A web-app generation benchmark. Each task is a multi-page UI taken from a public Figma community file. For every task we ship the textual page descriptions, the rendered mockup PNGs, the Figma node structure, the per-page click-annotations, and the distilled data-testid "anchors" that an evaluator uses to score a generated app.
The dataset is delivered in two layers:
Ground-truth layer — one folder per task (1_newsletter/, 2_real-estate/, ...) holding all… See the full description on the dataset page:
https://huggingface.co/datasets/JunJiaGuo/VIS-to-APP-Bench.
