Shelter/UI-testjev-data

UI-TestJev Data — UI Regression and Accessibility Decisions

UI-TestJev Data contains paired UI screenshots and structured decisions for local requirement checking, full-screen visual regression, and keyboard reachability. It is the exact frozen dataset used for Shelter/UI-testjev, a DiffusionGemma LoRA experiment.

The release contains 4,377 synthetic decision examples and 311 external Scry pairs in four unchanged archives, totaling approximately 593 MB. Screenshot images, public inputs, labels, source provenance, and original source notices are included. No model weights are included here.

Splits and intended use

Archive / split	Local checks	Full-screen decisions	Keyboard decisions	Total	Source coverage
jev-v3-train.zip	884	622	84	1,590	50 pages; Start Bootstrap, Flat UI, Material Dashboard, HTML5 UP
jev-v3-validation.zip	483	339	24	846	29 pages; Bulma Templates
jev-v3-test.zip	1,127	787	27	1,941	58 pages; AdminLTE
jev-v3-external-scry.zip	—	311 external pairs	—	311	52 app groups; Scry

Only train is used for fitting. Validation selects checkpoints. The test archive is a development diagnostic: the AdminLTE family was inspected during earlier diagnosis and influenced construction repairs. It remains source-separated but is not an untouched final holdout. Scry is external evaluation only and did not select checkpoints.

These are related decision counts, not independent applications. Local/full-screen questions and benign/failing variants may share a page or screenshot. Vendor-related templates stay in the same source family and split. coverage.json records task, label, page, and family distributions.

Sources actually included

Source	Material and role	Upstream terms
Start Bootstrap SB Admin 2 and Agency	Rendered interfaces; one training family	MIT; bundled asset notices retained
Designmodo Flat UI	Rendered components and example pages; training	MIT
Creative Tim Material Dashboard	Rendered dashboard pages; training	MIT
HTML5 UP by AJ (@ajlkn)	Phantom, Massively, Editorial, Forty; one training family	CC BY 3.0; original credits retained
Bulma Templates	Rendered interfaces; validation	MIT
AdminLTE	Rendered application-style interfaces; development test	MIT; bundled asset notices retained
Scrymore / Scry Design Diff Eval	Reference/implementation screenshots and annotated issues; external diagnostic	CC BY 4.0, as declared in its retained upstream card

Repository revisions are in source-repositories.json. HTML5 UP download hashes and modification notices are in additional-sources.json. HTML5 UP screenshots use the downloaded templates and supplied placeholder images, not separate live-demo photographs. Scry is pinned to b97d030c3ed977a9cdcc1d749f844a42f0213377.

This collection has component-specific licenses; it is not uniformly Apache-licensed. See LICENSE.md and source-notices. Original notices also remain inside every archive. RICO and ScreenParse notices are retained from an earlier acquisition bundle; their presence does not indicate that those records are included.

What the questions measure

• Local requirement check: given a public requirement, approved reference target, full screenshots, and matching reference-derived detail crops, decide pass/unknown/the relevant failure category. Contracts cover text clipping, minimum contrast, required content, occlusion, image aspect, and minimum control dimensions.
• Full-screen regression: given ordered reference/current screenshots without target-location hints, decide pass/unknown/a supported failure category. The constructed task assumes at most one newly introduced category. Full-screen occlusion is excluded.
• Keyboard access: given a target and a bounded observed Tab traversal, decide pass/fail/unknown. Unknown examples omit the trace; screenshots alone cannot establish reachability.

Where applicable, contracts specify 4.5:1 contrast, at most 3% aspect-ratio change, or a minimum 24 by 24 CSS-pixel control. These narrow contracts do not establish comprehensive accessibility compliance. A local pass concerns only the selected requirement.

Visual unknown is an allowed inference option but has no training targets in this release. Keyboard training has 28 examples each of pass, fail, and unknown. The experiment does not supervise generated explanations, severity, fixes, captions, or bounding-box prediction.

How existing material was processed

1. Freeze sources and isolate families. Record revisions and licenses, render with controlled viewport/device scale and locally available assets, and keep related source pages in one split.
2. Verify the reference. Check target visibility, ancestor clipping, and the baseline requirement. Reject unsuitable targets and unsupported evidence.
3. Construct paired outcomes. Introduce a controlled failing state and a changed but passing counterpart for the same requirement. Browser measurements establish the intended outcome; pixel checks confirm an observable change. A mutation name alone is not a label.
4. Keep task evidence separate. Public inputs contain permitted requirements, targets, image roles, viewport, crops, and observed traces. Labels, mutation details, source grouping, and oracle measurements stay in separate files and are excluded from model context.
5. Clean and preserve relationships. Reject ineffective/unexpected changes, quarantine ambiguous clipping, deduplicate public inputs, and retain related pair groups for sampling and uncertainty analysis. Training visits every admitted row once per epoch without repeating minority rows to inflate diversity.
6. Preserve visual evidence. Local crops use a margin of max(48 pixels, 20% of the reference-target dimension), clamped to the screen. The same reference-derived window is applied to both states; crops are not moved using the current fault location. Blind tasks retain full screens.

The completed model run used the pinned original DiffusionGemma processor with aspect-preserving resizing, at most 1,120 visual tokens per image, and an 8,192 input-token limit. Overflow raises an error. The archives contain original rendered screens and approved crops; uploading them does not resize or recompress them.

Keyboard examples use actual Tab traces: the baseline must reach the target in a complete bounded cycle; a controlled unreachable variant removes it from sequential focus and verifies absence on another complete traversal. The paired screenshot remains identical.

The Scry archive preserves reference/implementation roles, issue labels, and normalized issue boxes. Its stored schema remains the upstream-style issue task. The model evaluation helper converts its annotations into fixed category-presence questions at evaluation time. The resulting metric is recall of annotated categories plus a separate count of additional predictions, not localization accuracy or the full upstream Scry scorer. Untagged categories and zero-tagged pairs are not certified clean negatives.

Structure and leakage prevention

Each ZIP contains the following paths, with its own split name:

<split>/inputs.jsonl
private/<split>.targets.jsonl
private/<split>.provenance.jsonl
images/<sha256>.png
audit.json
coverage.json
duplicates.json
source-repositories.json
additional-sources.json
source-notices/...
DATASET_CARD.md

inputs.jsonl, targets.jsonl, and provenance.jsonl join by id. Image paths are relative to the extracted archive root. The stored image order is reference/current/reference_detail/current_detail for local checks; reference/current for full-screen checks; initial_screen for keyboard; reference/implementation for Scry.

The directory name private/ describes separation from model inputs, not access control. All labels and audit records in this public repository are downloadable. Use targets only for supervision/scoring and provenance only for audit/grouping. Never concatenate entire joined records into a model prompt. IDs and image filenames are lookup keys, not textual evidence. A keyboard input's target is the public control descriptor; it is distinct from the gold target in the targets file.

This release preserves the original archive layout rather than flattening heterogeneous tasks into an automatic Dataset Viewer table. The short DATASET_CARD.md inside each frozen archive is historical; this README gives the expanded release documentation. See DATA_PROVENANCE.md for the detailed processing and source audit.

Gaps found in existing datasets

We inspected several datasets before choosing the final source material. Their task labels were not pooled into generic UI-error truth:

• RICO Widget Captioning and ScreenParse: captions, elements, and boxes are useful for grounding but do not establish a violated requirement. Their normalized XYXY versus pixel XYWH conventions also require explicit conversion.
• DiffSpot and WUICC: changes and change descriptions do not inherently mean an error; font, spacing, or opacity changes can be intentional. Mutation metadata supplied as input would leak the answer.
• WebSight and WebUI: screenshot/HTML data support reconstruction, but rendering can depend on external assets. Sampled WebUI geometry used parallel arrays and parent indices that must stay aligned.
• UIJudgeBench and Scry: benchmark roles must remain separate from training. Labels require the corresponding evidence, and incomplete annotations do not establish clean negatives.

RICO, ScreenParse, WebSight, WebUI, DiffSpot, WUICC, and UIJudgeBench are not included as training records in this release. WebSight belonged to an earlier pilot. Inspection scope varied; source-inspection-summary.json records sample sizes and revisions. These observations concern suitability for this objective, not general defects in entire upstream datasets.

Audits and remaining limitations

The construction audit checked 6,536 distinct images, removed 36 train / 32 validation / 52 test duplicates, and quarantined 21 ambiguous or unconfirmed clipping pairs. It found zero cross-split exact image/public-input overlap and preserved source-family isolation. There were 2,564 local-state crop-containment checks before deduplication. See construction-audit.json.

Perceptual screening covered 935 reference/external screens with zero candidates under the recorded heuristic; this does not exclude semantic similarity or model-pretraining overlap. Stratified visual review covered 108 rendered pairs, not independent human adjudication of every label. The processor audit checked 4,688 prompt upper bounds (maximum 5,384 tokens) and 170 processed comparisons covering all 85 local aspect-failure examples; none became pixel-identical pairs. See the source inspection summary and publication-audit.json for the publication recheck of the frozen ZIPs.

Controlled synthetic mutations dominate. Natural bugs, native-app interaction, visual ambiguity/abstention, and keyboard diversity remain limited. Designed subtle/moderate changes are not empirically calibrated medium/hard difficulty. Related images and template components reduce the effective number of independent examples. One synthetic test family cannot establish cross-repository generalization, and that family already influenced development. Use a fresh, independently reviewed natural-bug holdout before making production claims.

Download the exact training archives

Install huggingface_hub, then download into the archive directory used by your training cell. No Hugging Face token is required for this public dataset:

import hashlib, json
from pathlib import Path
from huggingface_hub import HfApi, hf_hub_download

repo = "Shelter/UI-testjev-data"
# Resolve once so all files come from the same immutable commit.
revision = HfApi(token=False).dataset_info(repo).sha
destination = Path("/content/drive/MyDrive/JEV/archives-v3")  # edit to your cell's DATA_DIR
destination.mkdir(parents=True, exist_ok=True)
manifest_path = hf_hub_download(
    repo, "archives.json", repo_type="dataset", revision=revision,
    local_dir=destination, token=False,
)
for item in json.loads(Path(manifest_path).read_text(encoding="utf-8")):
    path = Path(hf_hub_download(
        repo, item["path"], repo_type="dataset", revision=revision,
        local_dir=destination, token=False,
    ))
    with path.open("rb") as stream:
        actual = hashlib.file_digest(stream, "sha256").hexdigest()
    if actual != item["sha256"]:
        raise RuntimeError(f"Checksum mismatch: {path.name}")
print(f"Verified all four archives at revision {revision}")

Mount Drive before using a Drive path in Colab. These are v3 archives; use the matching v3 training code rather than renaming them to v2. archives.json records original archive hashes and sizes; sha256.json covers the publication files. The trained adapter and measured results are available in Shelter/UI-testjev.