kuramot6f/classical-ja-pretrain

Classical Japanese & Literary Chinese Pretraining Corpus (pre-1868)

About 1 billion tokens of text written before 1868, the start of Japan's Meiji era. The corpus covers classical Japanese (古文), early-modern Japanese and Literary Chinese (漢文) as it was read in premodern Japan. It is cleaned, deduplicated and split for language-model pretraining.

The goal is to study a language model that has never seen modern text: what it knows, how it writes, and how it differs from models trained on today's web. This corpus was used to train classical-smollm2-trunk, a small model trained from scratch.

Quick start

from datasets import load_dataset

ds = load_dataset("kuramot6f/classical-ja-pretrain")             # splits: train / validation / test
row = ds["validation"][0]
print(row["title"], row["era"], row["license"])
print(row["text"][:200])

# Stream without downloading everything (~1.6 GB):
for row in load_dataset("kuramot6f/classical-ja-pretrain", split="train", streaming=True).take(3):
    print(row["source"], row["title"])

# Commercial use: drop the non-commercial (CBETA) rows
ds = ds.filter(lambda r: r["license"] != "cc-by-nc-sa-3.0-tw")

What's inside

source	what it is	docs	tokens	license
NDL	OCR of digitized books published ≤1867, from Japan's National Diet Library	30,213	452M	Public Domain Mark
Kanripo (漢籍リポジトリ)	Premodern Chinese classics: Siku Quanshu, Daoist canon, ...	59,856	406M	CC BY-SA 4.0
Kanripo, CBETA-derived	Buddhist canon texts based on the CBETA edition	12,975	133M	CC BY-NC-SA 3.0 TW
Wikisource (ja)	Volunteer transcriptions of classical Japanese works (e.g. 源氏物語, 土佐日記)	987	10.3M	CC BY-SA 4.0
Minna de Honkoku (みんなで翻刻)	Crowd-sourced transcriptions of handwritten Edo-period documents	10,657	2.4M	CC BY-SA 4.0
total		114,688	1.00B	mixed, per row

Token counts use the project's 32k SentencePiece tokenizer, classical-ja-tokenizer.

By era, the corpus is mostly Chinese texts (about 56%) and Edo-period Japan (1603–1868, about 31%). It also includes Muromachi, Azuchi–Momoyama, Kamakura, Nanbokuchō and Heian texts, plus about 8% of NDL books whose exact era is unknown but which were published no later than 1867.

Splits

split	docs	tokens
train	110,593	977.1M
validation	1,684	15.1M
test	2,411	11.6M

Splits are made by work, not by document. Every volume of a multi-volume work lands in the same split, so a model can't see volume 3 in training and be tested on volume 4. The split is stratified by source and era.

Columns

column	description
text	the document text (cleaned and Unicode-normalized)
id	unique id, <source>/<original path or id>
source	ndl, kanripo, wikisource or minhon
license	pdm-1.0, cc-by-sa-4.0 or cc-by-nc-sa-3.0-tw; see LICENSES
title	title from the source catalog
era	e.g. edo, muromachi, heian, cn:唐 (Chinese dynasty), unknown
era_basis	what the era is based on: publication year, catalog dynasty, ...
group	the work this document belongs to (the unit used for splitting)
split	train, validation or test
tokens	token count with classical-ja-tokenizer, without BOS/EOS
normalized_chars	character count
date_status, script_profile	how the date was determined; script mix (e.g. han_without_kana = pure kanbun)
source_document_id, work_candidate_id, unit_type, unit_label	structure ids (book, volume, page)
text_sha256	hash of text
metadata	JSON string with the rest of the provenance: date evidence, OCR-quality statistics, original file hashes

How it was built

1. Collect texts dated before 1868 from the four sources. NDL items were selected from NDL's open bibliographic dataset by an explicit publication year ≤1867.
2. Clean each source: strip markup, remove OCR noise and duplicates, normalize characters. NDL OCR also went through a quality audit.
3. Unify and filter: apply common normalization; drop Kanripo texts whose dynasty couldn't be determined and documents without reliable pre-1868 evidence.
4. Group volumes into works and split.

The per-source cleaned text, before unification, is in classical-ja-sources. The scripts are in code/.

Repository layout

data/       parquet shards (train-0000x-of-00010, validation, test)
LICENSES/   per-source license terms and attribution
meta/       build reports, the list of excluded documents, work groups, fixed evaluation windows
code/       the scripts that built this corpus (Apache-2.0); worklog/ has the research notes (mostly Japanese)

License

Every document keeps its source's license, shown in the license column. Details and required attribution are in LICENSES/README.md.

• pdm-1.0: NDL OCR text of public-domain books. Free to use; NDL asks you to note that the data was processed.
• cc-by-sa-4.0: Kanripo, Wikisource, Minna de Honkoku. Attribution and share-alike required.
• cc-by-nc-sa-3.0-tw: CBETA-derived Buddhist texts. Non-commercial only; keep the CBETA notice.

Limitations

• NDL text is uncorrected OCR of woodblock prints and manuscripts and contains recognition errors.
• Era labels come from publication dates or catalog dynasties, which can differ from when a work was written.
• The corpus is dominated by Chinese texts and Edo-period printed books. Heian and Kamakura literature is a small fraction.
• Texts reflect the views of their time, including religious, political and social views that may be offensive today.

Citation

If you use this corpus, please cite this repository and the original sources listed in LICENSES.