README.md

# Chandra OCR Output - davanstrien/ufo-ocr-chandra

This dataset was created by processing images from [davanstrien/ufo](https://huggingface.co/datasets/davanstrien/ufo)
using the [Chandra OCR model](https://huggingface.co/datalab-to/chandra).

## Model Information
- **Model**: Chandra (9B parameters, Qwen3VL-based)
- **Accuracy**: 83.1% on document benchmarks
- **Strengths**: Tables, old scans, long text documents
- **OCR Mode**: ocr_layout

## Processing Configuration
- **Temperature**: 0.0 (initial), 0.3 (retry)
- **Top-p**: 0.1 (initial), 0.95 (retry)
- **Max Output Tokens**: 12384
- **Batch Size**: 28
- **Bounding Box Scale**: 1024

## Output Columns
- `chandra_markdown`: Markdown text extracted from images
- `inference_info`: Metadata about the generation parameters

## Generation Script
This dataset was generated using the UV script from:
https://huggingface.co/datasets/uv-scripts/ocr

To reproduce:
```bash
uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/chandra.py \
    davanstrien/ufo --output-dataset davanstrien/ufo-ocr-chandra --ocr-mode ocr_layout
```

## Citation
If you use this dataset, please cite the Chandra model:
```
@article{chandra2024,
  title={Chandra: High-Accuracy OCR for Document Processing},
  author={DataLab Team},
  year={2024}
}
```