museado/newspapers

Metadata about American historical newspapers, built to find the right paper before opening a page: what each title is, who wrote its history, which years we actually hold, how legible the OCR is, and what each page looks like (columns, headline sizes, masthead, paper size).

It is built around the Library of Congress Chronicling America collection (the National Digital Newspaper Program, NDNP) and its bulk OCR batches. There are 2,997 OCR batches covering 4,420 titles. The OCR archives themselves and the page images are not in this repository; this is the layer of facts about them that was expensive to produce and is worth keeping.

Maintained by the Museado project. Produced by the harvest application (provider:cron-america:* commands); the code that made each file is named below.

Why this is worth persisting

1. The source is hard to re-read at scale. From our experience (2026-10): loc.gov's search and item pages refuse non-browser clients and datacenter addresses with a bot challenge (HTTP 403), the bulk-metadata search rate-limits to roughly ten requests per ten minutes, the human title pages are slow even in a browser, and the site was reporting "technical difficulties" while we worked. Every record below was fetched politely, one at a time, from a residential connection. Doing it once and keeping it means nobody has to do it again, and that a build never depends on LOC being reachable.
2. Some of it exists nowhere else in machine-readable form. The essay on a paper's history (written by state historical societies and libraries) and the title lineage (what a paper was called before and after) are in the title record, not in the OCR archives.
3. The page statistics cost days of CPU. Each OCR batch is a bzip2 tarball of ALTO XML (the biggest is over a gigabyte). Reading every page for confidence, hyphenation, block counts and geometry took roughly a hundred batches per hour. The output is a few kilobytes per page and is deterministic.
4. They answer questions the catalogue cannot. "Which Southern papers have readable text for 1900–1930?" needs year coverage as the archives really hold it (LOC's own counts lag; one batch claimed 123 issues and held 155), OCR quality, and layout, not just place and dates.
5. Provenance. Each derived file records the code version and the SHA-256 of the archive it was computed from, so a number can be traced and re-derived.

What is in this repository

Counts are as of the export that produced each file (see the file header / fetchedAt). "Status" says whether the file is in the repository yet.

File	Shape	One row per	Status
title-records.jsonl.gz	LOC title record	newspaper (LCCN)	uploaded for the first 133 titles; the full set (4,420) follows
batches.jsonl.gz	LOC bulk-batch inventory	OCR batch	uploaded (2,997 batches)
page-records/<lccn>.jsonl.gz	LOC page record (trimmed)	page of an issue	uploaded for 132 titles (547,848 pages); more follow
listings/<archive>.listing.jsonl.gz	page statistics (v3)	page	planned, still being built
geometry/<archive>.geometry.jsonl.gz	raw layout measures (v2)	page	planned, still being built
layout/<archive>.layout.jsonl.gz	derived layout profile	issue (pages nested)	planned
title-stats.jsonl.gz	OCR/coverage roll-up	newspaper	planned
layout-titles.jsonl.gz	layout roll-up	newspaper	planned

Files are gzip-compressed JSON Lines (.jsonl.gz), one JSON object per line, readable directly by the datasets library, pandas.read_json(lines=True), DuckDB and zcat | jq.

title-records.jsonl.gz — the title record

What: the Library of Congress's own record for a newspaper title, one row per LCCN. How made: provider:cron-america:title-records requests https://www.loc.gov/item/<lccn>/?fo=json&at=item once per title (about 3 s apart, with a descriptive User-Agent), caches the response, then --export folds the cache into this file. Index internals that change on every request (_version_, score, shard, index, timestamp, extract_timestamp, source_modified) are dropped; everything else is kept as returned.

Row: lccn, source (the loc.gov URL), fetchedAt, and item (the LOC object). Useful item fields:

• title, other_title, number_lccn, number_oclc, number_issn, call_number
• essay (HTML text), essay_contributor (who wrote it; keep the credit), number_hasessay
• dates_of_publication, number_first_issue / number_last_issue (issue URLs), number_issue_count
• publication_frequency, created_published, notes (frequency, volume ranges, publisher, microfilm provenance), description
• subject_headings, subjects, genre, language
• location, location_city, location_state, location_county, location_country, latlong
• preceding_titles, succeeding_titles, number_preceding_items, number_succeeding_items (lineage)
• holding_data, number_institution, number_holding_type (which libraries hold it, and in what format)
• rights, partof_collection, batch, digitized

Not every title has every field. Most (but not all) have an essay; check number_hasessay.

batches.jsonl.gz — the OCR batch inventory

LOC's bulk-data inventory of the OCR batches: code, issueCount, pageCount, lccns (titles in the batch), bytes, archiveUrl, sha256. Made by provider:cron-america:inventory. A batch is a delivery, not a title: one batch can hold several titles and a title can span several batches. The sha256 is the checksum we verified every archive against before reading it.

page-records/<lccn>.jsonl.gz — page records

The LOC collection-search record for each page of a title, 1,000 per request (/collections/chronicling-america/?q=<lccn>&fo=json&c=1000), kept to the fields we use: id, date, title, number_lccn, number_edition, segmentof (the issue), original_format, language, partof_title, partof_collection, contributor, location*, publication_frequency, image_url (the IIIF image-service segment, from which every JP2/PDF/ALTO URL is derived). Made by dataset:cron-america:metadata. These are page records from the search index, which is not the same document as an issue record from the item API.

listings/<archive>.listing.jsonl.gz — page statistics (version 3)

What: one row per page of an OCR batch, measured from the ALTO XML. How made: provider:cron-america:list-archive streams the batch (lbzip2 | tar), reads each page's ALTO once and verifies the archive's SHA-256 on the way through. The first line is a header (version, archive, sha256, bytes, pages, tool, decompressor, tarMembers, duplicatePages, geometry, builtAt, host); the rest are pages:

issue, lccn, date, edition, sequence, bytes, occurrences (duplicate page names in the archive are recorded, not dropped), altoWidth, altoHeight, altoUnit, quality (LOC's own Normal/Damaged flag, when present), position, pc (page confidence), imageNumber, printSpace, software, softwareVersion (the OCR engine), maxFontSize, sourceImage, words, lines, textBlocks, composedBlocks, hyphens, illustrations, wordConfidenceMean, lowConfidenceWords, xxh128 (hash of the ALTO file, to tell a revised page from an unchanged one), textBytes, reel, scanKey (<lccn>/<yyyymmdd><edition>/<image>, the join key to LOC's scan paths).

Notes: LOC's quality flag is missing on whole batches (those pages are unrated, not undamaged), and damage is concentrated in a few batches. reel is only recoverable for some batches.

geometry/<archive>.geometry.jsonl.gz — raw layout measures (version 2)

Produced in the same pass as the listing, so no page is read twice. Per page (issue, sequence):

• lh: words by line height, in permille of page height
• lx: words by left edge, in percent of page width
• lw: words by line width, in percent of page width
• fs: words by reported font size
• top: the six largest lines in the top third, as [x, y, w, h, words, text] in permille

These are raw measurements; interpretation is separate so it can be changed without re-reading archives.

layout/<archive>.layout.jsonl.gz — derived layout profile

Computed from the listing and geometry (never from the archive) by LayoutInterpreter; recomputable. One row per issue: issue, lccn, date, profile (issue roll-up) and pages, each with body (body text height), display (headline sizes), columns, columnEdges, columnWidth, columnConfidence, masthead (the nameplate box, anchored on the issue's own date line when the page text has it), and inches (paper size, from the ALTO units).

Rules are deliberately simple and visible in the code (minimum 150 measured words; body = the peak of line heights; columns = left-edge peaks; a second title-sized line must be at least 0.6× the first, so headlines are not mistaken for the masthead). They were checked against hand-reviewed pages, not proven. Expect errors on advertisement pages and unusual layouts; measurable: false marks pages with too little text to say anything.

title-stats.jsonl.gz and layout-titles.jsonl.gz — roll-ups per newspaper

title-stats: per LCCN, batches, batchesTotal, complete (whether every batch holding the title has been read), issues, pages, first, last, words, ratedPages, damagedPages, damagedShare (of the rated pages), pageConfidence, wordConfidence, scanKeys, engines. layout-titles: per LCCN, the modal columns and columnsAgreement (a paper that changed layout mid-run shows about 0.5), body and headline sizes, mastheadFound, masthead position, paper inches.

How to read it

from datasets import load_dataset
titles = load_dataset("museado/newspapers", "titles", split="train")
pages = load_dataset("museado/newspapers", "page_records", split="train")

zcat title-records.jsonl.gz | jq -r 'select(.item.essay) | [.lccn, .item.title] | @tsv' | head

Limitations

• Statistics describe the OCR text LOC supplies, not the printed page. A high confidence can still be a poor scan; a low one can be a clean scan in an unusual typeface (e.g. Yiddish, Fraktur).
• The title record is a snapshot (fetchedAt); LOC edits its catalogue.
• Layout profiles are heuristics (see above). Do not treat a masthead box as ground truth.
• Coverage is the 2,997 OCR batches LOC publishes in bulk, not every digitized page in Chronicling America.

Licensing and credit

• The newspapers in Chronicling America are, in the Library of Congress's words, in the public domain or have no known copyright restrictions; see the LOC rights statement in each record's rights field. The page records are LOC's catalogue facts about them.
• The numbers we computed (listings, geometry, layout and the roll-ups) were computed by Museado and are offered under CC0.
• The title essays (the essay field of title-records) were written by named contributing institutions (state historical societies and libraries) for the National Digital Newspaper Program and published by the Library of Congress. The LOC record carries no separate licence for them: its rights statement covers the newspapers. They are included here, with their credit (essay_contributor), for research and reference. They are not covered by our CC0 dedication; keep the credit when you reuse them. If you are a contributor and want an essay removed or its credit changed, contact the Museado project through https://museado.org and it will be done.

This is why the repository's licence is other and not a single standard licence.

Citation and provenance

Please cite the Library of Congress as the source of the underlying collection, and this repository for the derived files. Each derived file's header records the tool, version, host, build time and the SHA-256 of the source archive. Code: the harvest application, provider:cron-america:*.