zhwiki-ob

Dataset Summary

This corpus contains 656,879 cleaned Chinese Wikipedia sections, totaling 492,326,682 characters of plain text, extracted from the full page-history snapshot zhwiki-20260601-pages-meta-history (2026-06-01). One record is one cleaned wiki section of at least 200 characters. Record text is the latest revision of each page as of the snapshot; the author list of each section is attributed by diffing the page's full revision history.

What sets it apart from other large language-model corpora is record-level author provenance. Every record carries a section_hash that resolves, through a bundled OriginBlame ledger, to the exact list of wiki editors who wrote that text, plus the license and year. Major open corpora such as FineWeb or Dolma ship an id, a URL, and a date per document; none carry authors. Here the full attribution chain is one join away, and the ob export-copyright command generates attribution bundles for arbitrary subsets, which is designed to make CC BY-SA compliance push-button.

Metric	Value
Records	656,879 sections
Text volume	492,326,682 characters (corpus.jsonl, 1.3 GB)
Pages parsed	~1.65 million with full revision history, 0 parse errors
Distinct authors in ledger	2,729,243 (pseudonymized)
Provenance ledger	2.5 GB, OriginBlame format v0.2.2
Ledger-record reconciliation	verified, every record resolves

摘要

本数据集包含 656,879 个清洗后的中文维基百科章节，共计 492,326,682 字符，提取自 2026-06-01 快照的完整页面历史转储（约 165 万页面，0 解析错误）。

每条记录的 section_hash 可通过随附的 OriginBlame 溯源账本（.ob/，2.5 GB，格式 v0.2.2）解析出撰写该章节的编辑者列表、许可证与年份，为 CC BY-SA 署名合规提供记录级支持。

账本收录 2,729,243 名编辑者，以 {editor_id}@zhwiki.invalid 化名存储，不含真实邮箱、IP 等信息；账本与语料逐条对账一致。语料文本已通过 17/17 结构残留质量门（零残留验收）。

数据保留贡献者原始的繁简混写，不做字形转换。

Supported Tasks

• Language modeling: character-level Chinese text generation and pretraining. The text is plain characters; no tokenizer is shipped.
• Provenance and authorship-attribution research: per-section author lists support studies of collaborative writing, contributor turnover, and text reuse.
• Attribution-compliance tooling: machine-readable copyright notices can be exported for any subset of records.

Languages

Chinese (zh), predominantly Traditional Chinese as written by zhwiki contributors. The corpus preserves the original script mix: MediaWiki language-variant markup was resolved, but no Traditional/Simplified conversion was applied, so each section reads exactly as its authors wrote it. Text is plain UTF-8 prose at the character level.

Dataset Structure

Files

Path	Size	Contents
corpus.jsonl	1.3 GB	656,879 records, one JSON object per line
.ob/	2.5 GB	OriginBlame provenance ledger, format v0.2.2: authors/ + sections/ + document-index/ (256 hash shards each), log, ob-version — 771 files

Record schema

Each line of corpus.jsonl is a JSON object:

{"text": "玻璃市的名称来源...", "title": "玻璃市", "heading": "地名", "section_hash": "6c348738...", "n_chars": 5168}

Field	Type	Description
text	string	Cleaned section text
title	string	Page title
heading	string	Section heading
section_hash	string	64-hex content hash; the join key into the ledger
n_chars	int	Length of text in characters

Provenance ledger (.ob/)

The ledger ships unpacked as the .ob/ directory alongside corpus.jsonl. It uses OriginBlame format v0.2.2, an open, MIT-licensed format. Layout:

.ob/
├── authors/           # 2,729,243 distinct authors, 256 hash-sharded files
├── sections/          # 256 shards
├── document-index/    # 256 shards
├── log
└── ob-version

Author identities are pseudonymized as {editor_id}@zhwiki.invalid. No real email addresses, IP addresses, or user agents are stored. Per-section provenance records the content hash, source path, author list, license, and year, a superset of what major LM corpora carry. Ledger-record reconciliation was verified: every record's section_hash resolves, and counts match.

Querying the ledger

The MIT-licensed OriginBlame toolkit provides the ob CLI to read the ledger:

• show: provenance for a record
• blame: line-level authorship
• export-copyright: export attribution/copyright notices for arbitrary subsets, designed for CC BY-SA compliance
• revoke / purge: author retraction and physical deletion
• status, log, reconcile: ledger health and integrity

Dataset Creation

Source data

• Chinese Wikipedia full page-history dumps, snapshot 2026-06-01 (zhwiki-20260601-pages-meta-history): 66 .7z archives, ~17 GB compressed, from dumps.wikimedia.org.
• ~1.65 million pages parsed with full revision history; 0 parse errors. Section text comes from each page's latest revision as of the snapshot; per-section authors are computed by attributing every revision's diff to the section level.
• The build streamed through the archives and is resumable; the dumps were never fully decompressed to disk.

Cleaning pipeline

An ordered 15-layer pipeline (L0-L15), battle-tested across earlier builds, produced the text:

• Blacklisted sections dropped: references, external links, see-also, and similar.
• HTML comments, ref tags, tables, and templates removed. Templates that render visible text, such as {{Lang}} and {{tsl}}, are resolved to their display text first.
• Wiki links unwrapped to display text; bare-URL lines dropped.
• MediaWiki language-variant markers resolved first-listed-variant-wins, with no script conversion. The original Traditional/Simplified character mix is deliberately preserved as written by authors.
• HTML entities decoded in a safe order, so double escapes such as < are not over-decoded.
• Final character whitelist: a Unicode-category allowlist strips private-use characters, variation selectors, and control characters.
• The cleaner operates under a never-raise contract and is covered by adversarial unit tests.
• Sections shorter than 200 characters after cleaning were filtered out.

Quality gate

An acceptance harness measures 16 structural residue classes plus illegal characters as hard gates; all must be zero for ACCEPTED status. This corpus passed 17/17 gates. Organic punctuation such as =, ·, and — is monitored but not gated.

Personal & Sensitive Information

The only personal data is public wiki editor usernames, which are public by nature of public wiki editing. Identifiers are pseudonymized in the ledger as {editor_id}@zhwiki.invalid. No email addresses, IP addresses, or user agents are stored. The revoke / purge mechanism supports author retraction and physical deletion for removal requests. For takedown or attribution requests, open an issue on the dataset repository.

Bias & Limitations

• Wikipedia POV and notability bias carries over; coverage reflects the topics the zhwiki community chooses to document.
• The Traditional Chinese predominance reflects zhwiki community demographics, not a neutral balance of written Chinese.
• Single-domain encyclopedia prose; conversational, literary, and other registers are not represented.
• Tables and templates were intentionally removed, so some page information is lost.
• Records are sections, not full documents; surrounding-page context is not guaranteed.
• The 200-character minimum filter drops short sections.
• The snapshot is frozen at 2026-06-01; no updates are provided.
• The text is cleaned plain prose; citation and reference sections were deliberately excluded.

Licensing & Attribution

Read this section before redistributing.

Content licensing. The text is licensed by its contributors under CC BY-SA, via the Wikimedia Terms of Use dual-license with the GFDL. Material whose last relevant revision predates the 2023-06 Wikimedia ToU upgrade is CC BY-SA 3.0; later material is CC BY-SA 4.0. This dataset is distributed under CC BY-SA 4.0 in a way that honors the upstream terms, and the per-section license is recorded in the ledger.

Attribution. WMF ToU §7 accepts attribution via (1) a hyperlink to the article, (2) a link to a stable copy that includes the page history, or (3) a list of authors. This dataset ships the strongest of the three forms: every record carries its author list in the bundled OriginBlame ledger, joined by section_hash, and ob export-copyright generates attribution bundles for any subset. Downstream redistributors of the text should preserve the section_hash linkage or export attributions for the records they redistribute.

ShareAlike. CC BY-SA ShareAlike obligations apply to adaptations of the content. The attribution ledger is machine-readable and designed to make compliance push-button, but the license terms themselves still govern.

Compilation note. The compilation (selection and arrangement) does not claim a separate database license; content licensing governs. This differs from ODC-BY compilations such as FineWeb or Dolma, which disclaim content rights. Here, the content rights chain is the feature.

This card is not legal advice; consult the licenses and the Wikimedia Terms of Use for your use case.

Recommendations for downstream users

• Keep records joined to their section_hash, or run ob export-copyright on your subset before any redistribution of the text.
• Treat adaptations of the content as CC BY-SA adaptations and carry attribution notices with derived artifacts.
• Honor removal requests: if an author is retracted via ob revoke / ob purge, propagate the change to derived artifacts that still carry the affected text.
• Report residual quality issues or authorship concerns through the dataset repository issue tracker.

Versioning

• v2026.06.01: first release. Source snapshot 2026-06-01, ledger format ob-version 0.2.2.
• The dataset is fully regenerable from the public dumps at dumps.wikimedia.org; the provenance ledger uses the OriginBlame format.

Citation

The data belongs to Wikimedia contributors. Where a formal citation is impractical, credit the Chinese Wikipedia contributors and the dumps at dumps.wikimedia.org, and note this dataset's version. If you use the provenance ledger, please also cite the OriginBlame paper:

@misc{zhwiki_originblame_2026,
  title         = {zhwiki-ob: zhwiki sections with record-level author provenance (OriginBlame)},
  year          = {2026},
  note          = {Dataset v2026.06.01; built from zhwiki page-history dumps (2026-06-01) at dumps.wikimedia.org},
  author        = {Xue, Haolin},
  howpublished  = {https://huggingface.co/datasets/tzbkk/zhwiki-ob}
}

@misc{xue2026originblame,
  title         = {OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets},
  author        = {Xue, Haolin},
  year          = {2026},
  eprint        = {2607.13037},
  archivePrefix = {arXiv},
  note          = {arXiv:2607.13037}
}

Acknowledgements

• The Wikimedia Foundation and the thousands of zhwiki contributors who wrote the text, which remains theirs under CC BY-SA.
• Haolin Xue, author of OriginBlame (arXiv:2607.13037).
• dumps.wikimedia.org for hosting the public dumps.