The product of this repository is stable, citable addresses. Structured facts and translations live elsewhere and cite into this one. Everything here exists to make those addresses resolvable, unambiguous, and correct.

Addresses

turgot-1739                     a record
turgot-1739.p0                  its first page — the atlas key sheet
journal-de-paris-1789-01-03.p1  page 1 of that day's issue

Ids are flat and never contain /, so directories are cosmetic: folders can be reorganised without breaking a single citation. Ask the tool what an address points at:

cd tools && cargo run -- resolve turgot-1739.p0

The model

Four layers, named after their TEI equivalents. layer says which; type says the genre within it (newspaper, map, issue, supplement, engraving, …).

`layer`	what it is	example
`source`	the publication or work, independent of any copy	`journal-de-paris`
`copy`	one physical object that was digitised	`journal-de-paris-1789-vol1`
`document`	the citable intellectual unit	`journal-de-paris-1789-01-03`
`page`	one scanned side, carrying graphics	`…-01-03.p1`

Layers may be skipped: of points at any ancestor, so a one-off engraving is a single document file and does not pay four files of ceremony. A page is never a file of its own — it is declared inline as [[page]] in the record that owns it.

A file states only what is new. [rights], [[resp]], language, scan.file and the rest are inherited down the of chain; date, title, note, url and [[page]] are not.

Layout

sources/
  frc/                                      the Newberry French Revolution Collection
    ab/abondance00unse/
      abondance00unse.toml                  source, one per Internet Archive item
      abondance00unse.p1.ocr.md             OCR, one file per page
      abondance00unse.pdf                   the scan
    …38,377 items across ~700 two-character shards
  journal-de-paris/
    journal-de-paris.toml                   source
    1789/
      journal-de-paris-1789-vol1.pdf
      journal-de-paris-1789-vol1.toml       copy
      journal-de-paris-1789-vol2.pdf
      journal-de-paris-1789-vol2.toml       copy
  hardy/
    hardy.toml                              source, the eight-volume manuscript
    hardy-vol1.toml                         copy, 658 pages inline as [[page]]
    hardy-vol1.pdf                          the bound volume; pages address into it
    …hardy-vol2 … hardy-vol8
    wikimedia/                              GENERATED by `scans wikimedia hardy`
      hardy-1-work.qs … hardy-3-links.qs    QuickStatements, staged
      hardy-vol1.commons.wiki               one Commons file page per volume
  turgot/
    turgot-1739.toml                        source, 21 sheets inline as [[page]]
    turgot-sheet00.jp2 … turgot-sheet20.jp2
  verniquet/
    verniquet-1795.toml                     source, 73 sheets inline
    verniquet-sheet00.jp2 … verniquet-sheet72.jp2
schemas/source.json                         GENERATED from tools/src/model.rs
schemas/ocr.json                            GENERATED from tools/src/ocr.rs
tools/                                      the `scans` CLI
docs/                                       the design spec

Binary assets are in Git LFS.

The tool

The scans CLI is not in this dataset — it lives in the source repository at https://github.com/peasanttide/scans. What is here is the archive it produces: the records, the OCR, and the page images. Everything below describes what that tool does, for readers who want to regenerate or extend the archive.

tools/ in the source repository is a Rust crate producing the scans binary. One definition in tools/src/model.rs drives three things at once: Rust type checking, TOML deserialisation, and — via schemars — schemas/source.json. The schema is never hand-written, or it would drift from the validator and start lying.

cd tools
cargo run -- validate            # the ten semantic checks
cargo run -- validate --strict   # warnings fail the build too
cargo run -- resolve <address>   # what does this citation point at?
cargo run -- schema --check      # has the schema drifted from the types?
cargo test

validate never opens an image by default — the archive holds several gigabytes of JPEG 2000. The checks that read bytes (real image dimensions, real PDF page counts) are behind an optional feature:

cargo run --features probe -- validate --probe

Pixels

Two commands turn a citation into an image. Both take an address, so nothing ever names a page number inside a PDF by hand.

cargo run --release --features render -- render journal-de-paris-1789-01-03
cargo run --release --features render -- render journal-de-paris-1789-01-03.p1 --grid 2x2
cargo run --release --features render -- extract journal-de-paris-1789-01-03.p1

render rasterises the page. extract writes out the bitmaps stored on it, untouched. Output goes to .render/ at the repository root, which is gitignored: these are derived files, reproducible at any time, and at native resolution they outweigh the PDFs.

Resolution is native by default — the scan's own, taken from its largest embedded image. Journal de Paris pages come out around 3500x4900. --dpi overrides it.
--grid RxC cuts the page into tiles with --overlap percent of slack on the interior edges. A whole page shown to a reader that samples it down to fit is illegible type; for this paper's two-column setting, --grid 2x2 puts roughly one half-column in each tile.
Both are behind the render feature, which is what pulls in the rasteriser. Build --release: a debug build is about thirty times slower, and it shows over 888 pages.

hayro does the decoding rather than lopdf, because every page image in volume 1 is JBIG2 and volume 2 adds JPEG 2000 plates. It is pure Rust, so there is no Ghostscript, pdfium or C toolchain to install.

Each record carries a #:schema directive on its first line, which is what gets it validated in an editor. See schemas/README.md — the alternative, an editor schema association, was tested and silently does nothing.

Binding

render and extract take pages out of a container. pdf goes the other way: it binds a record's loose page scans into one, embedding the images without touching a pixel.

cargo run --release --features pdf -- pdf hardy-vol1
cargo run --release --features pdf -- pdf hardy          # every copy under the source

Page order comes from the record's [[page]] array, never from sorting the directory — a record is the thing that gets corrected when a leaf turns out to have been shot out of sequence, and a directory listing would ignore the correction. Page size comes from the image's own JFIF density, so the object states the physical size it really has.

Gallica serves progressive JPEG and PDF's DCTDecode is specified against baseline, so each stream is transcoded through jpegtran, which re-emits the existing DCT coefficients sequentially. That is lossless, not a re-compression: no requantisation, no generation loss, identical decoded pixels. The claim is checked rather than trusted — every page is decoded on both sides with zune-jpeg and compared, and a single differing byte fails the build.

jpegtran is an operator tool, like git-lfs: install libjpeg-turbo-progs (Debian), libjpeg-turbo-utils (Fedora) or jpeg-turbo (Homebrew). It is not linked against, so the crate stays pure Rust.

Wikimedia

wikimedia turns a source and the copies beneath it into the edits somebody would otherwise type by hand. It opens no socket and uploads nothing: the output is files, to be read and then pasted.

cargo run -- wikimedia hardy
cargo run -- wikimedia hardy --work Q12345678    # once stage 1 has been run

Three QuickStatements batches and one Commons file-description page per copy, written to wikimedia/ beside the source record. The batches are staged because QuickStatements v1 can address a new item only as LAST, so one batch cannot both create a work and create the volumes that point at it; $WORK and $VOLn mark where a Q-id belongs.

Every property and item id lives in wikimedia::vocab with the label it carried when it was verified, and every generated statement prints that label in a trailing comment. Property numbers are not memorable and are easy to get subtly wrong — P7101 is privacy policy URL, not Gallica ID — so a mistake shows up in review as a wrong word rather than a wrong number.

OCR

Word-level OCR lives in a sidecar beside the record it belongs to, one markdown file per scanned page, reached through the [[text]] that points at it.

---
of: 1789iemilseptcen00unse
page: 6
engine: ABBYY FineReader 11.0
lang: fr
w: 2597
h: 4418
dpi: 500
---

<the page's text>

Frontmatter says which page of what, and how big the image is. The body is the page. That is the whole format.

Scalars are bare unless they would lie. lang: fr, not lang: "fr" — the frontmatter is meant to read as YAML. But no is the ISO 639-1 code for Norwegian and, to every YAML 1.1 parser, the boolean false; a Python user loading that with PyYAML gets False. So a value is written bare when it round-trips as itself and quoted when it would not, which keeps the common case clean and the pathological case correct.

What is deliberately not here: word boxes, per-word confidence, baselines, and DjVu's nesting. An earlier version carried all of it and could write the source XML back out unchanged. The boxes were seven eighths of the bytes — 8.3 GB against about 1.2 GB for the text — and were not being used. Nothing is destroyed by that: the coordinates are still in XML_for_OCR/ in frc-data, one clone away, so a later version that wants them can have them without asking archive.org for anything.

cargo run -- ocr import x_djvu.xml --out ./pages
cargo run -- ocr check sources/frc/ab/abondance00unse

The French Revolution Collection

38,377 Internet Archive items from the Newberry Library, listed by NewberryDIS/frc-data. Records and OCR come from a local clone of that repository; only the PDFs are fetched.

git clone --filter=blob:none --no-checkout https://github.com/NewberryDIS/frc-data.git
cd frc-data && git sparse-checkout set --cone Metadata XML_for_OCR && git checkout master

cargo run --release -- frc ingest --frc-data ../frc-data      # records + OCR, ~3 minutes
cargo run --release --features fetch -- frc fetch             # the PDFs, ~21 hours

frc fetch makes one request at a time, a second apart, with a User-Agent naming this repository, and verifies every download against the size and MD5 archive.org publishes for it. It resumes from what is already on disk, so it can be killed at any moment.

Licence

The database — the records, the OCR sidecars, the structure and the addresses — is under the Open Database License (ODbL 1.0). Use it, adapt it, build on it; if you publish a derived database, share it alike and keep the attribution.

The contents are a separate question, and mostly not ours to license. Every work here is an 18th-century imprint or a 19th-century study of one, long out of copyright everywhere, and the digitisations are published as public domain by the institutions that made them. Each record says so for itself:

[rights]
work        = "PD-old-100-expired"
scan        = "PD"
attribution = "Digitised by the Internet Archive, sponsored by The Newberry Library."

So: quoting a pamphlet, reprinting a page image, or lifting one item's OCR carries no obligation from us — that material is public domain and stays public domain. ODbL bites on the collection: the selection, the identifiers, the metadata mapping and the structured OCR taken together. That distinction is the point of using a database licence rather than a content one, and it is why the per-record [rights] fields are not redundant with this section.

Scans and metadata originate with the Newberry Library and the Internet Archive, and the item list comes from NewberryDIS/frc-data.

Design

docs/superpowers/specs/2026-08-05-source-archive-schema-design.md is the authority: the four layers, the inheritance allowlist, EDTF dates, and the ten validation checks, with the reasoning for each.

Downloads last month: 5,360

Total file size:

108 GB