vision Large Language Models"

This dataset is published as the official supplementary resource for the research paper "Evaluating Software Architecture Diagram Connectivity with Vision Large Language Models". It provides the core evaluation and inference data focused on assessing the capability of Vision Large Language Models (VLLMs) to accurately count connectivity elements (specifically, arrowheads) within software architecture diagrams. All predictions included in this dataset were generated using the Gemma 4 (31-Billion Parameter) Vision Large Language Model, showcasing its visual reasoning and counting capabilities in technical diagrammatic domains. The data provides a direct mapping between the source diagrams and the model's predictions, allowing for a direct comparison of different spatial inference strategies. Dataset Features: image URL: The direct web link to the original software architecture diagram evaluated in the study. pred full image: The numerical prediction (arrowhead count) generated by the Gemma 4 31B model when analyzing the diagram as a single image. pred patch 2x2: The numerical prediction generated by the Gemma 4 31B model when the original diagram is subdivided into a 2x2 grid, allowing the model to process localized patches at a higher resolution.