Kotomiya07/kuzushiji-dataset-coco

Dataset Card for Kuzushiji Page Dataset

Dataset Summary

The Kuzushiji Page Dataset packages page images from Japanese historical books with character-level bounding boxes and labels. When supplied during conversion, it also includes reading-column and segment annotations for document-layout analysis. This repository contains the COCO variant; bounding boxes use [x_min, y_min, width, height] in pixels.

This card describes the generated repository Kotomiya07/kuzushiji-dataset-coco. Its statistics are calculated at conversion time rather than copied from the upstream collection:

Statistic	Value
Page images	5,390
Books	44
Character categories	4,341
Split	train only

The source material is the 日本古典籍くずし字データセット (Japanese Historical Character Dataset), owned by the National Institute of Japanese Literature (NIJL) and other institutions and processed by the ROIS-DS Center for Open Data in the Humanities (CODH).

Supported Tasks and Leaderboards

• Character detection / Kuzushiji recognition: use objects.bbox, category, and category_id to locate and classify cursive Japanese characters.
• Document layout analysis: use the optional columns and segments fields to detect reading columns and larger text regions.
• OCR preprocessing and evaluation: use the page image, book identifier, Unicode labels, and layout hierarchy to construct recognition pipelines.

There is no official train/evaluation split, benchmark protocol, or leaderboard for this converted dataset. Users must define evaluation splits appropriate to their task.

Languages

The documents are in Japanese (ja), primarily historical written Japanese represented in Kuzushiji. The metadata does not provide a verified language distribution by book, period, script type, or genre.

Dataset Structure

Data Instances

from datasets import load_dataset

dataset = load_dataset("Kotomiya07/kuzushiji-dataset-coco")
example = dataset["train"][0]

print(example["image_id"], example["book_id"])
print(example["objects"]["bbox"][:3])

A record has the following shape (sequences may be empty):

{
    "image": Image(),
    "image_id": str,
    "book_id": str,
    "width": int,
    "height": int,
    "objects": {
        "bbox": list[list[float]],
        "category": list[str],
        "category_id": list[int],
        "is_pua": list[bool],
        "pua_code": list[str],
        "pua_reading": list[str],
        "pua_memo": list[str],
        "char": list[str],
    },
    "columns": {
        "bbox": list[list[float]],
        "column_id": list[str],
        "char_ids": list[list[str]],
        "segment_id": list[str],
    },
    "segments": {
        "bbox": list[list[float]],
        "segment_id": list[str],
        "column_ids": list[list[str]],
    },
}

Data Fields

Field	Type	Description
image	Image	Full page image.
image_id	string	Source page identifier, without the image extension.
book_id	string	Identifier of the source book.
width, height	int32	Page dimensions in pixels.
objects.bbox	sequence of 4 floats	Character boxes in the repository's declared bbox format.
objects.category	sequence of strings	Unicode labels such as U+3042; PUA labels are preserved.
objects.category_id	sequence of int32	Integer IDs defined by label2id.json.
objects.is_pua	sequence of booleans	Whether each label is in a Unicode Private Use Area.
objects.pua_code	sequence of strings	PUA code, or an empty string for a standard Unicode label.
objects.pua_reading	sequence of strings	Optional reading imported from PUA metadata.
objects.pua_memo	sequence of strings	Optional note imported from PUA metadata.
objects.char	sequence of strings	Character obtained from the Unicode code point.
columns.bbox	sequence of 4 floats	Union box of the characters assigned to a reading column.
columns.column_id	sequence of strings	Column identifiers such as COL0001.
columns.char_ids	sequence of string sequences	Character IDs belonging to each column.
columns.segment_id	sequence of strings	Parent segment ID, or an empty string when unavailable.
segments.bbox	sequence of 4 floats	Union box of columns/characters assigned to a segment.
segments.segment_id	sequence of strings	Segment identifiers such as SEG0001.
segments.column_ids	sequence of string sequences	Column IDs belonging to each segment.

The parallel sequences within objects, columns, and segments are positionally aligned. columns and segments can be empty when the corresponding optional annotation files were not provided or a page was not annotated.

Data Splits

Split	Number of rows
train	5,390

No validation or test split is generated. To reduce leakage between pages from the same work, create downstream splits by book_id, not by randomly splitting individual rows.

Bounding Box Formats

Variant	Coordinates	Normalized
COCO	[x_min, y_min, width, height]	No; pixel units
YOLO	[x_center, y_center, width, height]	Yes; values in 0–1

All character, column, and segment boxes in this repository use the COCO variant. Do not infer the format from the values alone.

Label Mapping Files

• label2id.json: Unicode label to category ID.
• id2label.json: category ID to Unicode label.
• pua_metadata.json: PUA code to reading and memo, only when metadata was available.

PUA (Private Use Area) Characters

Some annotations use PUA (Private Use Area) code points for glyphs without a standard Unicode representation. The original code point is retained in category; the derived fields is_pua, pua_code, pua_reading, and pua_memo make these records explicit. Readings and notes are empty when no PUA metadata was supplied to the converter.

A notable example is the four-way distinction around the kōto ligature:

Character / code	Variant	is_pua	pua_reading
ヿ (U+30FF)	Katakana ヿ (standard Unicode)	false	—
U+E009	Hiragana ヿ	true	こと (koto)
U+E00A	Hiragana ヿ with dakuten	true	ごと (goto)
U+E00B	Katakana ヿ with dakuten	true	ゴト (goto)

Depending on the research question, users may merge these forms, preserve them as separate classes, or exclude them. This choice is not made by the dataset converter.

Dataset Creation

Curation Rationale

This conversion makes the upstream character coordinates directly usable with Hugging Face Datasets and preserves page/book provenance. Optional column and segment structures support layout-aware OCR and reading-order research without replacing the original character annotations.

Source Data

Initial Data Collection and Normalization

The converter reads each source page image and its coordinate CSV. It validates image availability, converts Unicode code points to display characters, assigns deterministic category IDs, and converts boxes to the selected COCO or YOLO representation. Boxes that extend beyond the page image are clamped to the image bounds, so every box in this repository lies within its page. Column and segment boxes are derived as unions of their member character boxes when matching local annotation CSV files are supplied. The images themselves are not resized by this step.

Upstream source:

• Dataset: 日本古典籍くずし字データセット
• Owners: National Institute of Japanese Literature and other institutions
• Processing: ROIS-DS Center for Open Data in the Humanities (CODH)
• DOI: 10.20676/00000340
• Website: codh.rois.ac.jp/char-shape

Who Are the Source Language Producers?

The text was produced by historical Japanese authors, scribes, printers, and publishers. Their identities and demographic attributes are not encoded in this converted dataset. The holding institutions and CODH provide and process the digitized source material.

Annotations

Annotation Process

Character labels and coordinates originate from the upstream dataset. The converter does not re-transcribe or independently verify them. Column annotations are optional local assignments of characters to reading columns. Segment annotations are optional derived or human-corrected groupings of columns. Their availability can therefore differ by book and page; empty sequences do not mean that the page contains no text.

Who Are the Annotators?

See the upstream dataset documentation for the provenance of character annotations. The converter does not store annotator identities for optional column/segment annotations, so their annotator composition and inter-annotator agreement cannot be determined from this repository alone.

Personal and Sensitive Information

Historical pages can contain personal names, addresses, ownership marks, or other information about historical individuals. No dedicated personal-information or sensitive- content audit is performed during conversion. Users should inspect the source material for their intended publication context and follow the policies of the holding institutions.

Considerations for Using the Data

Social Impact of the Dataset

The dataset can support preservation, search, transcription, and accessibility of historical Japanese materials. Automated recognition may also produce plausible but incorrect readings; outputs should not be treated as authoritative transcriptions without review, especially in historical, genealogical, or identity-related research.

Discussion of Biases

The collection reflects the books selected, preserved, digitized, and annotated by the source institutions rather than the full distribution of historical Japanese writing. Character frequencies, genres, periods, hands, print styles, page conditions, and institutions may be uneven. Rare characters and PUA labels are likely to be especially sparse. No demographic, geographic, genre, or performance fairness audit is included.

Other Known Limitations

• Only a train split is provided; reported row counts describe this generated revision.
• Annotation completeness and accuracy are inherited from upstream data and optional local column/segment files; they are not independently audited by the converter.
• Categories can be highly imbalanced, and PUA semantics depend on optional metadata.
• Boxes are axis-aligned and cannot fully describe rotated, touching, damaged, or highly irregular glyphs and regions.
• Column and segment annotations may be absent or partially covered across books/pages.
• category_id values are repository-specific; use the included mapping files rather than assuming IDs are stable across independently generated versions.
• A random page-level split can leak book-specific visual characteristics; split by book for a stronger estimate of generalization.

Additional Information

Dataset Curators

The source collection is curated and processed by NIJL, other holding institutions, and CODH. This Hugging Face packaging is generated by the kuzushiji-hf-converter project; consult the repository history for the maintainers of a particular published revision.

Licensing Information

This generated dataset is distributed under CC BY-SA 4.0. Users remain responsible for checking the upstream dataset terms, providing attribution, indicating modifications, and applying ShareAlike requirements to adaptations. This card is descriptive and is not legal advice.

Citation Information

Please cite the source dataset:

『日本古典籍くずし字データセット』（国文研ほか所蔵／CODH加工）
doi:10.20676/00000340

"Japanese Historical Character Dataset"
(Owned by NIJL and others, Processed by CODH)
doi:10.20676/00000340

When publishing a derived dataset or model, also cite the exact Hugging Face repository revision used (Kotomiya07/kuzushiji-dataset-coco) so that the generated schema and statistics are reproducible.

Contributions and Acknowledgments

Data is provided by the ROIS-DS Center for Open Data in the Humanities, the National Institute of Japanese Literature, and the other holding institutions credited by the upstream dataset.