README.md

---
annotations_creators:
- human
language:
- en
license: mit
pretty_name: VisionReasoner UI Dataset
size_categories:
- 1K<n<10K
source_datasets:
- original
task_categories:
- image-segmentation
task_ids:
- semantic-segmentation
---

# VisionReasoner UI Dataset

This dataset contains user interface (UI) images along with associated annotation prompts and solutions for fine-tuning the VisionReasoner model.

## Dataset Description

- **Size**: 245 samples
- **Task**: Image Segmentation (Semantic Segmentation)
- **Language**: English
- **License**: MIT

## Structure

- `images/`: Folder containing UI images (`.webp` format)
- `visionreasoner_dataset.parquet`: Metadata file containing:
  - `id`: Unique identifier for each sample
  - `problem`: Annotation prompt describing the UI element to segment
  - `solution`: JSON-formatted segmentation solution with bounding boxes and points
  - `image`: Reference to the image file
  - `img_height`: Image height in pixels
  - `img_width`: Image width in pixels

## Usage

To load and preview the dataset:

```python
from datasets import load_dataset

# Load the dataset with trust_remote_code=True
dataset = load_dataset("shirve13/Demo", trust_remote_code=True)
print(dataset["train"][0])
```

**Note**: Make sure to use `trust_remote_code=True` as this dataset uses a custom loading script.

## Dataset Loading Script

The dataset uses a custom loading script (`demo.py`) that:
- Loads metadata from the parquet file
- Handles image paths correctly
- Provides proper dataset features for Hugging Face compatibility

## Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{visionreasoner_ui_dataset,
  title={VisionReasoner UI Dataset},
  author={shirve13},
  year={2024},
  url={https://huggingface.co/datasets/shirve13/Demo}
}
```