Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

WebSight synthetic screenshot and HTML pairs

Document status: reviewed. Canonical source.

Summary

WebSight provides large synthetic HTML/screenshot pairs for screenshot-to-code training. Version 0.1 contained 823,000 synthetic pairs; version 0.2 expanded to roughly two million and changed generation toward real images and Tailwind CSS. Its scale makes it useful for pretraining or controlled ablations, but the synthetic distribution and version-specific generation choices prevent it from being the sole training or evaluation corpus for NUIF reconstruction.

Evidence

  • The official March 2024 project article describes v0.1 as 823,000 synthetic HTML/screenshot pairs and v0.2 as two million examples.
  • The article states that v0.2 introduced real images in screenshots and switched the generated frontend style to Tailwind CSS.
  • The paper record is arXiv:2403.09029. The authoritative dataset namespace is HuggingFaceM4/WebSight; mirrors are not equivalent provenance.
  • The dataset is intended for screenshot-to-HTML fine-tuning, not recovery of original production sites or binary assets.

Mechanism

Synthetic prompts and generated HTML/CSS are rendered by a browser to create paired targets. This provides exact generated structure for each synthetic screenshot and makes controlled perturbations possible. It also inherits the generator’s design patterns, framework choices, text distribution and resource simplifications.

NUIF relevance

Borrow the scale and exact synthetic pairing for perception pretraining and render-loop ablations.

Adapt by generating a larger share of data directly from canonical NUIF so every entity, operation, resource, layout context and provenance label is known. Hold out entire templates, component families, fonts and visual themes.

Reject treating synthetic HTML as proof of real-world generalization, training on unversioned mutable dataset snapshots, or assuming image pixels identify the original asset bytes.

Open questions

  • What are the exact redistribution, image-source and generated-output terms of the pinned WebSight revision selected for any training run?
  • Which visual/layout distributions are underrepresented compared with the project’s target corpus?
  • Does WebSight pretraining improve typed NUIF operations after controlling for model size and total training tokens?