Datasets:
Commit History
-
docs: the README describes the archive as it is now laid out
Three things had drifted. The counts: 39,324 PDFs and 137 GB were the figures before the last two fetch runs, and the file list has grown a tail of JP2, JPEG, PNG and DjVu the line never mentioned. It is 39,668 files and 155 GB, and the per-collection table is recounted from the records that name the files, since the archive itself no longer says which collection anything belongs to. Which is the second thing. The table read as a directory listing — `frc`, `marat`, `hardy` — and said the Newberry scans were "sharded two deep by the item's Internet Archive identifier". Neither is true since the flattening: there is one `media/` tree and the shard is the md5 of the filename. The README said the directories were cosmetic without ever saying what decides them, which left the one rule a reader needs to find a file undocumented. Third, "How the PDFs are made" described the binding pipeline as though every file went through it. Most did not: 38,377 of them are whole-work PDFs taken as the I…eb7a858 -
feat: the five digitisations behind the Henriot dating dossier
Chassin tomes I and III, Bailly's 1804 tome I, Ferrières tomes I-III and Réveillon's own Exposé justificatif. Tuetey is not here: the archive already holds tome I as the Getty copy, and a second photograph of the same printed volume is not worth 51 MB. Every file was verified against the Internet Archive's own size and md5 before it was placed on the shard. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>19bc9f3
-
245 more scans, fetched from what the corpus named
LizeLive +1: Claude Opus 5Internet Archive 143, Wikimedia Commons 94, Gallica 8. Every one sharded on the md5 of its filename like the rest; every one manifested in the corpus. Eight of them are DjVu, an extension `.gitattributes` had never seen, so they went in as raw blobs and Hugging Face refused the push — five were over its 10 MiB limit. `*.djvu` and `*.tif` are tracked now, and the eight are pointers. 39,663 files. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KNRQDxd9kCdeWWfgdLGXkidf22847 -
Flatten the archive and shard it on the md5 of the filename
LizeLive +1: Claude Opus 5The collection folders were cosmetic and said so, but they still encoded a guess — that a file belongs to one collection — and that stopped being true the moment a number of Marat's paper arrived from two libraries. A file now lives at `media/${ab}/${cd}/${filename}`, where ab and cd are the first two and second two hex characters of the md5 of its filename, which is how Wikimedia shards media and has the property that matters: the path is derivable from the name alone, so a file is found without an index and without knowing where it came from. 39,422 files, 39,418 of them hashed and manifested. Filenames were checked for collisions before anything moved — there are none, which is what makes a flat namespace safe. What describes these bytes is not here. It is in the corpus, at `scans/${ab}/${cd}/${filename}.json` — one manifest per file, sharded by the same key, holding what a Wikimedia Commons upload needs: where the file came from, what licence covers the work and what covers the p…3d968f0 -
The archive holds scans; the records moved to the corpus
LizeLive +1: Claude Opus 5880,483 files leave this repository and none of them was a scan. 840,810 pages of OCR, 38,725 bibliographic records, 1,743 files of plain text, the design specs, the schemas and the `scans` tool itself — all of it describes the facsimiles rather than being one, and all of it is now in the corpus that cites this archive, at github.com/peasanttide/history: the records in data/, the text in sources/, the tool in scans-tool/, the specs in docs/archive/. Nothing was deleted that had not been verified present there first, file by file: 38,725 of 38,725 records and 840,810 of 840,810 OCR pages. And nothing is gone in any case — it is all in this repository's own history. What remains is 39,324 PDFs and 94 JP2 map sheets, 137 GB, and a README that describes them. The split is worth stating rather than assuming. A facsimile is stable and large; a record about it is small and gets corrected — the Hardy record was corrected three times today. Keeping them apart means this can be mirrored, pinn…da080c1
-
feat: Jefferson tells Jay the commons have named themselves
LizeLive +1: Claude Opus 5The Library of Congress serves mtj1.011_0481_0487 as seven master JPEGs and a page viewer; this binds them into one object and gives them an address. jefferson-to-jay-1789-06-17.p7 resolves to the last leaf. The letter is mostly the survey of European courts its docket calls a summary of foreign affairs, and would be a marginal thing to hold here on that alone. The postscript is why it is in: written the next morning, it reports the vote of 17 June 1789 — "a majority of 500 and odd against 90 odd" — from a member who had left the hall before the question was put and whose brother stayed to hear the decision. That is the Third Estate constituting itself the National Assembly, reported within a day, by someone with no stake in how it read later. One record, no copy or document layer: a single letter that was digitised once does not need three files to say so. The seven leaves are inline as [[page]], each carrying its own id, its dimensions, the viewer page it came from and the master U…a0dbd99
-
docs: linearising a volume past 2 GB breaks the reader that opens it
LizeLive +1: Claude Opus 5Tried, measured, reverted. Recording the result because the idea is sound and will otherwise be had again. Linearisation itself was faultless: all 4,230 image streams byte-identical afterwards, /L exactly the real file length, qpdf --check reporting no syntax or stream errors, and page one reachable in the first 2.6-4.2 MB of a 2.1-2.6 GB volume instead of the whole thing. It still had to go. A linearization table states /L, /T and its hint-stream offsets as byte positions, and past i32::MAX those stop fitting the 32-bit int some readers keep them in. hayro — which this archive's own render and extract use — reported a linearised 2.5 GB volume as ONE page. poppler emitted 21 warnings per open before recovering through the main xref. Ghostscript was unaffected, and the same qpdf invocation on a 12.8 MB slice of the same volume upsets nobody, which is what pins it to size rather than to malformed output. Breaking the decoder the archive's tools depend on, to speed a first-page load fo…03dacbd -
chore: regenerate the Wikidata batches against Q141104178
LizeLive +1: Claude Opus 5The work item exists now, so stage 1 addresses it instead of CREATEing one and can be re-run: QuickStatements skips a statement the item already holds and adds the rest. Stages 2 and 3 carry the real Q-id rather than $WORK. Generated output, regenerated by `scans wikimedia hardy --work Q141104178`. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>4da0710
-
refactor: Hardy is a bound volume, not a folder of loose leaves
LizeLive +1: Claude Opus 5The eight PDFs move up to sources/hardy/ beside their records, the 4,230 page JPEGs are removed, and every [[page.graphic]] now addresses a page inside the volume rather than a file of its own: - file = "hardy-vol1-p001.jpg" + page = 1 This is the shape journal-de-paris already uses: [scan] file names the container, and a graphic gives its 1-based index into it. width and height stay, because they are the real pixel dimensions and --probe checks them. The IIIF url stays on every page too — with the local copies gone, that is the route back to the original, and it matters more now, not less. Repointing was not optional. Deleting the images while the records still named them would have left all 4,230 page addresses dangling at files that no longer exist. The JPEGs remain in history and are recoverable; note that removing them here does not reclaim their LFS storage, which is still referenced. Verified: validate --probe reads the real PDFs and reports no finding against an…33b54ef -
chore: ignore .hf-upload.log
LizeLive +1: Claude Opus 5Another derived run log, alongside the four already listed. The line was written once before and lost when the local rename commit was dropped in favour of the three the remote already had. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>78582e6 -
docs: the generated README states the licence tag it actually emits
LizeLive +1: Claude Opus 5The Commons pages use {{PD-scan|PD-old-100-expired}}, one template taking the work's tag as its argument. The README beside them described two separate tags, which is what an earlier draft emitted. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>3e28b7f -
feat: bind each Hardy volume into one PDF
LizeLive +1: Claude Opus 54,230 loose Gallica JPEGs become eight PDFs, one per manuscript volume, with the image streams embedded rather than re-encoded. Page order comes from each record's [[page]] array; page size from the images' own 400 dpi JFIF density. Gallica serves progressive JPEG and DCTDecode is specified against baseline, so each stream goes through jpegtran, which re-emits the existing DCT coefficients sequentially. Lossless: no requantisation, no generation loss. Every page was decoded on both sides with zune-jpeg during the build and compared, and streams pulled back out of the finished files were checked against the source JPEGs again — identical throughout. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>9d72393 -
feat: Wikidata batches and Commons pages for Hardy
LizeLive +1: Claude Opus 5Nine items' worth of QuickStatements — the work plus its eight manuscript volumes, none of which Wikidata holds today — and one Commons file-description page per volume. All generated from the records by `scans wikimedia hardy`, none of it uploaded: the output is files to be read before they are pasted. The batches are staged in three 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 was checked against the live API rather than written from memory, which caught two: P7101 is privacy policy URL, not Gallica ID (P4258), and Q174728 is centimetre, not millimetre (Q174789). Every Commons template was checked for existence, which caught a third — there is no Creator:Siméon-Prosper Hardy, so the author line is a plain name rather than a red link. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>25656af -
design: Hardy volume PDFs and Wikimedia records
LizeLive +1: Claude Opus 5Eight volumes of loose Gallica JPEGs become one PDF each, with the streams embedded unmodified. Gallica serves progressive JPEG and DCTDecode is specified against baseline, so each stream is transcoded through jpegtran — lossless, the DCT coefficients untouched — and the builder verifies pixel identity rather than trusting the flag. Alongside: a QuickStatements batch for nine new Wikidata items (the work plus its eight manuscript volumes, which Wikidata has none of today) and eight Commons file pages. Both generated from the records, both written to disk for review rather than pushed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>88f715d
-
rename: drop the old underscore sheet paths
Verified signatureThe sheets are now turgot-sheetNN.jp2 and verniquet-sheetNN.jp2, matching their ids. These are the paths they were uploaded under before the rename.a433350 -
rename: map sheets are verniquet-sheetNN
Verified signaturebd6f72d -
rename: map sheets are turgot-sheetNN
Verified signature0fd51a0 -
fix: page ids match their filenames
Verified signature68f5b9c -
fix: page ids match their filenames
Verified signature4ced16b -
fix: page ids match their filenames
Verified signature9d2a7ff -
fix: page ids match their filenames
Verified signature065bd46 -
fix: page ids match their filenames
Verified signaturec0bb3b9 -
fix: page ids match their filenames
Verified signature4519486 -
fix: page ids match their filenames
Verified signature42c6a2b -
fix: page ids match their filenames
Verified signature0739a01 -
fix: page ids match their filenames
Verified signature56effba -
fix: page ids match their filenames
Verified signaturecfd0cae
-
Add files using upload-large-folder tool
Verified signature1121af9 -
Add files using upload-large-folder tool
Verified signature7319f05 -
Add files using upload-large-folder tool
Verified signature91de0f1 -
Add files using upload-large-folder tool
Verified signature0972314 -
Add files using upload-large-folder tool
Verified signature443c23e -
Add files using upload-large-folder tool
Verified signature946f0bc -
Add files using upload-large-folder tool
Verified signature452c491 -
Add files using upload-large-folder tool
Verified signature8e68c8c -
Add files using upload-large-folder tool
Verified signaturebefeaf3 -
Add files using upload-large-folder tool
Verified signature44e5530 -
Add files using upload-large-folder tool
Verified signaturee33d1b6 -
Add files using upload-large-folder tool
Verified signaturec77f15f -
Add files using upload-large-folder tool
Verified signature9ff1630 -
Add files using upload-large-folder tool
Verified signature3ac53b4 -
Add files using upload-large-folder tool
Verified signaturee75ac9d -
Add files using upload-large-folder tool
Verified signature3140e9a -
Add files using upload-large-folder tool
Verified signaturedf49624 -
Add files using upload-large-folder tool
Verified signature5dcd94d -
Add files using upload-large-folder tool
Verified signaturea10269c -
Add files using upload-large-folder tool
Verified signature5a57353 -
Add files using upload-large-folder tool
Verified signaturea4662c5 -
Add files using upload-large-folder tool
Verified signaturea8517e8 -
Add files using upload-large-folder tool
Verified signature2b25746