SKocur/polish-pre1918-corpus

Polish pre-1918 corpus

294,369 documents, 22.77 GB of Polish text catalogued as published between 1800 and 1918, assembled by a fully code-driven pipeline for training temporally bounded ("time-capsule") language models from scratch. Every filter is a documented function, each cleaning step writes a machine-readable report with the hash of the script that produced it, and the corpus's defects are quantified in those reports.

This is the corpus behind the wieszcz-xix model ladder (47M / 107M / 349M parameters).

⚠️ Read Content warning and Rights before use. This is nineteenth-century print. It contains antisemitic, nationalist and colonial discourse that was ordinary in its time.

What is in it

Documents	294,369
Text	22.77 GB (UTF-8)
Tokens	6,745,956,943 under the corpus's own 8k byte-level BPE
On disk	10.93 GB in 43 shards as zstd-compressed parquet
Period	catalogued publication dates 1800–1918, 1918 included; 90 Wolne Lektury documents are later text (see Known limitations)
Language	Polish, original orthography preserved in the Internet Archive text (no normalisation to post-1936 spelling); the Wolne Lektury transcriptions are in modernised spelling
Sources	Internet Archive 291,655 documents (22.60 GB) · Wolne Lektury 2,714 (0.17 GB)

Fields, one row per document:

field
id	stable document identifier, joins to the provenance ledger
source	internet_archive or wolne_lektury
source_identifier	the identifier at that source, so any document can be traced back
text	the cleaned full text

Catalogue metadata for the Internet Archive documents (title, date, place of publication, type, creator and rights statement, as retrieved on 2026-10-06) is in the code repository at metrics/ia_catalogue_metadata_2026-10-06.jsonl.gz, keyed by the same identifier. ledger/document_metadata_2026-10-06.csv.gz in this repository joins it to every document as one table (split, provider, year, place, type, title, language, rights basis).

There is a single train split. The document-level train/validation split used by the paper is not applied here; it is published separately as identifier lists, so a user can reproduce it or ignore it. Applying it here would hand every reader one particular experiment's split as if it were a property of the corpus.

This release is exactly what the models were trained on

The corpus published here is the frozen build, document for document: 294,369 documents, the same stream the wieszcz-xix ladder was trained on. Nothing was removed after training and nothing is held back.

This holds because the temporal audit ran before tokenisation, so the documents it excluded never entered the training stream. exclusions/exclusions_2026-08-03.json lists the 3,733 documents removed at that point, 2.47% of the pre-exclusion crawl by bytes, under a rule that trades precision for safety. The list ships with the data, so the decision can be audited.

The exclusion battery is calibrated against period-legitimate controls. Czołg, lotnisko and samolot are attested by 1918, telefon appears in 72,719 documents of the raw freeze and aeroplan in 6,391, and hand-reading candidate contexts removed more terms from the battery than it admitted.

How it was built

1. Crawl: Polish text published 1800–1918 from the Internet Archive and Wolne Lektury, queried by publication-date range and frozen with a manifest.
2. Filter: OCR-noise and layout heuristics, each a documented function with a report.
3. Deduplicate: near-duplicate removal at the document level.
4. Temporal audit: a content battery (anachronistic vocabulary, dated self-reference) calibrated on its own false alarms against period-legitimate controls, plus a provenance check. 3,733 documents excluded before tokenisation.
5. Freeze: a per-document provenance ledger recording identifier, source and byte length, hashed.

Every step's report is in the code repository, and each carries the SHA-256 of the script that wrote it.

The ledger

ledger/provenance_ledger_2026-08-03_{train,val}.csv.gz records every document in the frozen corpus: document_id, source, source_identifier, bytes. If the full text ever has to come down, the ledger is what makes the corpus reconstructible from the sources. It is also what this release's self-check verifies against: each document is decoded from the frozen token stream and its UTF-8 length compared to the ledger, so a truncated shard or a misaligned stream fails the build instead of shipping unnoticed.

Sizes in the ledger are what the tokeniser saw. For Wolne Lektury that differs from the file on disk by the digitisation colophon, which tokenisation drops.

OCR quality

OCR quality was estimated on twenty passages of about 400 characters, drawn with a fixed seed from twelve Internet Archive documents and corrected by hand from context, without the page images. Period orthography is treated as correct: blizko, Jenerał, terrytorjum and niczem are right for the period, and counting them as errors would overstate the noise.

Eleven passages could be corrected: CER 0.68%, WER 4.71%. Nine, 45% of the sample, could not be corrected at all.

In those nine passages the corruption removes text instead of garbling it: columns interleaved line by line, a fragment carried in from a neighbouring article, a clause ending mid-word. Context does not determine what the page said. 0.68% is the error rate conditional on the text being legible as running prose, and that condition fails on nine of the twenty passages.

The sample is twenty passages read by one annotator.

Content warning

This corpus contains antisemitic, nationalist, colonial and misogynist discourse, at times as the explicit subject of a text and at times as unremarked background. It is present in the press material especially. The prejudice comes from the material itself; the cleaning did not introduce it and was not designed to remove it.

Models trained on it reproduce that content. If you train on this corpus, that is a property of what you build, and it is yours to disclose.

Rights

The Internet Archive texts are catalogued as published from 1800 to 1918. The catalogue records of 93% of those documents (84% of their bytes) carry the holding library's public-domain statement. Author death dates were not checked document by document, so a work whose author died after 1955 may still be protected where the term is life plus 70 years. The catalogue names a creator who died in 1956 or later for 915 documents, 849 of them periodical issues whose named creator is an editor. The Public Domain Mark applies where the holding library states public domain, and to the Wolne Lektury documents that library records as public domain. The other Internet Archive documents (7% of them, 16% of the bytes) carry no rights mark. 79 Wolne Lektury translations are still in copyright and keep the licence that library publishes them under, CC BY-SA 3.0 or the Free Art Licence 1.3. The per-document table in this repository (ledger/document_metadata_2026-10-06.csv.gz) gives each document's rights_basis (library_public_domain_statement, scanning_institution_assertion, wolne_lektury_public_domain_note, free_licence or none) and names author, translator and licence for the Wolne Lektury rows. CC0 is asserted over the compilation layer (selection, cleaning, arrangement), which is the part that could attract the EU sui generis database right.

Redistributing a harvested collection is also bound by platform terms of use and by that database right. The per-source rights review is archived with the provenance record (doi:10.5281/zenodo.22099302).

Takedown: if you hold rights in a work you believe is included in error, write to legal@szymonkocur.com or open an issue in the code repository, and it will be removed. The full text lives here and not in the archival record for exactly this reason: a Zenodo version is effectively immutable, so the archival deposit (doi:10.5281/zenodo.22099302) holds only facts (identifiers, checksums, reviews) while the texts sit where a verified claim can actually remove one. The temporal audit has a known residue, 90 to 1,116 Wolne Lektury documents (0.04 to 0.38% of the corpus bytes, see Known limitations), and a residual risk beyond it.

Not legal advice. For publication or commercial use, consult a specialist.

Known limitations

• The temporal bound is audited and carries a residual risk. The audit removes 2.47% of the freeze by bytes, and after cleaning every strong marker counts zero, but post-1918 text that avoids modern vocabulary and dated self-reference evades it.
• Coverage is inherited. Whatever the digitising institutions under-represent (by region, publisher or genre) this corpus inherits in the same proportion. Four fifths of the text comes from the Jagiellonian Digital Library, and 62% of the Internet Archive documents were published in Austria-Hungary under the borders of 1914.
• Polona's press holdings are absent, because its OCR endpoint is authentication-gated. 94.2% of the Internet Archive documents are periodical issues, and they are the ones the remaining libraries digitised.
• The Internet Archive dominates, at 99% of documents. Wolne Lektury contributes clean transcriptions but a small fraction of the text.
• The Wolne Lektury part holds later text. It is dated by the library's epoch for the work. 696 of its 2,714 documents are translations, and 90 documents (0.04% of the corpus bytes) are certainly later than 1918: 79 translations still in copyright, which the library publishes under CC BY-SA 3.0 or the Free Art Licence 1.3, and eleven volumes of Proust in Boy-Żeleński's translation. Counting every translation or work by someone who lived past 1918 gives an upper bound of 1,116 documents, 0.38% of the corpus bytes. 141 documents are in other languages. The per-document table in this repository marks each of them.

Loading

from datasets import load_dataset

ds = load_dataset("SKocur/polish-pre1918-corpus", split="train", streaming=True)
print(next(iter(ds))["text"][:500])

At 22 GB, streaming=True is usually what you want.

Citation

If you use this corpus, cite the paper, arXiv:2610.10592:

@misc{kocur2026wieszcz,
  title  = {Wieszcz-XIX: A 3.1-Billion-Word Corpus of Pre-1918 Polish and Temporally Bounded Language Models Trained From Scratch},
  author = {Kocur, Szymon},
  year   = {2026},
  eprint = {2610.10592},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url    = {https://arxiv.org/abs/2610.10592},
}

To cite the exact data release, add the archival record:

@dataset{kocur2026corpus,
  title     = {Polish pre-1918 corpus (wieszcz-xix)},
  author    = {Kocur, Szymon},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22099302},
}

See also

• Models trained on this corpus: the wieszcz-xix ladder at 47M, 107M and 349M parameters, each released at three training points, archived at doi:10.5281/zenodo.22099358 (Apache 2.0).