README.md

---
license: cc-by-4.0
task_categories:
  - image-to-text
  - optical-character-recognition
language:
  - vi
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: "data/train-*-of-*.parquet"
      - split: test
        path: "data/test-*-of-*.parquet"
---

# UIT_HWDB - Vietnamese Handwritten Text Dataset

## Dataset Description

UIT_HWDB is a comprehensive dataset for Vietnamese handwritten text recognition.
This version combines both line-level and paragraph-level handwriting samples.

Each sample contains:
- A handwritten text image (line or paragraph)
- Ground truth transcription in Vietnamese
- Writer identification
- Unique image identifier

## Dataset Structure

### Data Instances

Each instance contains:
- `image`: A PIL Image of the handwritten text
- `text`: The ground truth transcription in Vietnamese
- `writer_id`: Unique identifier for the writer
- `image_id`: Unique identifier for the image

### Data Splits

| Split | Samples | Writers | Shards |
|-------|---------|---------|--------|
| train | 8,141 | 249 | 8 |
| test  | 232 | 6 | 1 |
| **Total** | **8,373** | **255** | **9** |

## Usage

### Basic Loading

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("YOUR_USERNAME/UIT_HWDB")

# Access samples
for sample in dataset['train']:
    image = sample['image']      # PIL Image
    text = sample['text']        # Vietnamese text
    writer_id = sample['writer_id']
    image_id = sample['image_id']
```

### Parallel Loading (Recommended for Large Datasets)

```python
# Use multiple workers for faster loading
dataset = load_dataset("YOUR_USERNAME/UIT_HWDB", num_proc=8)
```

### Streaming (For Limited Memory)

```python
# Stream data without downloading entire dataset
dataset = load_dataset("YOUR_USERNAME/UIT_HWDB", streaming=True)

for sample in dataset['train']:
    # Process one sample at a time
    pass
```

### Use in Training

```python
from torch.utils.data import DataLoader

# Convert to PyTorch format
dataset = dataset.with_format("torch")

# Create DataLoader
train_loader = DataLoader(
    dataset['train'],
    batch_size=32,
    shuffle=True,
    num_workers=4
)

for batch in train_loader:
    images = batch['image']
    texts = batch['text']
    # Your training code here...
```

## Dataset Creation

This dataset was converted to Parquet format following HuggingFace best practices:
- ✅ Sharded into ~450MB chunks for optimal performance
- ✅ Memory-efficient streaming conversion process
- ✅ Proper image encoding using HuggingFace Image feature
- ✅ Compatible with HuggingFace Dataset Viewer
- ✅ Supports parallel loading and streaming

## Citation

If you use this dataset, please cite the original UIT_HWDB paper:

```bibtex
@inproceedings{uitHWDB,
  title={Vietnamese Handwriting Database (UIT-HWDB)},
  author={Your Authors Here},
  booktitle={Conference Name},
  year={Year},
  organization={Organization}
}
```

## License

This dataset is released under CC-BY-4.0 license.