Honkoku-Lines

A million-scale line-level image–transcription dataset for handwritten historical Japanese (kuzushiji), built from crowdsourced transcriptions.

Honkoku-Lines pairs images of individual text lines cut from premodern Japanese documents with diplomatic transcriptions of those lines. It is derived from Minna de Honkoku, a citizen-science platform on which volunteers transcribe and peer-review historical documents, by automatically aligning the page-level transcriptions with line regions detected in the corresponding IIIF page images. No additional line-level manual annotation was required.

This Hugging Face release bundles the redistributable subset: line images whose source rights permit redistribution (CC-BY-4.0, Public Domain Mark, CC0). Metadata for every line — including those whose images cannot be bundled — is included, together with a script that fetches the remaining images directly from each institution's IIIF endpoint.

Dataset summary

Lines (bundled here)	1,018,511 line images with transcriptions
Lines (full dataset)	1,169,304 (metadata for all of them is included)
Source items	4,140 items (3,750 bundled)
Source pages	79,086 page images
Transcribed characters	18.2 M (mean 15.5 per line)
Providers	29 institutions and aggregators over 31 IIIF endpoints
Format	WebDataset — 103 tar shards, ≈ 1 GB each
Image	JPEG, height 256 px, quality 92, 4:4:4, max width 2048 px
Recommended splits	train 1,061,121 / val 52,095 / test 56,088 lines, separated by item

Loading

from datasets import load_dataset

ds = load_dataset("yuta1984/honkoku-lines", split="train", streaming=True)
for sample in ds:
    image = sample["jpg"]    # PIL image, 256 px tall, written horizontally
    text  = sample["txt"]    # transcription, original notation preserved
    meta  = sample["json"]   # per-line metadata (see below)

Or with the webdataset library:

import webdataset as wds
ds = wds.WebDataset("https://huggingface.co/datasets/yuta1984/honkoku-lines/resolve/main/data/data-{000000..000102}.tar")

Sample structure

Each sample is three members of a tar sharing one key, <item_id>_<page_index>_<line_index>:

Member	Content
<key>.jpg	The line image. Lines are written vertically in the source; each crop is rotated to horizontal and scaled to 256 px in height.
<key>.txt	The transcription of that line, with the original notation preserved.
<key>.json	Per-line metadata.

The .json member carries the same fields, with the same names and meanings, as the rows of lines.jsonl.gz, so the two can be used interchangeably:

{
  "image_id": "92123B4F15883F0D4ECE22CC8D56AE01_43_012",
  "item_id": "92123B4F15883F0D4ECE22CC8D56AE01",
  "project_id": "aizu",
  "iiif_host": "iiif.adeac.jp",
  "image_index": 43,                   // page index within the item
  "line_index": 12,                    // line index within the page, in reading order
  "iiif_image_url": "https://.../full/full/0/default.jpg",
  "iiif_region_url": "https://.../X,Y,W,H/full/0/default.jpg",
  "bbox": [x, y, w, h],                // line region in page-image pixels
  "text": "懐《割書：ふところ｜むね》悔",  // transcription as entered
  "plain_text": "懐ふところむね悔",       // notation stripped
  "plain_len": 8,
  "ocr_text": "懐ふところむね悔",         // reference recogniser output (alignment key, not ground truth)
  "edit_distance": 0.0,                // normalised Levenshtein, transcription vs ocr_text
  "length_ratio": 1.0,                 // |ocr_text| / |transcription|
  "det_score": 0.79,                   // line-detector confidence
  "image_license": "CC-BY-4.0",
  "image_license_url": "https://creativecommons.org/licenses/by/4.0/deed.ja",
  "holding_institution": "会津若松市立会津図書館",
  "split": "train"
}

text keeps the markup entered by transcribers — warigaki (interlinear double-column glosses, 《割書：…｜…》), furigana, kaeriten and okurigana for reading Chinese-style text, editorial marks for damaged or uncertain characters, and so on. plain_text is the same line with that markup removed. The notation is specified in koji_notation.md in the Zenodo record.

Companion files

File	Content
lines.jsonl.gz	Per-line metadata for all 1,169,304 lines, including the 150,793 whose images are not bundled. Same schema as the tar .json members, plus image_on_hf.
items.tsv	Per-item bibliography, counts, license and split (4,140 rows).
hosts.tsv	Per-IIIF-endpoint aggregation with the provider name and license mix (31 rows).
examples/fetch_iiif.py	Fetches the non-bundled line images from their institutions.

Assembling the full dataset

Images for items whose rights do not permit redistribution are not in the tars, but every line is described in lines.jsonl.gz with a IIIF region URL. To fetch them:

python examples/fetch_iiif.py --lines-jsonl lines.jsonl.gz --output-dir ./extension --rate 3.0

The script writes WebDataset tars in the same layout, waiting at least 3 seconds between requests to any one host. Please keep a courteous rate: these are live services run by libraries and archives.

How the dataset was built

1. Page images are fetched once from each institution's IIIF Image API and kept, so that later stages can be re-run without touching the source servers again.
2. Line regions are detected with an RTMDet-s detector at 1024 px. Every box is stored with its confidence, unfiltered, so that the threshold can be revisited without re-running inference. Boxes below 0.25 are dropped downstream, and the survivors are ordered right-to-left by column, top-to-bottom within a column.
3. Each line is cropped twice — tightly (10 px margin) for recognition, and generously (45 px) for the released image — and recognised with PARSeq (NDL Kotenseki OCR-Lite).
4. Alignment. The recogniser output is compared against every line of the page's crowdsourced transcription by normalised Levenshtein distance, and the closest one is taken as the candidate pairing. Recognition here is a matching key, not ground truth: the transcription is what is released.
5. Acceptance. A pair is kept when the distance is ≤ 0.4 and the character-count ratio between recogniser output and transcription is within 0.85–1.20. The ratio test is what rejects truncated lines: a distance threshold alone accepts a line that is missing a large part of its text, because omission and misrecognition are indistinguishable to it.
6. Packing. Accepted lines are grouped into shards such that an item never spans two shards, and images are normalised (rotated to horizontal, scaled to 256 px in height, JPEG quality 92 without chroma subsampling).

Lines whose transcription has fewer than three characters after markup removal are dropped, as are two IIIF endpoints and four platform projects whose terms do not permit this use.

The resulting distances are concentrated near zero: 43.5 % of accepted lines match the recogniser exactly, 65.5 % fall below 0.1, and 3.4 % lie in the 0.3–0.4 band.

The construction code is archived in the Zenodo record.

Splits

Splits are assigned per item, so that no two pages of the same source volume land on different sides of the boundary:

import hashlib
def split_of(item_id, seed=0):
    b = int(hashlib.md5(f"{seed}:{item_id}".encode()).hexdigest()[:4], 16) % 1000
    return "train" if b < 900 else "val" if b < 950 else "test"

This is the split field in every sample, in items.tsv and in lines.jsonl.gz. Shards mix splits, so filter on the field rather than by shard.

Licensing

Images carry the license of the institution that provides them; the field image_license states it per record. This release bundles only records under:

• Public Domain Mark (PDM-1.0) — 603,739 lines. No rights reserved.
• CC-BY-4.0 — 414,772 lines. Attribution required.

The transcriptions (text, plain_text) are contributed by the Minna de Honkoku community and are released under CC BY-SA 4.0, as is the curation metadata. The construction code is MIT.

When you redistribute or publish work using this dataset, attribute the holding institutions in accordance with their terms.

Providers

Institution / provider	License	Items	Lines
国立国会図書館 National Diet Library, Japan	PDM-1.0	668	273,033
東京大学附属図書館（地震研究所図書室・総合図書館ほか）The University of Tokyo Library System	CC-BY-4.0 (825), (unspecified) (15)	840	140,982
伊那市立高遠町図書館・合同会社AMANE Takato Library, Ina City / AMANE LLC	CC-BY-4.0	1,385	127,100
福井県文書館 Fukui Prefectural Archives	PDM-1.0 (113), (unspecified) (21)	134	116,232
国立公文書館・高野山金剛峯寺 National Archives of Japan / Kongōbu-ji, Kōyasan	PDM-1.0	88	84,889
ADEAC 提供機関（会津若松市立会津図書館ほか）Institutions on ADEAC (Aizu Library, Aizuwakamatsu City, and others)	CC-BY-4.0 (121), (unspecified) (6)	127	81,017
国文学研究資料館 国書データベース National Institute of Japanese Literature (Kokusho Database)	(unspecified) (65), CC-BY-SA-4.0 (38), CC-BY-4.0 (38), PDM-1.0 (34), CC-BY-NC-SA-4.0 (8), CC-BY-NC-4.0 (2), CC-BY-NC-ND-4.0 (2), CC-BY-ND-4.0 (1)	188	79,405
国文学研究資料館 古典籍データセット National Institute of Japanese Literature (Kotenseki Dataset)	CC-BY-SA-4.0 (59), CC-BY-4.0 (27), CC-BY-NC-ND-4.0 (1), CC-BY-ND-4.0 (1)	88	61,832
京都大学附属図書館 Kyoto University Library	FREE LICENSE with Attribution	79	47,568
琉球大学附属図書館 University of the Ryukyus Library	PDM-1.0	140	41,425
東京学芸大学教育コンテンツアーカイブ Tokyo Gakugei University Educational Content Archive	PDM-1.0 (67), (unspecified) (1)	68	36,288
九州大学附属図書館 Kyushu University Library	PDM-1.0	66	31,906
国立歴史民俗博物館 khirin（鳴門教育大学ほか所蔵）National Museum of Japanese History (khirin)	CC-BY-4.0 (55), CC-BY-SA-4.0 (3)	58	21,486
京都大学古地震研究会 Kyoto University Historical Earthquake Research Group	PDM-1.0 (93), (unspecified) (1)	94	11,716
石川県立図書館 Ishikawa Prefectural Library	PDM-1.0	26	6,042
国立科学博物館 National Museum of Nature and Science	CC-BY-NC-4.0	56	1,693
慶應義塾大学メディアセンター Keio University Libraries	(unspecified)	3	1,638
メトロポリタン美術館 The Metropolitan Museum of Art	PDM-1.0	1	946
筑波大学附属図書館 University of Tsukuba Library	RS-NOC-CR	2	866
国立国語研究所 National Institute for Japanese Language and Linguistics	CC-BY-4.0	2	628
大阪大学附属図書館 Osaka University Library	(unspecified)	3	570
千葉大学附属図書館 Chiba University Library	RS-NOC-CR	9	557
龍谷大学図書館 Ryukoku University Library	(unspecified)	2	508
神戸大学附属図書館 Kobe University Library	(unspecified)	3	480
立命館大学アート・リサーチセンター Art Research Center, Ritsumeikan University	(unspecified)	2	254
東京大学地震研究所 Earthquake Research Institute, The University of Tokyo	(unspecified)	1	169
算額アーカイブ Sangaku Archive	CC-BY-NC-SA-4.0	2	47
岡山大学附属図書館 Okayama University Library	(unspecified)	4	23
バイエルン州立図書館 Bayerische Staatsbibliothek	PDM-1.0	1	4

Line counts cover the full dataset; the bundled subset is restricted to the CC-BY-4.0 and PDM-1.0 records. "(unspecified)" means the source metadata points only to an institutional terms-of-use page from which no standard license identifier could be derived — those images are not bundled here.

Limitations

• Transcriptions are produced and peer-reviewed by volunteers rather than specialists. The platform reports approximately 98.5 % character-level accuracy, so a small fraction of characters is wrong.
• Alignment is automatic. A line image can be paired with a neighbouring line's transcription, and this is more likely in the higher distance bands.
• The character-count test rejects badly truncated lines but not mildly clipped ones; a line may be missing a character at either end.
• Notation is preserved as entered and is not applied uniformly across volunteers.
• Old and new character forms are not normalised, and the transcriptions follow the platform's conventions rather than a scholarly diplomatic standard.
• Coverage is uneven: the four largest providers account for more than half of all lines, and document genres follow what the community chose to transcribe.

Related resources

• Source platform — Minna de Honkoku — https://honkoku.org/
• Source transcriptions — honkoku-data
• Construction code and per-line metadata — Zenodo: https://doi.org/10.5281/zenodo.20656128

Acknowledgements

We thank the more than ten thousand volunteer transcribers of the Minna de Honkoku community, whose peer-reviewed work over nearly a decade is the substance of this dataset, and the institutions that provide IIIF access to their holdings. We acknowledge the National Diet Library Lab for the Kotenseki OCR-Lite recogniser used in alignment. This work was supported by JSPS KAKENHI grants 25H01243 and 26H02553.

Citation

@dataset{hashimoto_honkoku_lines,
  author    = {Hashimoto, Yuta},
  title     = {Honkoku-Lines: A Million-Scale Line-Level Image--Transcription
               Dataset for Historical Japanese OCR Built from Crowdsourced Transcriptions},
  version   = {2.0},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20656128},
  url       = {https://huggingface.co/datasets/yuta1984/honkoku-lines}
}

The DOI above is the concept DOI and always resolves to the latest version; cite 10.5281/zenodo.21801040 to pin version 2.0.

Contact

Please use the Community tab for questions and problem reports.