Human-reviewed snapshots of public web pages (full-page screenshots + the pages' own assets + navigation graphs), built as source material for screenshot-to-code tasks. Access is granted for research use and reviewed manually.

By requesting access you agree to use this archive for non-commercial research only, to respect the copyright of the original website owners, and to honor takedown requests for any site included here.

Curated Web Page Archive for UI-to-Code

6,233 human-reviewed web page captures, each with the page's own images, fonts and a navigation graph — ready to be turned into screenshot-to-code / web-reproduction tasks.

Every site here survived both an automated gate and a manual visual review. If you are building a "here is a screenshot, rebuild this page" dataset, the expensive part is not crawling — it is throwing away the 60% of captures that are unusable. That part is already done.

What makes these captures usable

Crawling a page and getting a faithful, reproducible capture are very different things. Each of the following was a real failure we hit and fixed:

Problem	What it does to a reproduction task	Handled
Ads in the screenshot	The reference is whatever ad was auctioned that second — not reproducible	Ad/tracker requests blocked at capture time; ad slots render empty
Asset served by content negotiation	`photo.jpg` actually contains AVIF bytes; decoders that trust the suffix fail to open it	Extension rewritten from the file header, so the suffix is always the true format
Webfonts pruned as "unused"	Cross-origin `@font-face` rules are unreadable, so fonts looked unreferenced and got deleted	Fonts kept unconditionally (a downloaded font was used)
Below-the-fold iframes	Chromium throttles off-screen frames — full-page shots come back with white holes	Off-screen frames re-shot individually and patched back in
Pages taller than 8000px	A truncated screenshot as ground truth teaches the model to stop early	Dropped, never truncated
Assets referenced but not captured	Reference shows content the solver was never given	Page dropped on any missing visible asset
Cookie/consent overlays	Overlay covers the page being reproduced	DuckDuckGo autoconsent (778 CMP rules) + structural fallback
Sites with no responsive design	A "mobile" shot that is really a scaled desktop page inverts the training signal — writing real breakpoints scores worse than faking it	Measured `innerWidth` must match the nominal width within 2%, else the whole viewport group is dropped

Then a human looked at every surviving page and dropped the site if any page was bad.

What is in each site directory

<site_slug>/
├── meta.json            # start_url, capture timestamp, crawler version, FineWeb lineage
├── pages.jsonl          # per page: url, title, depth, page height, screenshot path, sha1
├── navigation.jsonl     # click-edge graph: from → to, trigger text, bbox, selector
├── screenshots/         # full-page PNG, 1920px wide, DPR 1, animations frozen
└── resources/
    ├── images/ fonts/ medias/
    └── manifest.jsonl   # per asset: source url, local path, and every place it renders
                         # (page id, CSS selector, rendered bbox, context, alt text)

The rendered-bbox information in manifest.jsonl is what lets you decide which assets are content and which are decoration — we use it to pick what ships with a task.

Multi-viewport pools additionally carry screenshots_tablet/ and screenshots_mobile/: the same page captured at 1920×1080 / 1024×768 / 375×812, each in its own browser context (not a resize — a resized page keeps serving desktop image variants and desktop JS branches, so the capture silently lies).

How pages were selected

URLs sampled from FineWeb (cleaned Common Crawl, ODC-BY) → static filters (English, token density, no red-flag domains) → liveness + robots.txt probe → crawl → automated gates → manual visual review.

Sites that were crawled but never used for a task have been removed, so what remains is the set that actually produced tasks — the selection is auditable rather than a pile of raw crawl.

Format

Per-pool .tar volumes (~20 GB each), each with a .manifest.json listing the sites inside and their file counts and byte sizes, so you can verify integrity after download.

Provenance and copyright

Each site's meta.json records its source URL, capture time and crawler version. The captured pages remain the copyright of their original owners; this archive is redistributed for research reproducibility only. The CC-BY-4.0 license covers the archive's structure, metadata and organization — not the third-party page content within it.

If you own a site included here and want it removed, open a discussion on this repository.

Downloads last month: 37

Total file size:

327 GB