e1879/showui-web-processed

---
language: [en]
license: cc-by-4.0
task_categories: [image-classification]
tags: [ui-grounding, web, showui, processed]
source_datasets: [showlab/ShowUI-web]
size_categories: [10K<n<100K]
---

# ShowUI-Web Processed

Flattened, normalized, and scenario-split version of [showlab/ShowUI-web](https://huggingface.co/datasets/showlab/ShowUI-web).

Each row is a single (instruction, UI element) pair with normalized bounding-box coordinates.

## Schema

| Column | Type | Description |
|---|---|---|
| `sample_id` | string | Unique row identifier (`{row}_{element}`) |
| `screenshot_id` | string | Groups elements from the same screenshot |
| `image_relpath` | string | Relative path to the screenshot image |
| `scenario` | string | Website/domain inferred from the image path |
| `instruction` | string | Natural-language grounding instruction |
| `bbox_xyxy` | list[float] | Normalized bounding box `[x1, y1, x2, y2]` in `[0, 1]` |
| `point_xy` | list[float] or null | Normalized click point `[x, y]` |
| `element_type` | string or null | UI element type label |

## Splits

| Split | Rows | Strategy |
|---|---|---|
| train | majority | Scenario-based holdout |
| validation | ~10% scenarios | Domain holdout |
| test | ~15% scenarios | Domain holdout |

## Repository Layout

The dataset repo contains both row-level parquet artifacts and image files:

- `flat.parquet` — full flattened table (all rows)
- `splits/train.parquet` — train split
- `splits/val.parquet` — validation split
- `splits/test.parquet` — test split
- `splits/splits.json` — split metadata
- `images/...` — screenshot and UI metadata files

## Images

Screenshot images are hosted in the `images/` directory of this repository.
Use `image_relpath` to construct the path or fetch individual images on demand:

```python
from huggingface_hub import hf_hub_download
from PIL import Image

path = hf_hub_download(
repo_id="e1879/showui-web-processed",
repo_type="dataset",
filename=f"images/{row['image_relpath']}",
)
img = Image.open(path)
```

## Usage

Load train split via `datasets`:

```python
from datasets import load_dataset

ds = load_dataset("e1879/showui-web-processed")
print(ds["train"][0]["instruction"])
```

Or load parquet artifacts directly from the dataset repo:

```python
import pandas as pd
from huggingface_hub import hf_hub_download

train_path = hf_hub_download(
repo_id="e1879/showui-web-processed",
repo_type="dataset",
filename="splits/train.parquet",
)
train_df = pd.read_parquet(train_path)
print(train_df.shape)
```

## Credit

Source dataset: [showlab/ShowUI-web](https://huggingface.co/datasets/showlab/ShowUI-web)

语言：[英语]
许可证：CC-BY-4.0
任务类别：[图像分类（image-classification）]
标签：[UI锚定（UI-grounding）、网页、ShowUI、预处理]
源数据集：[showlab/ShowUI-web]
样本规模：[10000 < 样本数 < 100000]

# 预处理版ShowUI-Web数据集

本数据集是[showlab/ShowUI-web](https://huggingface.co/datasets/showlab/ShowUI-web)的扁平化、归一化且按场景划分的版本。每一行均为一组经过归一化边界框坐标处理的「指令-UI元素」配对数据。

## 数据结构
各列的定义如下：
1. `sample_id`：字符串类型，唯一行标识符，格式为`{row}_{element}`
2. `screenshot_id`：字符串类型，用于对同一张截图中的UI元素进行分组
3. `image_relpath`：字符串类型，指向截图图像的相对路径
4. `scenario`：字符串类型，由图像路径推导得到的网站/域名
5. `instruction`：字符串类型，自然语言UI锚定（UI-grounding）指令
6. `bbox_xyxy`：浮点型列表，归一化的边界框坐标`[x1, y1, x2, y2]`，取值范围为`[0, 1]`
7. `point_xy`：浮点型列表或空值，归一化的点击点坐标`[x, y]`
8. `element_type`：字符串或空值，UI元素类型标签

## 数据集拆分
各拆分集的详情如下：
1. **训练集**：占绝大多数样本，采用基于场景的留出划分策略
2. **验证集**：覆盖约10%的场景，采用域留出划分策略
3. **测试集**：覆盖约15%的场景，采用域留出划分策略

## 仓库文件结构
本数据集仓库包含行级Parquet格式文件与图像文件：
- `flat.parquet`：完整扁平化数据表（包含所有样本行）
- `splits/train.parquet`：训练集拆分数据
- `splits/val.parquet`：验证集拆分数据
- `splits/test.parquet`：测试集拆分数据
- `splits/splits.json`：拆分元数据文件
- `images/...`：截图及UI元数据文件

## 图像文件说明
截图图像托管于本仓库的`images/`目录中，可通过`image_relpath`构建文件路径或按需获取单张图像：
python
from huggingface_hub import hf_hub_download
from PIL import Image

path = hf_hub_download(
repo_id="e1879/showui-web-processed",
repo_type="dataset",
filename=f"images/{row['image_relpath']}",
)
img = Image.open(path)

## 使用方式
可通过以下两种方式加载数据集：
1. 通过`datasets`库加载数据集：
python
from datasets import load_dataset

ds = load_dataset("e1879/showui-web-processed")
print(ds["train"][0]["instruction"])

2. 直接从数据集仓库加载Parquet格式文件：
python
import pandas as pd
from huggingface_hub import hf_hub_download

train_path = hf_hub_download(
repo_id="e1879/showui-web-processed",
repo_type="dataset",
filename="splits/train.parquet",
)
train_df = pd.read_parquet(train_path)
print(train_df.shape)

## 致谢
本数据集的源数据集为：[showlab/ShowUI-web](https://huggingface.co/datasets/showlab/ShowUI-web)