Przejdź do treści

Sourcing the human half

The retired rule pack needed a paired corpus, and the human half was the blocking one. This document is the survey behind that half: what Polish is obtainable, and what each body of it can and cannot support. The track it was written for is gone, and the survey outlives it because the grammar is measured over Polish too: the audit corpus it argues for is where the register's own ambiguity is counted.

The answer it arrives at is that no existing corpus serves. The register olski is scoped to is nearly absent from the corpora that exist, the corpus that would say whether a text is a translation does not record it, and the typography a linter would measure has been normalized out of the largest one. So the good side is assembled from named parts rather than chosen, and it is two corpora rather than one, because a rule reporting a rate wants prose an editor worked on and a rule whose hits get read wants prose nobody touched after the author.

What a corpus has to say about itself

Five things, each demanded by something already decided.

  • Licence. Milestone 1 exits on numbers taken over a corpus anyone can fetch and a run anyone can redo, so a corpus that cannot be redistributed or downloaded can produce a number nobody can check.
  • Size. A threshold is a point in a distribution, and a distribution needs enough text to have a shape. A rule whose hits get read instead of counted needs no shape, so what is a starved corpus for one demand is a working one for the other.
  • Register. A pack is scoped to a register, and a threshold read off a different one accuses the writing it was built to protect.
  • Provenance. The rules against anglicisms and calques need Polish that was written in Polish. A baseline translated from English licenses exactly what they flag.
  • Stage of production. linter.md argues a typographic rule measured against typeset prose is measured against nothing, because the corpus could not have held what the rule looks for.

Stage of production is the column that splits the answer in two, because no body of Polish is both worked on by an editor and untouched since its author, and the two shapes of rule want one each.

How the counts here were taken

Every figure below that is not attributed to somebody else was taken on 2026-08-07, over the files as they stand, by fetching a corpus with the command printed beside it and counting words and characters over what came down. Each count says which files it ran over and which selection of them, so that a second person picks the same text.

One of the corpora comes with the program that did the counting as well. The Markdown one, KSeF, is cloned by the command audit-corpus.md prints rather than by one printed here, because that document owns which repositories the corpus holds. It is extracted by harness/markdown.py:

python3 -m harness.markdown ksef-docs --into proza/ksef

The counting was done by the rule engine that ran the typography pack, which is retired (linter.md), so the figures over it are dated and a second person redoing them writes the count.

The rest are XML, JSONL, PO files and plain-text exports, which that extraction does not read, so their figures come from a program written for this survey that is not in this repository. A figure of that kind is checkable by rewriting the count, which the stated selection makes possible and does not make free, and todo/ holds the question of whether the other formats get an extraction of their own.

These are counts of characters and words rather than rule firings either way, and they are here to characterize a corpus rather than to calibrate anything. Every extraction has a price, and this one's is in extraction.md. The visible half of it here is that dropping a corpus's tables drops what the padding inside them was counted as: KSeF carries 456 runs of two or more spaces as it stands and 5 once extracted, so a survey that skips the step reports table layout as a typing defect.

The National Corpus of Polish

Narodowy Korpus Języka Polskiego is the reference corpus of contemporary Polish, built 2007–2011 by IPI PAN, the Institute of Polish Language, PWN and Łódź. It holds 1.5 billion words, of which a 250-million-word subcorpus is balanced by readership and a 1-million-word subcorpus of that is manually annotated.

Only the 1-million-word subcorpus is downloadable. It carries CC BY, and the rest is reachable through the Poliqarp and PELCRA search interfaces and not otherwise.

curl -L -o nkjp1m.tar.gz \
  'http://clip.ipipan.waw.pl/NationalCorpusOfPolish?action=AttachFile&do=get&target=NKJP-PodkorpusMilionowy-1.2.tar.gz'
tar xzf nkjp1m.tar.gz

Its register composition, from the file it ships

statystyki.txt inside the archive gives the breakdown by text type, over 1,003,956 words in 3,889 samples, and the taxonomy that names each type is in NKJP_1M_header.xml:

type words
typ_publ 495,662 49.4% journalism
typ_lit 164,543 16.4% fiction
typ_konwers 58,459 5.8% conversational
typ_inf-por 54,284 5.4% informative and instructive writing
typ_fakt 52,600 5.2% non-fiction novel
typ_net_interakt 50,510 5.0% interactive internet
typ_urzed 30,539 3.0% legal and official
typ_qmow 25,212 2.5% quasi-spoken
typ_media 23,408 2.3% spoken from the media
typ_nd 19,438 1.9% academic writing
typ_net_nieinterakt 18,937 1.9% non-interactive internet
typ_nklas 9,890 1.0% unclassified non-fiction book
typ_listy 411 0.04% letters
typ_lit_poezja 63 0.006% poetry

Two of those rows could hold olski's register, and neither does. Reading the titles of the 56 typ_inf-por samples in their headers gives dream dictionaries, travel guides, astrology, cookery and self-help: Sny – klucz at 3,577 words is the largest, then Wielkopolska : przewodnik, Przewodnik po Polsce, Wielki sennik współczesny. One sample out of the 56 is documentation of a computer system, TAG 3.1 dla opornych, and it runs to 974 words. The 19,438 words of academic writing are university textbooks — Meteorologia dla geografów, Glacjologia, Podstawy bankowości — and those are the closest thing in the corpus to the expository Polish a pack for technical documentation is scoped to.

So the register is present at 7.3% of the subcorpus if textbooks count, and at about a thousand words if only documentation does.

It does not record whether a text is a translation

Each sample's header.xml carries a title, an author, a publisher, a publication date and the text class above. Across all 3,889 of them, no header names a translator, an editor or a responsibility statement:

grep -l 'respStmt\|<editor\|translat' */header.xml | wc -l   # 0

The question the calque rules ask cannot be answered from this metadata, and answering it by hand means identifying the source of each of 3,889 samples.

Its text layer has been character-normalized

Concatenating the <ab> paragraphs of every text.xml gives 985,308 words carrying 10,055 straight " against 530 „, and 9,535 en dashes against 241 em dashes. The straight marks are in the running text of 1,088 of the 3,889 samples, books and dailies alike, and they are ordinary quotations: autor "Krótkich dni", W "Babim lecie".

Whatever produced that — the source digitizations or the corpus build — it settles what NKJP can be used for: a quote-straight rate measured here measures the encoding. It is also why linter.md asks these rules for prose that reached its reader through no step that rewrote its characters, rather than for prose from before a typesetter.

Two smaller limits. Of the 349,370 words whose header carries a publication date, the median falls in 2000, and the range runs 1957 to 2008, so the corpus predates the prose olski is aimed at by a generation. And the manually annotated subcorpus is the balanced one, so its 49.4% journalism is a design target met rather than an accident to be sampled around.

Wolne Lektury

A library of Polish literary works, in the public domain or under CC BY-SA 3.0 or the Free Art Licence 1.3, with a JSON API and per-work text, XML, EPUB and audio exports. The catalogue holds 7,651 works.

curl -s 'https://wolnelektury.pl/api/books/?format=json'

Three properties of it are in the catalogue's own metadata.

It is mostly verse. 5,275 of the 7,651 works are Liryka, 2,032 Epika and 342 Dramat. A corpus of Polish prose extracted from it is a quarter of what the count suggests.

It is period-marked. By epoch the catalogue runs Współczesność 2,614, Romantyzm 1,056, Modernizm 960, Pozytywizm 817, Dwudziestolecie międzywojenne 726, Renesans 510, Oświecenie 305, Barok 166, Starożytność 137, Średniowiecze 15. That is the trap fiction.md names for cliché mining, and its cause is the licence: a library of public-domain Polish is a library of what copyright has released.

A seventh of it is translated. 1,148 of the 7,651 detail records name a translator, which is 15.0%. The library commissions translations as well as hosting old ones, so this is a standing property rather than a legacy of digitization.

Typographically it is the opposite of NKJP. The first forty Epika entries the catalogue returns — thirty-eight of them, the other two being collections with no text export — are 1,434,945 words carrying 5,591 „ and 5,562 ” against 12 straight ". They also carry 32,782 em dashes, which is 22.8 per thousand words, against the 10 per thousand em-dash-density allows. Polish prose introduces dialogue with a dash, so a literary corpus measures dialogue punctuation on that statistic and not the construction the rule is aimed at — the same confusion as the link lists in generated-polish.md, with an editorial convention rather than markup doing it.

Wolne Lektury is therefore not the human side for a technical pack, on register, on provenance and on period at once. It would be the obvious corpus for a fiction pack, which linter.md files as a wish.

PolEval, and the one dataset in it that pairs

PolEval is an evaluation campaign for Polish running task editions from 2017 onward, each releasing its own data under its own terms — the DiaBiz sample used for punctuation prediction in 2022, for instance, is CC BY-SA-NC-ND, and prose pulled out of a corpus is a derivative of it, so a no-derivatives task set can be measured and cannot be shipped. The campaign is a source of task data rather than of a register-balanced corpus, with one exception, and the exception is aimed at olski's problem.

Śmigiel is the dataset of PolEval 2025's first task, machine-generated text detection for Polish. It is CC BY 4.0, published on Zenodo, and it holds 64,000 postprocessed fragments, 32,000 human-written and 32,000 machine-generated, with a raw corpus of over 462,000 generations behind them.

Its human side is drawn from six domains, each from a named dataset with a permissive licence: literature (the Polish Library of Science Corpus, open Polish coursebooks, and the public-domain texts of Wikiźródła), reviews (PolEmo, Allegro Reviews, Filmweb), social media (TwitterEmo, BAN-PL), Polish Wikipedia, Polish Wikinews, and ParlaMint parliamentary transcripts. Its machine side comes from eight open-weight models in three size tiers, from Bielik-7B to Llama-3.3-70B, each prompted with a prefix taken from a human fragment.

Two things about it matter here, and they pull in opposite directions.

The machine half was generated for the purpose. That is the condition generated-polish.md sets on the generated half and the corpus it measures fails. The pipeline the paper describes is prompting, filtering — about a third of the generations were discarded — and length normalization, with no editing pass in it, so what shaped the surviving fragments was a filter rather than a writer.

Neither half is in the register, and neither half is a document. The six domains do not include technical documentation. The fragments are passages, length-normalized, and every machine fragment continues a human prefix rather than answering a brief, which is not how the prose olski checks is written.

The typographic state of the training split says the same thing with numbers. Over 35,763 fragments, split by the released label file:

fragments words " „ — em dashes per 1000
human 17,873 1,809,311 13,631 3,176 1,308 0.72
machine 17,890 1,739,720 18,472 1,358 587 0.34

Both sit two orders of magnitude below the threshold em-dash-density sets, and the machine half sits below the human one, which is the reverse of the direction generated-polish.md found at 35.4 per thousand over generated Polish written as whole notes rather than as continuations of somebody else's sentence. One of the thirteen human source labels reaches the threshold's neighbourhood: the fragments marked classics run 1,035 em dashes over 105,643 words, or 9.80 per thousand, for the dialogue reason Wolne Lektury gives above. A rate is a property of a register, a writer and a task at once, and a corpus assembled for detection holds none of the three olski is aimed at.

The state also varies by source inside one half: human Wikipedia carries 2,072 „ and 2,045 ” over 678,486 words, which is Polish quotation used correctly, while human Filmweb reviews carry 5,672 straight " against 38 „. Stage of production is a property of each source and not of the corpus.

Polish technical documentation, original and translated

This is the register the linter was scoped to and the grammar is aimed at, and it is the hardest of the five to satisfy, because the bodies of it that are easiest to fetch are translations.

The translated pool announces itself. python/python-docs-pl, the Polish translation of the Python documentation, is 501 PO files holding 1,999,845 words of English source against 169,936 Polish words in 11,589 of 86,735 strings. Every Polish word in it sits in a msgstr beside the msgid it came from, so the storage format records the provenance that NKJP's headers do not. The Kubernetes documentation localizes into Polish the same way, contributor guide and terminology glossary included, so the arrangement is a convention rather than one project's choice. A corpus built from these carries a hazard worth naming: a rule against w oparciu o measured against prose produced by rendering based on into Polish reports the translator's habit as the norm, and the same holds for a coverage figure taken over translated documentation.

The original pool has to be gathered repository by repository. The largest single item found is CIRFMF/ksef-docs, the API documentation of the Polish national e-invoicing system, published by the Ministry of Finance under the MIT licence. It is written in Polish because the system, its law and its readers are Polish, which is the property that makes a repository worth adding and the one no licence field records. audit-corpus.md holds it, with the command that fetches it at the commit the figures here are taken at. Beyond it the pool thins. pot-gov-pl/rit-dokumentacja, in that list as well, is the only other item found, and the catalogue of open-source repositories of Polish state institutions covers 21 institutions across 25 GitHub accounts without describing a single one of their repositories as documentation.

Its typographic state is the reason this material is worth the gathering. Straight quotation marks outnumber „ by 314 to 13, en dashes outnumber em dashes by 131 to 11, and the straight marks are around Polish phrases in running prose — "Uwierzytelnianie zakończone niepowodzeniem z powodu błędnego tokenu", "Profil Zaufany", "Certyfikat zawieszony". This is prose caught at the stage a linter would run at, holding what NKJP and Wolne Lektury cannot hold.

The distribution of those hits is the caution that goes with the finding. 257 of the 314 are in api-changelog.md, which quotes the wording of API error messages release by release, so one document with one habit supplies four fifths of them. A pack audited against this repository alone would be audited against a handful of authors, which is the argument for the corpus being a list of repositories rather than a repository.

The expository Polish that is obtainable

Between the register that barely exists and the corpora that are the wrong register sits edited expository Polish, and this is what there is of it.

PLSC, the Polish Library of Science Corpus, is 159,767 records of title, abstract, journal and discipline from Polish scientific journals, 194 MB of them, under CC0. It is the only body in this survey whose licence asks nothing of whoever redistributes it. Two things it is not. An abstract is a paragraph written to a house convention, not running documentation. And a Polish journal publishing in English still carries a Polish abstract, which is then a translation, so the pool is Polish-first in part and not throughout.

Polish Wikipedia is 1,703,870 articles and 561,399,817 article words under CC BY-SA 4.0, with dumps and an API.

curl -s 'https://pl.wikipedia.org/w/api.php?action=query&meta=siteinfo&siprop=statistics&format=json'

It is expository and edited by more than one person, and the per-article provenance is not recorded any more than NKJP's is. An article's interlanguage links and its own edit history are where an answer for one article would come from, and neither has been read here for any of them.

Polish Wikibooks is 7,216 articles and 5,802,842 words under CC BY-SA 4.0, by the same query against pl.wikibooks.org. That is a hundredth of Wikipedia, in the instructive register documentation is nearest to.

Open Polish coursebooks, the open-coursebooks-pl dataset, is 1,528 chapters of Polish coursebook material, 4.8 MB, under CC BY-SA 4.0, written in Polish rather than translated into it.

SpeakLeash aggregates Polish text with a per-source manifest recording licence, category and size for each dataset it holds, which makes it the place to look for slices of the above rather than a corpus itself. Which of its datasets are in this register, and under what terms, is not checked here.

What the survey settles

The target register is nearly absent from the corpora that exist. Seven percent of NKJP's downloadable subcorpus is instructive or academic writing, one sample of it is documentation of a computer system, and Śmigiel's six domains do not include the register at all. Anyone wanting a distribution over Polish technical documentation assembles it; nobody has assembled it already.

Provenance is recorded in the data only where the data is translated. NKJP names no translator, and Wikipedia names no source article, while a PO file keeps the English beside every Polish string it holds. A rule against calques therefore needs material whose Polish-first origin somebody established when they gathered it, which makes provenance a property of the gathering rather than of the corpus, and makes it something this repository has to write down rather than something a corpus can be asked for.

Stage of production is a property of the corpus build, not of the typesetter. NKJP's text layer and Wolne Lektury's prose disagree by a factor of three and a half thousand on how often a Polish quotation opens with „ rather than with ", which is one statistic over one language, so at most one of them describes Polish practice and nothing in either says which. A typographic rule can be audited only over text that reached its reader through no step that rewrote its characters, and a repository cloned from version control is the case where that can be checked rather than assumed.

Negative material costs nothing to gather. Olski admits a declared subset and rejects everything outside it by construction, so a corpus of ungrammatical Polish would pass the run that first read it and go on passing, and whoever wrote it would be whoever is writing the grammar rather than somebody writing in the register. The rejections over the corpora above are the negative material, and over eight sentences in ten of Składnica are one (corpus.md), in somebody else's words and under a pinned release. What a rejection wants is therefore a reading of it rather than more of it.

The composition this argues for

Two corpora, because the two shapes of rule in linter.md wanted different prose, and one corpus cannot be both. Only the first of the two was ever assembled, and it is the one the grammar is measured over as well.

The audit corpus: Polish documentation in version control

Original Polish technical documentation, cloned from a list of repositories this repository keeps, prose extracted by the step the harness already has, which reads Markdown, and by whatever second extraction the format of a later member calls for. Every rule whose answer depended on the site rather than on the rate audited here: the whole typography pack, and any later rule with the same property. What reads it now is the grammar, sentence by sentence.

It is small on purpose, and it is a list rather than a body. An audit reads its hits, so tens of thousands of words is a working corpus, and one repository of that size supplies more hits than an afternoon settles. What one repository cannot supply is authors: four fifths of KSeF's straight quotation marks are in one of its files. So the corpus grows by adding repositories rather than by adding words, and a repository joins if its documentation was written in Polish first. The list is where that test is written down and applied, with the reason recorded beside each member and with the three further things a repository has to show.

The distribution corpus: edited original expository Polish

For every rule reporting a rate against a norm. Proportions by words, with the reason each is capped where it is:

share source why it is capped there
50% PLSC abstracts, CC0 the largest, and the only licence that asks nothing; abstracts rather than running prose, and Polish-first only in part
25% Polish Wikipedia, technical and scientific categories, CC BY-SA 4.0 supplies the running prose abstracts cannot; per-article provenance unread
15% Polish Wikibooks and open coursebooks, CC BY-SA 4.0 the instructive register, which is the nearest one to documentation; between them a few million words
10% NKJP typ_nd and typ_inf-por, 73,722 words, CC BY printed-book prose an editor worked on, in a balance somebody else designed; a median dated text from 2000, and an instructive half that is dream books and travel guides

The principle behind the numbers is that each source's share is bounded by the defect it is known to carry, so that no single defect can set a threshold on its own. The numbers themselves are a starting point rather than a result, and the run that corrects them is cheap: recompute every threshold with each source dropped in turn, and a threshold that moves when one source leaves is a threshold measuring that source.

Wolne Lektury is in neither corpus. It is verse, it is period-marked, a seventh of it is translated, and on the one statistic a shipped rule measures it sits at twice the rule's threshold for a reason that is Polish orthography.

The generated half

Śmigiel is the only generated Polish in this survey that is openly licensed and describes a pipeline with no editing pass in it, and it cannot be the only generated half. It meets the condition generated-polish.md sets, and it fails the register and the task, being fragments continuing human prefixes in six domains that exclude documentation. So it enters as the second opinion, the corpus that says whether a rate belongs to the register or to the model, and the primary generated half remains one produced in the register, from briefs.

One overlap has to be kept in view when it is used that way. PLSC and the open coursebooks are in the distribution corpus above and in Śmigiel's human half both, so the two human sides are not independent samples of anything, and a difference between them understates by however much they share.

That also answers what generated-polish.md asks about the corpus it measures. With an unedited generated half obtainable, a body edited against style detectors is not needed as the generated half, and stays what that document calls it: a floor, and the harder case.

Not yet decided

  • Whether a Wikipedia article's provenance can be established at scale. Interlanguage links and the article's own history are the obvious places to look, and neither has been checked.
  • Whether PLSC's abstracts are Polish-first. A Polish journal publishing in English carries a translated Polish abstract, and the dataset records the journal, which is where the answer starts.
  • Which repositories join the audit corpus beyond the two in the list, and how many authors it takes before a hit rate in it stops describing whoever wrote the largest file.
  • Whether SpeakLeash's manifests hold a slice of this register that would be cheaper than gathering one.
  • Whether the register of statutes belongs in this survey. It answers the five questions above where nothing here does: no copyright to clear, one register declared by the rules acts are drafted under, and a text that cannot change under the address it is fetched by. What it costs to read, and the one thing it cannot support — the typography, which its publisher normalizes away — is in ustawy.md.

Sources