Blame for README.md

---
pretty_name: ScreenProof
license: other
license_name: cc-by-4.0-with-third-party-images
license_link: LICENSE.md
language:
- en
- zh
task_categories:
- image-text-to-text
- object-detection
tags:
- gui
- user-interface
- mobile
- screenshot
- design-consistency
- ui-critique
- benchmark
size_categories:
- 10K<n<100K
configs:
- config_name: default
  default: true
  data_files:
  - split: test
    path: data/consistency_test-*.parquet
  - split: train
    path: data/train-*.parquet
  - split: validation
    path: data/validation-*.parquet
- config_name: text_layout
  data_files:
  - split: test
    path: data/text_layout_test-*.parquet
---

# ScreenProof: Where Is This Screen Wrong? Finding UI Faults Without a Reference

ScreenProof is a benchmark and a training set for **reference-free UI critique** of phone screenshots.
A critic gets one screenshot and no design file, no target brand and no second screen. It must say whether the
screen is correct and put a box on each element that is wrong: a general UI fault (text overflow, overlap, low
contrast, a small tap target and others) or an element that is inconsistent with the rest of the page (a component
from another design language, or a local change of icon, accent colour, corner radius or material).

- Project page and leaderboard: https://hms97.github.io/screenproof/
- Paper: [arXiv:2611.10234](https://arxiv.org/abs/2611.10234)
- Evaluation: `evaluate.py` in this repository (MIT License)

![Eight test screenshots: for each, a zoom on the labelled box (red) and the whole screen.](https://huggingface.co/datasets/HMSSS55/ScreenProof/resolve/main/assets/examples.jpg)

*Examples from the consistency test and the text layout test. Each card shows a zoom on the labelled box (red) and
the whole screen. Labels record code-injected edits; an edit is not by itself proof of a screenshot-visible fault. Nominally clean pages are expected to receive no findings; label validity has not been independently established for every page.*

## Evaluation revision: eval-2026-11-10

The original Parquet splits, labels and main leaderboard remain unchanged. This additive revision provides [exclusion flags and cluster IDs](robustness/records.jsonl), [results and limitations](robustness/README.md), and [a reproducible statistical analysis](robustness.py). It is not a fully cleaned or independently human-validated dataset.

- Original 2,765-page pool: ProofJudge F1 0.481 versus Argon 0.372. Paired base-page bootstrap: difference +0.109, 95% CI [+0.079, +0.140]; coarser requirement/source groups: [+0.045, +0.177].
- Restoring 106 model-disputed pages: F1 0.478 versus 0.360. Aggregate nominal-clean flag rates become 9.0% and 30.0% (original retained pool: 7.4% and 17.8%).
- Excluding author-flagged records and drop/reorder leaves 2,581 pages: F1 0.506 versus 0.395. Excluding component transplants and local-icon too leaves 1,766: 0.627 versus 0.397. These subsets change coverage and do not replace the main benchmark.
- Removing only ProofJudge's inference cut changes F1 from 0.481 to 0.455 and aggregate clean flags from 7.4% to 51.6%; weights are unchanged.

```bash
python robustness.py --data-dir . --output recomputed.json
python evaluate.py --data-dir . --predictions predictions.jsonl --subset structural_quarantine
```

`robustness.py` requires NumPy and reproduces intervals from released per-record sufficient statistics. `evaluate.py` requires PyArrow and scores new predictions against the original Parquet records, with optional subset filtering. The 106 previously excluded records are represented by counts only; their screenshots are not added to the dataset. Pin the Hub commit or the `eval-2026-11-10` tag when available, and report the selected subset with every score.

Known limitations: drop/reorder labels may use whole-screen boxes and depend on a reference; style variation is not necessarily a defect. The existing audit was performed by one author with injection boxes visible, not independent blind annotation. Two evaluated models determined disputed-clean exclusions. Cluster intervals do not correct test-informed model selection. Box matching does not validate explanations. No new training-data ablation or independent human study is claimed. Text-layout evaluation is English-only; implemented CJK rules are not evaluated here.

## Leaderboard

Consistency test: 2,765 screenshots, 607 clean (120 real, 487 generated); IoU >= 0.5. All judges get the same prompt
(`prompt.txt`) and no target. Local models run greedy without thinking; API judges run at default reasoning.
"Clean flagged" is the share of clean pages that the judge flags (lower is better).
Bold: the best value in the column among judges that find faults (F1 >= 0.10).

| Judge | Real clean flagged | Generated clean flagged | Recall general | Recall local style | Precision | F1 | F1 @ IoU 0.3 |
|---|---:|---:|---:|---:|---:|---:|---:|
| ProofJudge (our baseline, trained on ScreenProof-Train) | **2%** | **9%** | **59%** | 25% | **62%** | **0.48** | **0.50** |
| Gemini 4 Argon | 18% | 18% | 38% | 41% | 35% | 0.37 | 0.43 |
| Claude Opus 5.5 (medium) | 62% | 82% | 49% | **49%** | 22% | 0.31 | 0.35 |
| gpt-6.1-sol | 38% | 36% | 45% | 32% | 25% | 0.30 | 0.36 |
| Gemini 3.8 Flash | 38% | 66% | 40% | 26% | 22% | 0.26 | 0.29 |
| GPT-6 Luna | 50% | 45% | 31% | 17% | 20% | 0.22 | 0.27 |
| Qwen3.8-27B | 98% | 88% | 29% | 20% | 11% | 0.15 | 0.17 |
| GLM-4.6V (106B MoE) | 64% | 25% | 14% | 6% | 5% | 0.07 | 0.08 |
| InternVL3.5-38B | 95% | 99% | 19% | 17% | 3% | 0.05 | 0.06 |
| Qwen3-VL-32B | 79% | 36% | 26% | 28% | 3% | 0.05 | 0.06 |
| Qwen3.5-9B base (no training) | 83% | 89% | 21% | 33% | 2% | 0.04 | 0.05 |
| InternVL3.5-8B | 93% | 94% | 13% | 27% | 2% | 0.04 | 0.05 |
| Qwen3-VL-30B-A3B | 27% | 28% | 8% | 8% | 2% | 0.03 | 0.04 |
| Qwen3-VL-8B | 23% | 0% | 2% | 0% | 3% | 0.02 | 0.02 |
| Mistral-Small-3.2-24B | 99% | 100% | 2% | 4% | 0% | 0.01 | 0.02 |
| UI-TARS-1.5-7B (GUI agent) | 8% | 0% | 0% | 0% | 0% | 0.00 | 0.00 |

Text layout test: 367 faults, 75 clean pages, 71 hard negatives; IoU >= 0.5.

| Judge | Faults found | Clean flagged | Hard negatives flagged | Word broken | Needless ellipsis | Characters touch | Lone last word |
|---|---:|---:|---:|---:|---:|---:|---:|
| Claude Opus 5.5 (medium) | 72% | 80% | 96% | 98% | 82% | 73% | 55% |
| gpt-6.1-sol | 54% | 49% | 68% | 100% | 81% | 27% | 25% |
| Gemini 4 Argon | 47% | 32% | 56% | 86% | 81% | 17% | 22% |
| ProofJudge (baseline) | 43% | 15% | 13% | 94% | 96% | 5% | 14% |

A judge's rank on the consistency test predicts its rank on OwlEye (600 real Android screenshots): Spearman
rho = 0.58 between the F1 scores and 0.91 between the shares of clean screens flagged (21 judges and variants).
The interactive leaderboard is on the [project page](https://hms97.github.io/screenproof/#leaderboard).

## Splits

| Config / split | Rows | Images | Content |
|---|---:|---:|---|
| `default` / `test` (consistency test) | 2,765 | 2,765 | 607 clean pages (120 real, 487 generated; 181 of them hard negatives), 920 general faults, 1,238 local style faults (662 transplanted components). One fault per faulty page. |
| `text_layout` / `test` | 513 | 513 | 367 text layout faults, 75 clean pages, 71 hard negatives. English pages. |
| `default` / `train` (ScreenProof-Train) | 17,285 | 11,424 | From 990 pages: 542 generated and 448 public web pages. A row is an image with its label; an image can appear in several rows. |
| `default` / `validation` | 1,571 | 1,571 | Six held-out page requirements. Used to choose checkpoints and thresholds, never the test. |

No test or validation page requirement occurs in the training set. Three page requirements per design language are
held out for testing. The design languages are iOS, Material, One UI, HarmonyOS, ColorOS, MagicOS, OriginOS and
HyperOS.

```python
from datasets import load_dataset

test = load_dataset("HMSSS55/ScreenProof", split="test")                     # consistency test
text = load_dataset("HMSSS55/ScreenProof", "text_layout", split="test")      # text layout test
train = load_dataset("HMSSS55/ScreenProof", split="train")
```

## Fields

| Field | Meaning |
|---|---|
| `id` | Stable row id (`ct-`, `tl-`, `tr-`, `va-` prefix). Predictions for `evaluate.py` use it. |
| `image` | The screenshot (PNG or WebP, full resolution). |
| `page` | The page the image comes from. Faulty copies of one page share it. |
| `design_language` | `ios`, `material`, `oneui`, `harmonyos`, `coloros`, `magicos`, `originos`, `hyperos`; empty for public web pages. |
| `source` | `generated` (our pages and their engine-made copies), `vendor_screenshot` (a vendor's public screenshot), `web_page` (a public web page rendered at phone size, and its engine-made copies). |
| `source_url` | The original URL for `vendor_screenshot` and `web_page` rows. |
| `image_license` | `CC-BY-4.0` or `third-party` (see License). |
| `fault` | The engine's fault name, or `clean`. |
| `fault_group` | Test sets: the fault group used in the paper (`display`, `overlap`, `two_selected`, `typeface`, `reordered_sections`, `low_contrast`, `small_tap_target`, `dropped_item`, `text_overflow`, `transplanted_component`, `icon`, `accent`, `material`, `corner_radius`, `text_layout`, `clean`, `hard_negative`). Train and validation: the training family (`clean`, `designed`, `multi`, `swap`, `local`, `text`, `page`, `general`). |
| `follows` | `true` when the page has no fault. |
| `num_faults` | Number of labelled boxes. |
| `label` | JSON in the same form as a critic's answer (`prompt.txt`): `{"follows": bool, "inconsistencies": [{"element", "box", "kind", "aspect"}]}`. `kind` is `common` or `inconsistent`. Boxes are `[x0, y0, x1, y1]` in 0-1000 screen coordinates. |
| `human_check` | Consistency test only: the verdict of a hand check, where one was made (`wrong` = the labelled fault is real, `clean` = the clean page is clean, `faulty`, `hidden`, `fine` = the label is not right, `unsure`). |

A hard negative changes every sibling alike (for example all list rows), so the page stays consistent and a critic
must stay silent.

## How the labels are made

![The data engine on one generated ColorOS page: each copy holds one fault, made in code.](https://huggingface.co/datasets/HMSSS55/ScreenProof/resolve/main/assets/engine.jpg)

*The data engine on one generated ColorOS page. Each copy holds one fault, made in code; the box is the changed
element's rendered box. Brand faults (left) need a named target; general faults (right) need none.*

Faults are made in HTML page code, so the edit is the label. The data engine makes exactly one change, reads the
changed element's rendered box from the DOM, and re-renders the page identically. Clean labels rest on base pages
that two model judges found on-spec. Clean test pages that both outside judges boxed at the same place were removed
(106 pages). A hand check of 481 test items found 85 mislabelled; `evaluate.py --exclude-human-flagged` leaves them
out (2,680 pages), and every F1 in the paper changes by at most 0.01.

## Evaluation

```bash
python evaluate.py --split consistency_test --predictions preds.jsonl
```

`preds.jsonl` holds one line per screenshot: `{"id": "ct-00012", "answer": {...}}`, or `{"id": ..., "raw": "<model text>"}`.
A reported box finds a fault when IoU >= 0.5, one-to-one and greedy by IoU (`--iou 0.3` for the loose table).
Precision is over all reported boxes, clean pages included. The script also reports consistency recall, general
recall and the share of clean real and clean generated pages that the critic flags. Send the prompt in `prompt.txt`
with the image and nothing else.

## Source manifest

`manifest/source-manifest-2026-10-23.jsonl` is the frozen collection manifest: for each public image its source page,
date of collection, terms of use, size and hash, and for files from public UI datasets their licence.

## License

- Labels, the source manifest and all `generated` images: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
- `vendor_screenshot` and `web_page` images (`image_license = third-party`): copyright stays with their owners.
  They are included only for non-commercial research and the evaluation of this benchmark. See `LICENSE.md`.
- `evaluate.py`: MIT License.

If you own an image in this dataset and want it removed, write to the contact below and we remove it.

## Intended use and limitations

Intended use: evaluate and train critics that find UI faults on phone screenshots without a reference, and study when
such a critic can serve as a first pass before human review.
Out of scope: judging a designer's work or a brand's conformance without human review; use as a production gate on
its own; screens with interaction or motion.
Limitations: single phone screens only; the text layout test is English only; clean labels come from model judges.

## Citation

```bibtex
@article{sun2026screenproof,
  title   = {ScreenProof: Where Is This Screen Wrong? Finding UI Faults Without a Reference},
  author  = {Sun, Huiming and Li, Yue and Yang, Kangning and Li, Ruineng and Cai, Jie and Fu, Lan and
             Fu, Yunwan and Li, Jinlong and Shen, Jinglin},
  journal = {arXiv preprint arXiv:2611.10234},
  eprint  = {2611.10234},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  year    = {2026}
}
```

Contact: Huiming Sun, huiming.sun1@innopeaktech.com