Datasets:

wangekxy
/

classical-chinese-punctuation

license: other
license_name: commercial-paid
language:
  - lzh
  - zh
pretty_name: Classical Chinese Punctuation Restoration (文言文断句·标点) — Commercial
task_categories:
  - text-generation
  - token-classification
tags:
  - classical-chinese
  - 文言文
  - punctuation-restoration
  - 句读
  - sft
  - instruction-tuning
  - commercial-dataset
size_categories:
  - 1M<n<10M
configs:
  - config_name: sample
    data_files: sample.jsonl

Classical Chinese Punctuation Restoration · 文言文断句·标点 💰 (Commercial Dataset)

This is a commercial dataset. A free 200-record sample is provided below (sample.jsonl); the full 5.3M-pair corpus is available upon request.

📧 To license / purchase or request a quote, email wangeksy@gmail.com. ✅ Cleared for commercial use — built from public-domain classical works.

The task

Restore punctuation and sentence segmentation (句读) to unpunctuated Classical Chinese — a capability current LLMs perform poorly and inconsistently. Each record is an auto-verifiable supervised pair: the input is unpunctuated hanzi, the output is the same text with gold punctuation.

{"task":"punctuate_classical_chinese",
 "instruction":"为下面的文言文断句并加上标点。",
 "input":"韩信之后也当淮阴钟室难作之时有客匿其孤求抚于萧相国…",
 "output":"韩信之后也。当淮阴钟室难作之时，有客匿其孤，求抚于萧相国…",
 "title":"坚瓠集","dynasty":"清","author":"褚人获","canon":"子藏"}

What you get (full set)

5,314,004 supervised pairs across 10 canons of classical Chinese (history, philosophy, poetry, religion, medicine, arts), spanning the 汉–清 dynasties.
Train / validation / test splits, plus an 81,929-pair held-out evaluation set (contamination-controlled; licensed separately).
Per-work attribution: title, dynasty, author, canon.
Parquet (zstd) shards, one per canon. See CONTENTS.md for the full breakdown.

Why it is high quality

Auto-verifiable labels. Each input is produced by stripping an authoritative edition's punctuation from the output, so input == output minus punctuation by construction (QC: 0 mismatches). No second LLM touches the labels — clean provenance, no synthetic distillation.
Leakage-controlled. Split by work, not by passage. Mojibake-rejected, punctuation-density filtered, globally de-duplicated; cross-canon duplicate works collapsed to one canonical copy.

This repo (listing)

Public 200-record sample only (sample.jsonl). The full corpus, per-canon shards, and held-out eval are delivered on purchase. Payment and delivery are handled off-platform — email wangeksy@gmail.com.

Provenance & rights

Base texts are public-domain classical works (汉–清). An editorial-rights-clean, independently re-punctuated tier (labels produced from the public-domain text without any third-party 整理本) is available on request for redistribution use cases.

Companion commercial datasets from the same corpus: 校勘 Variant Collation (aligned recensions) and an expert-graded 文言→白话 Translation + QA evaluation set. Email for the catalog.