Aerbote88/nom-ocr-data

This is an ongoing project that is being updated frequently with new data.

Dataset Description

A corpus of historical Hán-Nôm manuscript pages annotated for OCR training. Each page has a high-resolution image, per-character bounding boxes with corrected text labels, candidate alternates from two upstream OCR engines (Kandianguji and Nôm Na Việt), reading-order column polygons (text / binding / marginalia / commentary), and per-character uncertainty / IDS / labeler-note metadata.

• Curated by: Albert Errickson (aee2126@columbia.edu)
• Funded by: Self-funded
• Shared by: Albert Errickson
• Language(s) (NLP): vi (Vietnamese), lzh (Literary Chinese / Han); script: Han + Nôm (CJK Unified Ideographs + CJK Ext. B SIP characters)

Licensing

Annotations in this dataset (per-page JSON labels, bounding boxes, column polygons, character labels, IDS strings, labeler notes, and all derived export formats) are licensed under Creative Commons Attribution 4.0 International (CC-BY-4.0).

Most source page images are reproductions from the Hán-Nôm Sino-Vietnamese Digital Collection at Columbia University Libraries, redistributed here with permission as part of work conducted by the Digitizing Vietnam project. The scans were originally produced by the Vietnamese Nôm Preservation Foundation (VNPF, 1999–2018) in partnership with the National Library of Vietnam, and acquired by Columbia University Libraries in 2021. Source manuscripts are held by the National Library of Vietnam, the Thắng Nghiêm Temple (Hà Nội), and the Phổ Nhân Temple (Hưng Yên).

A small number of additional pages come from other publicly available sources (e.g., Internet Archive). All source manuscripts in this dataset predate modern copyright, and any included scans are redistributed in good faith on that basis.

When reusing material from this dataset, please attribute:

• The originating holding institution (per document)
• Vietnamese Nôm Preservation Foundation (VNPF) — original digitizer
• Columbia University Libraries — current digital steward
• This dataset, for the OCR annotations layered on top

Acknowledgements

This dataset is part of the Digitizing Vietnam project — a partnership between Columbia University Libraries and the Vietnam Studies Center at Fulbright University Vietnam — and is built on digitization work performed by the Vietnamese Nôm Preservation Foundation (VNPF, 1999–2018) in partnership with the National Library of Vietnam. Source manuscripts are held by the National Library of Vietnam, the Thắng Nghiêm Temple (Hà Nội), and the Phổ Nhân Temple (Hưng Yên). The OCR annotations in this dataset are added on top of those scans to support training of Hán-Nôm character recognizers, line recognizers, and layout models.

Initial character-detection and first-pass OCR during annotation was performed by Kandianguji (kandianguji.com). All published labels were subsequently corrected by a human annotator. Kandianguji's site states that its resources are for learning and reference. Users of this dataset should consult Kandianguji directly if they intend to use this corpus for commercial-scale model training.

Uses

Direct Use

• Training/fine-tuning Nôm character recognizers.
• Training line-level recognizers — CRNN, TrOCR, PyLaia — on rotated column strips.
• Training layout/segmentation models for woodblock pages with text vs. binding vs. marginalia vs. commentary columns.
• Benchmarking existing Hán-Nôm OCR systems.

Out-of-Scope Use

• Modern Vietnamese (Quốc Ngữ) OCR.
• Non-Hán-Nôm CJK OCR (Japanese kana, modern Simplified Chinese typefaces).
• Author/scribe attribution or dating.
• Generic document-AI pretraining where modern layouts are expected.

Dataset Structure

• chars — one row per character: cropped image + label + column kind + upstream OCR confidence + provenance (slug / page / offset). Use this for character classifiers.
• lines — one row per text column: rotated column-strip image + full transcription + column metadata. Use this for line-level recognizers (CRNN, TrOCR, PyLaia).
• pages — one row per page: full-resolution page image + all column polygons + all character bboxes/labels/confidences + concatenated raw text. Use this for layout models or end-to-end OCR.

All three configs are derived from the same canonical per-page JSON. The on-disk repository layout (documents/<slug>/pages/NNN.jpg + NNN.json) and additional export formats (nnv, COCO segmentation, verbatim dump) are documented in the labeling-tool repository (soon to be published).

Dataset Creation

Curation Rationale

Public Hán-Nôm OCR training data is scarce, especially with column-level layout labels and labeler-uncertainty metadata. The bulk of the source scans for this dataset come from the Vietnamese Nôm Preservation Foundation's two-decade digitization project (1999–2018). VNPF's collection (~1,100 texts) was acquired by Columbia University Libraries in 2021. This dataset extends that preservation mission by adding machine-readable OCR ground truth on top of those scans, so downstream researchers can train Nôm character recognizers, line recognizers, and column-aware layout models against a single, attributable corpus.

Data Collection and Processing

Most source page images come from the Hán-Nôm Sino-Vietnamese Digital Collection at Columbia University Libraries — digitized by the Vietnamese Nôm Preservation Foundation (VNPF, 1999–2018) in partnership with the National Library of Vietnam, and acquired by Columbia in 2021. A smaller portion of pages come from other publicly available digital collections (such as the Internet Archive) covering manuscripts of similar vintage. This dataset is part of the Digitizing Vietnam project (Columbia × Vietnam Studies Center, Fulbright University Vietnam). Pages are imported into the labeling tool via IIIF Presentation manifests (or, where applicable, as image / PDF uploads of the same scans), with the largest available image resolution per canvas pulled down for local annotation. Optional preprocessing (grayscale + contrast 200% by default) is applied before OCR. First-pass character detection and recognition uses Kandianguji; each detected bbox is then re-OCR'd by Nôm Na Việt, with SIP-range Nôm characters preferred and the alternate kept in choices[]. A human annotator then corrects every character in a custom Next.js workspace, with auto-save to per-page JSON.

Who are the source data producers?

Original manuscripts: Historical Vietnamese scribes (dates mostly unknown; the broader VNPF collection covers material from roughly 1667 to 1957). Originals are held by the National Library of Vietnam (Hanoi), the Thắng Nghiêm Temple (Hà Nội), or the Phổ Nhân Temple (Hưng Yên), depending on the document; the small number of pages from non-Columbia sources cover manuscripts of comparable provenance and age.

Digital scans: Vietnamese Nôm Preservation Foundation (1999–2018), in partnership with the National Library of Vietnam.

Current digital steward: Columbia University Libraries (since 2021).

Annotations

Annotation process

Annotations are produced semi-automatically: bboxes from Kandianguji, character text from Kandianguji and Nôm Na Việt re-OCR (kept as alternates when not chosen), then human correction per character. Annotators flag low-confidence chars as uncertain (excluded from training exports by default), and may attach free-form notes. Column polygons + reading order are also generated semi-automatically and corrected by the annotator.

Who are the annotators?

Single annotator (Albert Errickson) so far

Personal and Sensitive Information

None known. Source material is historical printed text; no living individuals' personal data is captured.

Bias, Risks, and Limitations

• Genre / collection bias: source manuscripts primarily come from three institutions (NLV, Thắng Nghiêm Temple, Phổ Nhân Temple) via the VNPF digitization project; certain types of texts may be overrepresented in the data.
• Single-annotator labels for many pages → no inter-annotator agreement metric.
• Some characters are unencoded in Unicode and will be represented via IDS placeholders (in-development); downstream models must decide how to handle them.

Citation

BibTeX:

@misc{errickson2026nomocr,
  title  = {Nôm OCR Training Corpus},
  author = {Errickson, Albert},
  year   = {2026},
  howpublished = {Hugging Face Datasets},
  url    = {https://huggingface.co/datasets/aerbote88/nom-ocr-data}
}

APA:

Errickson, A. (2026). Nôm OCR Training Corpus [Data set]. Hugging Face. https://huggingface.co/datasets/aerbote88/nom-ocr-data

Glossary

• Nôm (chữ Nôm): logographic script for Vietnamese, built from Han characters + Vietnamese-coined glyphs (many in CJK Ext. B SIP range).
• IDS (Ideographic Description Sequence): Unicode notation for describing unencoded characters by structural composition (e.g. ⿰口巴).
• Column kind: text (main body), binding, marginalia (margin notes), commentary (interlinear smaller-font notes).
• Kandianguji / Nôm Na Việt: the two upstream OCR engines whose outputs seed the human-correction workflow.

Dataset Card Author

Albert Errickson

Dataset Card Contact

aee2126@columbia.edu