Shelter/UI-testjev-data

UI-TestJev Data

Paired screenshots and requirement-based decisions for UI testing.

UI-TestJev Data contains 4,377 web UI decisions and 311 external Scry screenshot pairs. It supports local requirement checks, full-screen regression testing and keyboard reachability. It is the dataset used to train UI-TestJev.

Why this dataset

A visible change is not always a bug. Captioning and element-grounding datasets describe what is on screen; change annotations alone do not establish a violated requirement. UI-TestJev pairs defects with harmless changes, so models learn to check requirements rather than flag every difference.

Local target checks and full-screen search are kept separate. Related page variants stay in the same split to reduce leakage.

Sources and splits

Split	Local / full-screen / keyboard	Total	Source material
Train	884 / 622 / 84	1,590	50 pages: Start Bootstrap, Flat UI, Material Dashboard, HTML5 UP
Validation	483 / 339 / 24	846	29 pages: Bulma Templates
Development test	1,127 / 787 / 27	1,941	58 pages: AdminLTE
External Scry	Category-presence evaluation	311 pairs	52 app groups: Scrymore Scry Design Diff Eval

Only training examples update model weights; validation selects checkpoints. The AdminLTE family was inspected during earlier repairs, so its test split is a development diagnostic. Scry is evaluation-only. Related variants and shared template families stay together; a count of rows is not a count of independent apps. Source revisions · Additional source attribution.

What a decision contains

• Local check: a public requirement, reference target, ordered reference/current screens and matching detail crops. Labels cover clipping, contrast, missing content, occlusion, image aspect and control size.
• Full-screen check: reference/current screens without a target hint. The task allows at most one introduced failure category; full-screen occlusion is excluded.
• Keyboard check: a target and an observed Tab trace. The labels are reachable, unreachable or unknown when no trace is supplied.

Contracts include a 4.5:1 contrast threshold, an image-aspect tolerance of 3%, or minimum control dimensions of 24 x 24 CSS pixels where applicable. A passing local result covers that requirement only.

How the examples were built

1. Render and verify the reference. Freeze the template revision and viewport, use locally available assets, and confirm that the target is visible and satisfies the requirement.
2. Create both outcomes. Build a failing variant and a changed-but-passing counterpart. Browser measurements check the intended outcome; pixel checks confirm that the change is visible. The mutation name alone does not determine the label.
3. Preserve the evidence. Local crops extend the approved reference target by max(48 pixels, 20% of its dimension), clamped to the screen. The same window is used for both states. Full-screen tasks retain complete screenshots.
4. Keep answers out of inputs. Requirements, image roles and observed traces form the public record. Gold answers, mutation details and oracle measurements are stored separately. Filenames and IDs are lookup keys, not prompt evidence.
5. Check and deduplicate. Remove duplicate public inputs and ambiguous outcomes. Keep source-page and pair groups for sampling and evaluation. Screenshots are stored without additional resizing or recompression.

Keyboard labels use actual Tab traversal. A complete bounded cycle must reach the baseline target; an unreachable variant must remain absent on a complete traversal. Scry keeps its original reference/implementation roles and issue annotations. The model evaluator converts those annotations to category-presence questions.

Dataset checks and coverage

The release audit decoded and hashed 6,536 distinct images, removed 120 duplicate rows (36 train, 32 validation, 52 test), and quarantined 21 ambiguous clipping pairs. It checked matching input/answer IDs, image roles, crop windows and source groups. No exact image or public-input hashes cross splits.

These checks do not certify every semantic label or rule out similar template components across families. Visual review sampled 108 rendered pairs. The training distribution is uneven: local defects include 123 occlusions and 105 control-size failures, but only 30 clipping cases. Keyboard training has 28 examples per label; visual tasks have no unknown targets. Designed subtle/moderate mutations are not calibrated difficulty levels.

Why existing datasets were not pooled into training

Material inspected	Gap for this task
RICO Widget Captioning, ScreenParse	Element captions and boxes identify content, not whether a requirement is violated. Coordinate conventions also need explicit conversion.
DiffSpot, WUICC	A described or generated change can be intentional; change labels alone are not defect labels.
WebSight, WebUI	Screenshot/HTML reconstruction does not supply pass/fail evidence; rendering and geometry need validation.
UIJudgeBench	Reviewed benchmark labels could not be used without their corresponding frozen page evidence; it was excluded.

These sources were inspected to assess suitability, but none contributes training records to this release. The inspection was sampled, not an exhaustive audit of each corpus. Scry is retained only as an external diagnostic, with incomplete negative labels.

Use

python -m pip install huggingface_hub
hf download Shelter/UI-testjev-data --repo-type dataset --include "*.zip" "archives.json" --local-dir ui-testjev-data

Extract each ZIP. Records join by id; image paths are relative to the extraction directory.

<split>/inputs.jsonl
private/<split>.targets.jsonl
private/<split>.provenance.jsonl
images/<sha256>.png

Use inputs.jsonl as model evidence and targets.jsonl for supervision or scoring. Provenance preserves source and pair groups. The private/ name marks input separation; these files are publicly downloadable. Archive checksums · Model input format

Scope

Most examples are synthetic. Natural bugs, native-app interaction, visual abstention and keyboard diversity remain limited. Scry annotations are incomplete: an unannotated difference is not necessarily a false positive. Its category-presence evaluation is not a localization score.

The model technical report connects this coverage to precision, recall, false alarms and per-category failures.

Attribution

Sources: Start Bootstrap, Designmodo Flat UI, Creative Tim Material Dashboard, HTML5 UP by AJ, Bulma Templates, AdminLTE, and Scrymore.

Component-specific licenses and original credits are retained. License and attribution.