This dataset is derived from the public Minna de Honkoku v1 and v2 transcription data (2024 ver.) for OCR model training created by National Diet Library (NDL). It contains page-level OCR examples and bounding-box-level cropped text examples.

Configs

pages: one row per page. Columns include image, text, page metadata, dimensions, and bboxes_json.
bboxes: one row per annotated text bounding box. Columns include cropped image, text, page metadata, resized-page coordinates, original coordinates, and annotation flags.

Size

Pages: 20,329 total (train: 18,276, test: 2,053)
Bounding boxes: 371,023 total (train: 333,490, test: 37,533)
v1 pages: 4,679
v2 pages: 15,650

Processing

Images were downloaded from URLs listed in v1_metadata.csv and v2_metadata.csv when the corresponding Minna de Honkoku JSON annotation exists. Permanently unavailable links are skipped. Page images are resized to a maximum width of 1080 px before Parquet export; bounding-box crops are cut from the resized page image. Splits are deterministic by page id.

Source

Minna de Honkoku data: https://github.com/yuta1984/honkoku-data
Minna de Honkoku data with clean text and borderbox: https://github.com/ndl-lab/ndl-minhon-ocrdataset
Existing image URLs are retained in the image_url column for provenance.

Citation

Please cite the original Minna de Honkoku project and the holding institution listed in each row's attribution metadata when using this dataset.

Downloads last month: 10

Total file size:

5.34 GB