README.md

---
license: cc-by-nc-4.0
task_categories:
- image-to-text
- visual-question-answering
language:
- en
tags:
- gui
- gui-agent
- ui-understanding
- screenshot
- visual-grounding
pretty_name: UIPro RefExp
size_categories:
- 10K<n<100K
---

# UIPro-RefExp-Data-v1

Part of the **UIPro** GUI-agent training suite (ICCV 2025). This repository packages the
**RefExp** source into the unified UIPro instruction-tuning format, with coordinates
normalized to a **[0, 1000]** grid.

> Referring-expression grounding on mobile UIs (element + intent grounding).

## Dataset at a glance

| | |
| :--- | :--- |
| **Total samples** | 15,545 |
| **Valid images** | 4,630 |
| **Avg. samples / image** | 3.36 |
| **Coordinate scale** | 0–1000 |
| **Source dataset** | [RefExp](https://github.com/google-research-datasets/uibert) |

### Samples by task

| Task | Count |
| :--- | ---: |
| IntentGnd | 14,755 |
| ElemGnd | 790 |

## Repository file structure

| File | Description |
| :--- | :--- |
| `RefExp_s1000_15k.json` | The dataset: a JSON **list** of 15,545 sample objects (schema below). |
| `RefExp_s1000_15k_sample.json` | A small preview slice of the same schema, for quick inspection without downloading everything. |
| `RefExp_s1000_15k_images.zip` | All screenshots referenced by the `image` field, preserving the relative paths stored there. |
| `RefExp_s1000_15k_info.json` | Full generation report — per-task counts, image statistics, invalid-element breakdown, and the exact processing config. |

Unzip `RefExp_s1000_15k_images.zip` and each sample's `image` path resolves relative to the extraction root.

## Sample schema — every field explained

Each element of the main JSON list is one training sample. This dataset's samples use the
following fields:

| Field | Meaning |
| :--- | :--- |
| `conversations` | The vision-language dialogue: a list of turns, each `{"from": "human"|"gpt", "value": ...}`. The **human** turn holds the instruction/question and contains the `<image>` placeholder marking where the screenshot is inserted; the **gpt** turn is the ground-truth answer. |
| `id` | Unique sample identifier, formatted `autogui_<dataset>_<task>_<n>`. The `<task>` segment (e.g. `intentgnd`, `textloc`, `ocr`, `elemgnd`, `elemref`) tells you which task the sample belongs to. |
| `image` | Path to the screenshot **inside `_images.zip`**, relative to the archive root. Load the image by joining this path with your extraction directory. |
| `task_attr` | The task's target attribute in plain form — for grounding tasks the referred element's text/instruction; for OCR/referring tasks the queried coordinate string. Useful for filtering or building custom prompts without parsing the conversation. |
| `unnormalized_box` | Ground-truth bounding box in **original image pixels**, as `[x1, y1, x2, y2]` (top-left, bottom-right). Present when a box is available. Note: the answer in the `gpt` turn is **normalized to 0–1000**, while this field is the raw-pixel box — divide by width/height and multiply by 1000 to reconcile them. |

> **Note:** Fields tied to a bounding box (e.g. `unnormalized_box`) are only present on samples
> that have a box; point-only answers (e.g. some intent-grounding samples) may omit them.

### The `conversations` field in detail

`conversations` is a list of turns that a vision-language model consumes directly:

- Each turn is `{"from": "...", "value": "..."}`.
- `from` is either **`human`** (the prompt) or **`gpt`** (the ground-truth response).
- The token **`<image>`** inside a human turn marks where the screenshot is spliced into the
  prompt — replace it with the actual image when tokenizing.

### Coordinate system

- Answers are **normalized to the `0–1000` range** relative to image width/height.
- A **point** answer is formatted `(x,y)`; a **bounding box** answer is `(x1,y1,x2,y2)`.
- Prompts ending in `(with point)` expect a point; `(with bbox)` expect a box.
- `unnormalized_box`, when present, is the same box in **raw pixels** — combine it with `wxh`
  (or the image's true size) to convert between pixels and the normalized grid.

## Example

```json
{
  "id": "autogui_RefExp_intentgnd_2413",
  "conversations": [
    {
      "from": "human",
      "value": "<image>\nI want to click the settings button on the web page. Please locate the target element I should interact with."
    },
    {
      "from": "gpt",
      "value": "(951,125)"
    }
  ],
  "image": "rico/71692.jpg",
  "unnormalized_box": [
    974,
    194,
    1080,
    287
  ],
  "task_attr": "click the settings button on the web page"
}
```

## Usage

```python
import json, os, zipfile
from PIL import Image
from huggingface_hub import hf_hub_download

repo = "HongxinLi/UIPro-RefExp-Data-v1"
samples = json.load(open(hf_hub_download(repo, "RefExp_s1000_15k.json", repo_type="dataset")))

images_zip = hf_hub_download(repo, "RefExp_s1000_15k_images.zip", repo_type="dataset")
with zipfile.ZipFile(images_zip) as zf:
    zf.extractall("images/")

s = samples[0]
print(s["conversations"])
img = Image.open(os.path.join("images", s["image"]))   # screenshot for this sample
print("image size:", img.size)
```

## About UIPro

UIPro is a generalist GUI agent trained on 20.6M understanding tasks across 13 task types,
followed by agent continued fine-tuning. See the project repository for the full data
pipeline, training recipes and evaluation scripts:
**https://github.com/ZJULiHongxin/UIPro**

## License

Released under **CC BY-NC 4.0** (non-commercial research use). The underlying screenshots and
annotations remain subject to the terms of their original source, [RefExp](https://github.com/google-research-datasets/uibert).

## Citation

```bibtex
@inproceedings{uipro2025,
  title     = {UIPro: A Generalist GUI Agent},
  author    = {Li, Hongxin and others},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025}
}
```