---
pretty_name: RelationGUI and RelationQA (updated release v2026.09)
language:
- en
task_categories:
- visual-question-answering
- image-to-text
tags:
- gui
- relational-reasoning
- grounding
- desktop
- mobile
- web
size_categories:
- 10K<n<100K
---

# RelationGUI and RelationQA

Dataset resources for **Beyond element-level understanding: Explicit relational understanding for GUI agents**, *Pattern Recognition* 176 (2026), 113262.

[Paper](https://doi.org/10.1016/j.patcog.2026.113262) · [Code and complete release audit](https://github.com/malonghui/RelationGUI)

## Version qualification

This is the **updated release v2026.09**, rebuilt and quality-filtered from the original source annotations. It is an updated dataset, rather than a recovered copy of the paper's original 37,724-sample experimental release. Mobile and web source-group counts match the paper; the historical desktop crop selection remains unverified. Filtering removes invalid annotations and excludes benchmark-source screenshots from training. Do not equate these rebuilt counts with the paper's counts or claim published performance from this release without retraining and evaluation.

## Files

| File | Contents |
|---|---|
| `RelationGUI-cleaned.zip` | Source-group annotations, referenced screenshots, training JSONL files, full audit and exclusions |
| `RelationQA-cleaned.zip` | All 1,009 original questions, gold labels, original/marked/cropped images, source exclusion manifest |
| `SHA256SUMS` | Archive integrity hashes |

Only packaged data and its metadata belong in this repository. Training, inference and evaluation code are maintained on GitHub. Extract both archives into the same parent directory to obtain `RelationGUI/` and `RelationQA/`.

## Statistics

| Platform | Accepted relation groups | Groups eligible for training |
|---|---:|---:|
| Desktop | 1,450 | 473 |
| Mobile | 20,273 | 19,761 |
| Web | 8,947 | 7,859 |
| **Total** | **30,670** | **28,093** |

Generated training files contain 75,754 grounding instructions, 75,754 referring instructions, and 84,278 relation instructions. Several instructions can derive from one source group. Relation instructions represent spatial, functional and hierarchical dimensions; these totals are not counts of independent screenshots.

RelationQA retains the original **1,009 questions** across **350 source screenshots**: desktop 587, mobile 200 and web 222. Label counts: trigger 269, complement 260, parallel 230 and none 250. No new dev/test partition is introduced.

## Format and use

`RelationGUI/annotations.jsonl` contains stable IDs, source IDs, relative image paths, dimensions, normalized element boxes, captions, relationship annotations and explicit `train` or `benchmark_source` tags. `benchmark_source` rows document the original annotation provenance and **must not be used for training**. Use the files in `RelationGUI/train/` to avoid introducing benchmark leakage.

Training rows use `id`, `source_id`, `images`, `instruction` and `output`. Image paths are relative to `RelationGUI/`. Bounding boxes use `[x1, y1, x2, y2]` in `[0, 1]`.

`RelationQA/annotations.jsonl` contains `id`, `source_id`, `platform`, `question`, `options`, `answer` and an `images` mapping with `original`, `marked` and `crop`. These paths are relative to `RelationQA/`. The marked view uses the source's `line_img_path`; the source's `img_path` was unmarked. Evaluate exact label matches with the scorer in the GitHub repository. Gold labels are included for local evaluation and must not enter model prompts or training.

## Filtering and limitations

- Correct swapped image width/height metadata in 7,210 mobile annotation files using decoded image dimensions.
- Exclude 1,650 groups with invalid/out-of-frame element boxes, 66 desktop groups with ambiguous group alignment, and two singleton groups.
- Exclude every benchmark source screenshot and exact decoded-pixel match from all training tasks.
- Derive grounding coordinates from original geometry; do not reuse the legacy normalization bug.
- Preserve original relation descriptions and labels. These automated checks do not establish new human semantic validation.
- Omit desktop crop augmentation and synthetic negatives whose historical inclusion is unverified.

The full source-hash inventory and individual exclusion records are included in the RelationGUI archive. This updated version contains no model checkpoints, API credentials, crawler scripts, or unrelated downstream datasets.

Mobile imagery originates from [MobileViews](https://huggingface.co/datasets/mllmTeam/MobileViews). Desktop and web screenshots depict third-party software and websites. This card does not assign a new license to those underlying materials; their applicable terms remain in force.

## Citation

```bibtex
@article{ma2026relationgui,
  title={Beyond element-level understanding: Explicit relational understanding for GUI agents},
  author={Ma, Longhui and Zhao, Di and Wang, Siwei and Lv, Zhao and Wang, Miao},
  journal={Pattern Recognition},
  volume={176},
  pages={113262},
  year={2026},
  doi={10.1016/j.patcog.2026.113262}
}
```