README.md

---
license: mit
---
**791 annotated images for PaddleOCR text detection — all in vertical (top-to-bottom) Chinese ancient literature layout.**

![0204](https://cdn-uploads.huggingface.co/production/uploads/6985de6c5725615efb85ea8a/4-uhxnzN7TMsJPrBjAhww.jpeg)

This dataset is built specifically for the classic Chinese vertical typesetting you see in ancient books and documents. Instead of the usual left-to-right rows, the text flows down in columns — and PaddleOCR sometimes misses a lot of it in that format (see the original issue here: https://github.com/PaddlePaddle/PaddleOCR/issues/17856).

Every sample gives you:
- The original vertical scan/photograph
- A matching PaddleOCR detection annotation file (standard label format with rectangle coordinates for text regions)

## Why this dataset?
Most public OCR datasets are horizontal and modern. Ancient Chinese vertical text is a blind spot for a lot of models, so I put this together to help close the gap. It’s perfect if you’re working on digitizing old books, historical archives, or any project that needs solid vertical detection.

## Size
- 791 images + annotations
- Purely Ancient Chinese vertical layout (no mixed directions or languages)

## How to use it
1. Download `vertical_annotations.zip` (361 MB).
2. Unzip — you’ll get the images and their paired annotation files.
3. Drop them straight into your PaddleOCR training pipeline for the detection (det) task. 

It builds from my [ocr-producer](https://github.com/alrowilde/ocr-producer) tool (and the [synthesis system](https://github.com/alrowilde/ocr-producer/blob/main/synthesisSystem/README.md)) to generate extra synthetic vertical examples.

## Structure (inside the zip)
The files are organized in the standard PaddleOCR detection style — one image + one annotation file per sample (exact folder layout matches what PaddleOCR expects).

## License
MIT — use it freely for research, commercial work, or whatever you need. Attribution is appreciated but not required.

This one comes from the same workflow as my [invoice-checkmark-annotations](https://huggingface.co/datasets/AlroWilde/invoice-checkmark-annotations) and the ocr-producer repo.

## Contact
Questions, ideas for more vertical/ancient-text datasets, or just want to chat about OCR? Reach out at hi@support.alrowilde.com or open an issue on the ocr-producer GitHub.

Happy training — hope this helps your vertical OCR hit the next level! 📜