PULdischo/china-archive-guides

Chinese Archival Handbooks and Wenshi Ziliao Indexes

Machine-readable tables of contents, fonds lists and index entries from two collections at Princeton University Library's East Asian Library, digitized in the library's DPUL "East Asian Library Tables of Contents" exhibit:

• Chinese archival handbooks (档案馆指南): guides published by Chinese archives describing their holdings. Princeton holds the largest known collection. The scans cover each handbook's table of contents and fonds list (全宗介绍 / 全宗名册), not the whole book.
• Wenshi ziliao (文史资料) indexes: the provincial volumes of 中国文史资料总目录电子版, a general index of the "literary and historical materials" series published by local committees of the Chinese People's Political Consultative Conference: memoirs and local history, by locality, series and issue.

The dataset is built for discovery: finding which archives hold records on a subject, period or place, and which local-history volumes contain relevant articles. Every row links back to the original page image in DPUL, which remains the authoritative source.

Version 0.1.2 (2026-10-17). Text and structured fields are machine-extracted and have not been reviewed by people. See Limitations.

Contents

Config	Rows	One row per
guides (default)	372	Handbook (343) or wenshi ziliao provincial index (29): catalog metadata, province, links
volumes	1,591	Volume: each handbook is one; wenshi indexes have 1,248 locality volumes
pages	51,409	Scanned page: HTML and plain-text transcription, page type, OCR engine, IIIF image service, DPUL viewer link
fonds	55,356	Fonds (全宗) row from a handbook's fonds register: number, name, counts, date range
toc_entries	60,147	Table-of-contents entry in a handbook (in fonds guides, mostly fonds names)
wenshi_articles	1,025,796	Wenshi ziliao index entry: series, issue (辑), title, author, page
places	253	Place (province, prefecture, county) from the library's earlier map of the collection, with its catalog links

The collections cover 32 provinces and national-level archives; handbooks were published between 1984 and 2018.

Not included: page images (use the IIIF image_service or viewer_url in pages); full texts of wenshi ziliao articles (only the index entries); the full text of the handbooks (only the scanned contents and fonds lists).

Keys

• guides.id: the DPUL identifier (https://dpul.princeton.edu/ealtoc/catalog/{id}).
• volumes.id: the Figgy (Princeton's repository) resource UUID of the volume's IIIF manifest.
• pages: (guide_id, page_n) is unique. page_n numbers scans across a guide's volumes and matches the site's #p{n} anchors. canvas_id is the IIIF canvas URI.
• fonds, toc_entries and wenshi_articles rows point to their page by guide_id + page_n (and canvas_id); wenshi_articles.volume_id gives the locality volume.

Field notes

• pages.html: sanitized transcription using a small set of tags (h2-h4, p, table with rowspan/colspan, ul class="toc" with span class="au" for authors and span class="pg" for page numbers). pages.text is plain text derived from it.
• pages.page_type: toc, fonds_table, holdings_table, wenshi_index, front_matter, prose, other, or empty for pages read by the fallback OCR engine.
• fonds.date_start / date_end are cleaned: Republic of China years (民国1-38) are converted to 1912-1949, reversed ranges are swapped, and years outside 1368-2030 are dropped. The values as extracted are kept in date_start_raw / date_end_raw, with a date_note saying what changed.
• wenshi_articles.page is the page number printed in the index (the page in the indexed volume, not a scan number): the first page when a range is printed. It is empty where the index prints "( )", no number, or something that is not a page. page_raw keeps the printed value whenever it is more than a single page number (e.g. "116—120", "1 首 13", "期刊 4/94").
• wenshi_articles.series / issue come from the index's series and issue headings; imprint and editor lines (出版：…, 编　辑：…) are not treated as series.
• extraction is vlm (vision-language model) or rules (rule-based parsing of fallback OCR).

Loading

from datasets import load_dataset

guides = load_dataset("PULdischo/china-archive-guides", "guides", split="train")
articles = load_dataset("PULdischo/china-archive-guides", "wenshi_articles", split="train")

Or query the Parquet files directly, e.g. with DuckDB:

SELECT g.title_zh, f.fonds_no, f.name, f.date_start, f.date_end
FROM 'hf://datasets/PULdischo/china-archive-guides/data/fonds.parquet' f
JOIN 'hf://datasets/PULdischo/china-archive-guides/data/guides.parquet' g ON g.id = f.guide_id
WHERE f.name LIKE '%粮食局%';

Downloads without code

The downloads/ folder has the same data for spreadsheet and SQL users:

• Flat CSV files (downloads/csv/): guides, volumes, fonds, contents entries and places, with the guide title, province and DPUL links on every row, so no joins are needed. Files are UTF-8 with a byte-order mark (Excel shows Chinese correctly).
• Wenshi ziliao index entries, one CSV per province (downloads/csv/wenshi_articles/). All of them together (over 1M rows) are in downloads/wenshi_articles.csv.gz.
• Everything in one file: downloads/china-archive-guides-csv.zip (all CSVs, data dictionary, README, license) and downloads/china-archive-guides.sqlite.zip (an SQLite database with all tables (page text, not HTML) and joined views fonds_flat and wenshi_flat).
• downloads/manifest.json lists every file with row counts and SHA-256 checksums.

How it was made

1. Harvest. Catalog metadata and IIIF manifests were harvested from DPUL and Figgy item by item (Figgy has no bulk export), with incremental change detection. Collections: DPUL browse categories "Chinese Archival Handbooks" and "Wenshi Ziliao".
2. Transcription. Page images were requested at 1,500 px width from Princeton's IIIF Image API. 97.5% of pages were transcribed by the vision-language model Qwen3-VL-Plus (Alibaba Cloud DashScope, qwen3-vl-plus). Handbooks were asked for HTML plus structured fonds and contents rows (prompt v2). Wenshi index pages were asked for HTML only (prompt wenshi-v3); their entries are parsed from the HTML. 1,294 pages were read instead by RapidOCR (PaddleOCR models, run locally) with rule-based extraction: the pages the model service refused (see below). Eight pages on which the model failed repeatedly (very wide, mostly empty tables; one very dense index page) were rescued with targeted requests: fonds rows only, or the page split into two halves. pages.ocr_notes records both cases.
3. Normalization. HTML was sanitized to an allow-list of tags; fonds dates were cleaned as above; catalog dates with MARC placeholder digits (e.g. 2004-699) were normalized to the first year.
4. Places. Each guide has a province from its administrative-division code (行政区划代码) in the catalog record, or from its matched place. Handbooks were matched to prefecture and county places from the library's earlier ArcGIS map by catalog record ID, then admin code, then place name. Wenshi volumes were matched by locality name within their province.

The pipeline is open source: https://github.com/PULdischo/china-archive-guides

Limitations

• Machine extraction, not reviewed. Transcriptions and structured rows can contain recognition errors, missed or merged entries, and misread numbers. Check the page image before relying on a detail. In a 40-page test on wenshi index pages, entries parsed from the model's HTML matched counts from an independent OCR engine on most pages, and every entry was correct on a page checked by hand. There is no hand-keyed gold standard yet.
• Content moderation. The model service refused 1,292 pages (2.5%) as "inappropriate content", unpredictably and often on politically sensitive material (e.g. a fonds list including materials on 1989). These pages were read by the fallback OCR engine, so they are present but with plainer layout and rule-based (less complete) structured rows; pages.ocr_notes records this. A dataset built only from the model's output would silently under-represent sensitive material; this one does not drop it.
• Scope of the scans. Handbook scans contain contents and fonds lists only; descriptions of fonds in the body of the books are not included.
• Index counts. The wenshi indexes yield about 23 entries per page (1,025,796 in total), more than the roughly 500,000 articles often cited for this series. Entries are not deduplicated across volumes (reprinted articles appear more than once); about 0.7% are exact repeats within a volume.
• Restricted volumes. Volumes restricted in DPUL were skipped: 中国文史资料总目录电子版. 湖北卷: 黄石市.
• Places. Matching to the earlier map's places is incomplete for townships and districts, and historical place names (e.g. 地区 later renamed 市) are not yet reconciled.
• Dates. Wenshi index items carry the CD-ROM publication year (2009-2012), not the years of the indexed series.

Rights and license

This dataset (structured data and transcriptions of contents, fonds-list and index pages) is released under CC0 1.0. It consists of bibliographic and descriptive information extracted for discovery; it does not include page images or the full texts of indexed works.

The source scans are made available by Princeton University Library through DPUL, which carries their own rights statements: the archival handbooks are marked No Known Copyright and the wenshi ziliao CD-ROM indexes Copyright Not Evaluated. No rights statement appears on the scanned index pages or in their catalog records. Please link to DPUL when presenting material from the dataset.

Maps

Province boundaries used by the project's website come from geoBoundaries (CHN ADM1, public domain, via Wikimedia Commons). The website shows only provinces that have guides in the collection and draws no national outline. geoBoundaries' Xinjiang boundary includes Aksai Chin, which is also claimed by India.

Citation

@dataset{princeton_china_archive_guides_0.1.2,
  title     = {Chinese Archival Handbooks and Wenshi Ziliao Indexes (Princeton East Asian Library)},
  author    = {{Princeton University Library}},
  year      = {2026},
  version   = {0.1.2},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/PULdischo/china-archive-guides},
  note      = {Machine-extracted from scans in the Princeton University Digital Library (DPUL)}
}

Changelog

• 0.1.2: Fixed wenshi_articles: series is no longer taken from imprint or editor lines (出版：…, 编　辑：…; 107,006 entries in 0.1.1); printed page ranges and other non-page values are no longer run together into one number (1,165 entries over 2000 in 0.1.1). page is the first page of a range and the new page_raw column keeps the printed value.
• 0.1.1: Downloads for spreadsheet and SQL users (flat CSVs, per-province wenshi CSVs, a ZIP of everything, an SQLite database). Site bundle now ships the static search index. Data tables unchanged.
• 0.1.0: First release: all 343 archival handbooks (6,160 pages) and all 29 wenshi ziliao provincial indexes (1,248 volumes, 45,249 pages).