scans

Facsimiles of primary sources of the French Revolution. Scans, and nothing else — 39,668 files, 155 GB: 39,536 PDFs, 94 JP2 map sheets, 30 loose page images and 8 DjVu volumes.

The records that describe these files, the transcriptions and OCR of what is printed on them, and the tool that fetched and bound them all live in the corpus that cites this archive: https://github.com/peasanttide/history. This repository is the bytes.

That split is the point. A facsimile is stable and large; a record about it is small and gets corrected. Keeping them apart means the archive can be mirrored, pinned and left alone, while the description of it moves as often as it needs to.

What is here

collection	files	size	what it is
Newberry French Revolution Collection	38,377 pdf	83 GB	the Newberry Library's collection, digitised by the Internet Archive. One PDF per item.
Marat's papers	944 pdf	41 GB	every surviving number of Jean-Paul Marat's four papers, 1789–1793, one PDF per issue. Page images from the ARTFL Project at the University of Chicago, photographed from the New York Public Library's seven-volume set; seven issues and one whole number come from Gallica, which has leaves ARTFL's server does not.
Hardy, Mes loisirs	8 pdf	19 GB	Siméon-Prosper Hardy's autograph journal, 1764–1789. Eight manuscript volumes, 4,227 leaves, from Gallica (BnF, Français 6680–6687).
individually-named items	238 files	7.5 GB	single works fetched one at a time because some record came to cite them — memoirs, decrees, pamphlets, and the nineteenth-century editions the corpus quotes documents from. Chassin, Tuetey, Bailly, Ferrières, Morris, Pellet, Jefferson's despatch of 17 June 1789.
Atlas du plan général de la ville de Paris	73 jp2	2.4 GB	Verniquet, 1795.
Plan de Turgot	21 jp2	1.3 GB	1739, from the David Rumsey Map Collection.
Journal de Paris	3 pdf	1.0 GB	the 1789 run of the first daily newspaper in France.
Les Actes des Apôtres	4 pdf	116 MB	nos. 1–119, four volumes.

Those groupings are a description, not a directory structure — see below. Every file in the archive is named by at least one record in the corpus; none is orphaned.

Where a file lives

media/${ab}/${cd}/${filename}

${ab} and ${cd} are the first two and second two hex characters of the md5 of the filename — not of the file's contents, and not of the id of the record that cites it:

md5("4Z2264INV2356FA.pdf") = 3f576d2b…   ->   media/3f/57/4Z2264INV2356FA.pdf

The directories carry no meaning beyond that. They exist to keep any one directory from holding forty thousand entries, and the archive can be reorganised without breaking a citation, because nothing cites a path.

The same hash of the same filename is what joins the three trees that describe a file, so all three land on the same leaf — the other two being in the corpus repository, not this one:

scans   media/3f/57/4Z2264INV2356FA.pdf                  the bytes (here)
history scans/3f/57/4Z2264INV2356FA.pdf.json             the upload manifest
history sources/3f/57/4Z2264INV2356FA.pdf.pNNNN.toml     the text of one page

The record itself shards separately, on the md5 of <id>.toml, and so lands nowhere near its own scan. The string in the record's [[scan]].file is the whole interface between a record and the bytes it describes.

Addressing a page

A file is addressed by the record that names it, over in the corpus. Ids there are flat and never contain /:

lami-du-peuple-1790-07-01.p3     page 3 of that issue
turgot-1739.p0                   the atlas key sheet

Ask the corpus what an address points at, and what is printed on it:

history show lami-du-peuple-1790-07-01 --page 3

How the PDFs are made

Every page is one /DCTDecode XObject holding the JPEG exactly as the library served it — nothing resampled, requantised or colour-converted, because the point of an archival scan is that it is the bytes the holding institution published. The page box comes from the image's own JFIF density, so a leaf shot at a different size stays a different size rather than being normalised into a uniformity the object does not have. Page order comes from the record, never from sorting a directory: the record is what gets corrected when a leaf turns out to have been shot twice or out of sequence.

Where a source serves progressive JPEG — Gallica does — each stream is re-emitted with jpegtran -copy all. -copy none would drop the APP0 density and every page box would become a guess.

Linearisation is skipped above 2 GB. A linearisation table states byte offsets, and past i32::MAX the decoders disagree about what they are looking at: one reports a 2.5 GB volume as a single page, another emits twenty-one warnings per open. Two of the Hardy volumes are affected.

Not every file here was bound by that pipeline. Where a library already publishes a whole-work PDF — the Internet Archive does — the published file is taken as it stands and verified against the size and md5 the publisher states for it, because refetching page images to rebuild a container the holder already made would be a worse copy, not a better one.

Rights

The works are out of copyright. The digitisations are not all offered on the same terms, and the corpus records what each holding institution claims, unchanged, rather than rewriting it to match what Wikimedia Commons would accept. Gallica asserts a non-commercial restriction over its images; the Internet Archive's Newberry scans carry no such claim. The [[copyright]] block on a record says which applies to it.

Binary assets are in Git LFS.