VCG-Bench Task 1 is a public Vision-to-XML dataset for visual-centric structured diagram generation. Each example pairs a diagram image with domain labels, a structured visual description, and restored Draw.io / mxGraph XML when available.
VCG-Bench is designed to support cold-start training and evaluation for models that generate executable, editable diagram code from visual inputs. The full benchmark and evaluation toolkit are maintained in the companion code… See the full description on the dataset page:
https://huggingface.co/datasets/sxy1620348809/VCG-Bench.