Datasets:

SlayerLab
/

polish-dynaword

Adds plwiki_talk: 43,054 records and 55,568,733 measured cl100k_base proxy tokens (144,604,697 characters) from the pinned official Wikimedia dump plwiki-20260901-pages-meta-current. These are user-generated talk pages: editorial discussion about article content (ns=1, 41,828 pages) and about the project itself (ns=5, 1,226 pages).

Source dataset: https://huggingface.co/datasets/PiotrSty/plwiki-talk-pages/tree/20c74a6da0cbaef5d2a644d0accf7e90262ea24c

What the texts are

Genre: editorial discussion - source requests, wording and naming disputes, style fixes, merge/delete debates, project-wide consultations. Register: informal written Polish with typos, wiki jargon ("WP:OR", "SDU"), signed comments (user name + timestamp). Time span: per-record created = dump last-revision date, 2003-02-13 – 2026-09-01. User-generated, not OCR'd, not machine-translated. Record model: one whole talk page.

Sample (verbatim, from the built shard)

Article talk - bibliographic source request and replies (ns=1):

id: plwiki_talk_1903252
source: plwiki_talk
added: 2026-09-11
created: 2023-12-04
token_count: 315
license: CC-BY-SA-4.0
author: plwiki contributors

Strona przekierowująca 101 Rezerwowy Pułk Piechoty dla kategorii Pułki piechoty II RP i Ostrów Mazowiecka. --grzes1966 (dyskusja) 11:31, 14 lut 2010 (CET)

Witam.
W bibliografii na temat 84pSP nie widzę opracowań źródłowych dr.hab. Andrzeja Wesołowskiego.
"Szczerców Góry Borowskie"
"Mokra - Działoszyn"
"My Strzelcy Polescy"
Czy kolega nie miał czy też nie ma dostępu do tych publikacji?
ps.
w tym pułku walczył mój dziadek. http://www.bohaterowie1939.pl/polegly,wira,franciszek,16023.html
Pozdrawiam Robert Kielek Wrześniowiec (dyskusja) 15:42, 11 lis 2022 (CET)
Zgodnie z Twoim życzeniem uzupełniono historię walk wrześniowych pułku w oparciu również o przytoczoną przez Ciebie bibliografię. Pozdrawiam, Grzegorzg1960 (dyskusja) 21:46, 26 mar 2023 (CEST)

Article talk - naming/translation dispute (ns=1):

id: plwiki_talk_6047211
source: plwiki_talk
added: 2026-09-11
created: 2026-03-11
token_count: 212
license: CC-BY-SA-4.0
author: plwiki contributors

Nazwa artykułu

Obecnie wszystkie źródła w artykule są anglojęzyczne, dlatego ten przekład ang. double empathy problem to prawie twórczość własna (WP:OR). Słyszałem krytykę przekładu „problem podwójnej empatii” – to problem jest podwójny, a nie empatia. Hasło „podwójna empatii” można znaleźć w Google, ale chyba w niczym naukowym. Z drugiej strony „podwójny problem empatii” też znajduję tylko w mało oficjalnych źródłach typu strony szkoleń. Ktoś może ma dostęp do polskich źródeł psychologicznych? Tarnoob (dyskusja) 23:15, 11 mar 2026 (CET)

Project talk - classification proposal debated project-wide (ns=5):

id: plwiki_talk_4488
source: plwiki_talk
added: 2026-09-11
created: 2008-07-20
token_count: 674
license: CC-BY-SA-4.0
author: plwiki contributors

Podoba mi się część propozycji nr 2 (Cywilizacja, Dzieje świata) i - co dziwne - Inne z propozycji nr 3. RomanK.

3 strasznie szatkuje biologie i informatyke, za to 1 i 2 wrzucaja je w jakies zakamary.
Moim zdaniem trzeba eksponowac takie tematy - na temat informatyki niewatpliwie bedzie tu proporcjonalnie o wiele wiecej niz na przecietnej papierpedii, a biologia jest bardzo rozlegla, wiec rowniez bedzie tu duzo artykułów.

A po za tym informatyka nie za bardzo jest nauką ścisłą. --Taw

Proponuję połączenie propozycji 2 i 3
Akir

Akir,
Tworzenie tego typu "twardych" klasyfikacji jest trudne żmudne podatne na błędy i utrudniające tworzenie artykułów.
To już było "przerabiane" na angielskiej wikipedii i zdecydowano że nie powinno być tego typu klasyfikacji.
BTW rozumiem Cię i mnie też czasem korci żeby klasyfikować, ale jak się dobrze zastanowić to nie jest to takie proste i dużą część pracy nasi
Wikipedianie musieliby spędzać nad rozmyślaniem o tym jak to zaklasyfikować, potem nad wpisywaniem klasyfikacji, a część i tak robiłaby to źle i inni znów musieli by to poprawiać itd.
Pozdrowienia--Kpjas
[…]

PII - which scrubber ran

Pattern redaction in src/build_plwiki_talk_pl.py (e-mail, labelled phone, IPv4, parsed IPv6 incl. compressed notation, labelled PESEL/NIP/REGON, 26-digit account candidates -> [REDACTED:*]). This is not SlayerLab/NERGAL. Residual-PII spot check (60 records, seed 20260928): 25 [REDACTED:IP] placeholders remain (the applied redaction, incl. unregistered editors' signatures); the one long-digit-run hit is an academic DOI citation (10.1017/9781108668675.006) - not PII; 0 hits on all other residual classes; manual reading of 8 records found only public wiki attribution and editorial discussion. These are user-generated talk pages - no release before a NERGAL pass; merge with release: None is fine meanwhile. Full report in the review thread and in the datasheet.

Inclusion argument and overlap (item 3)

What this adds that wikipedia (ns 0) does not: the editorial process around the articles - multi-voice disagreement about sources, wording, naming, merges, deletions. The namespace partition is exact: wikipedia = ns 0, this slice = ns 1+5, plwiki_user_talk (#40) = ns 3 - no page is in two slices.

Text-level overlap against both (normalized exact-line/paragraph containment, ≥30/≥100 chars, 64-bit blake2b, verbatim only): against wikipedia (ns 0) 418 of 567,171 lines match — 75,163 chars = 0.053% of this slice's line chars (1 paragraph, 271 chars); the feared verbatim article quoting is negligible at this unit size. Against plwiki_user_talk (#40): 5,556 lines (1,207,378 chars = 0.85%), 1,111 paragraphs (547,695 chars = 0.39%); 1,965 of 43,054 records (4.6%) contain ≥1 matched line. The matched lines are almost all page-unique (4,941 of 5,138 distinct matched lines on exactly one talk page) — genuine thread fragments mirrored between article talk and user talk, not template mass-duplication (the reverse of the ns-3-side profile in #40, where 47 shared notice lines dominate). Net: this slice is ~99% text-novel vs both references at verbatim unit level; cross-shard verbatim dedup at integration would move ~0.9% of chars. (Exact-verbatim only; quotes below the unit thresholds are not counted.)

Rights and method

Same CC BY-SA 4.0 basis as the shipped wikipedia shard (Wikimedia Terms of Use section 7: https://foundation.wikimedia.org/w/index.php?title=Policy:Terms_of_Use/pl&oldid=584706), on a namespace complement. Dump shards sha1-verified against dumpstatus.json. Wikitext stripped via mwparserfromhell; >=1%-frequency boilerplate lines removed; email/labelled-phone/IPv4/parsed-IPv6/labelled-PESEL-NIP-REGON/26-digit-account-candidate patterns redacted (not complete anonymization); langid vote; exact + near dedup within source (deterministic prefix-indexed capped-shingle Jaccard >= 0.90).

Rejection breakdown

Reason	Count
too_little_text (<500 chars)	500,652
non_polish_text	677
too_little_text_after_redaction	39
low_letter_ratio	92
normalized_duplicate	36
near_duplicate	41

Overlap and Slayer ontology

The overlap audit compared source-prefixed record IDs only (0 collisions); the text-level numbers above replace that weak evidence. Benchmark checks, abuse screening and controlled ablations remain pending. This PR proposes a source, not a merged or stable DynaWord release.

Limitations

Talk pages are informal: typos, edit disputes, occasional vandalism. Record model is a whole talk page; comments are not split.

Source snapshot

Repo: https://huggingface.co/datasets/PiotrSty/plwiki-talk-pages
Commit: 20c74a6da0cbaef5d2a644d0accf7e90262ea24c, tag v1.0.0
35 files verified remotely against local sha256

Test plan

python3 -m pytest src/test_build_plwiki_talk_pl.py — 16 passed (rebased onto main @ 34e7ea3)
Residual-PII spot check (above)
Text-level overlap vs wikipedia and plwiki_user_talk (above)
Reviewer: inclusion argument accepted
Reviewer: NERGAL-before-release note accepted

Add plwiki talk pages (ns 1+5) - conversational Polishd3c7b399

PiotrSty

Fabryka AI org 28 days ago

@kacperwikiel Review requested.

Polish Wikipedia talk pages (ns 1+5) — conversational register

43,054 records, 55.6M cl100k_base tokens, 144.6M characters
Source: pinned plwiki-20260901-pages-meta-current dump, 7 sha1-verified shards
Scope: namespaces 1 (article talk) + 5 (project talk); ns=0 covered by existing wikipedia shard, ns=3 excluded by scope decision (not measured)
License: CC BY-SA 4.0 per Wikimedia Terms of Use section 7; per-record page URL preserves contributor-history attribution

Processing

Wikitext stripped via mwparserfromhell; >=1%-frequency cross-document boilerplate lines removed; Unicode/whitespace normalization; PII pattern redaction (email, labelled phone, IPv4, parsed IPv6 including compressed notation, labelled PESEL/NIP/REGON, 26-digit account candidates — not complete anonymization); three-window langid vote; exact + near dedup within source (deterministic prefix-indexed capped-shingle Jaccard >= 0.90).

Rejection breakdown

Reason	Count
too_little_text (<500 chars)	500,652
non_polish_text	677
too_little_text_after_redaction	39
low_letter_ratio	92
normalized_duplicate	36
near_duplicate	41

Before/after evidence

Three complete before/after samples are in data/plwiki_talk/plwiki_talk.sample.jsonl and the source repo's artifacts/sample.jsonl. They show prose, paragraph boundaries, signatures and disagreement preserved; whitespace normalized; IP patterns masked.

Overlap

The overlap audit compared source-prefixed record IDs only; zero collisions do not establish text novelty. Comments can quote article text, so quoted-fragment text dedup remains a target integration gate.

Limitations

Talk pages are informal: typos, edit disputes, occasional vandalism. Record model is a whole talk page; comments are not split. Benchmark contamination, abuse screening and controlled training ablation remain pending.

Source snapshot

Repo: https://huggingface.co/datasets/PiotrSty/plwiki-talk-pages
Commit: 20c74a6da0cbaef5d2a644d0accf7e90262ea24c
Tag: v1.0.0
35 files verified remotely against local sha256

ppuzio

Fabryka AI org 11 days ago

Review against AGENTS.md ("What the PR must show", "Normalize the text before you build"). Needed before merge:

Sample inline. Paste 2–3 real documents into the description, one fenced block each: metadata as key: value lines, a blank line, then the text with its real line breaks, truncated with […] (format as in #24). A linked .sample.jsonl renders on the Hub as one escaped line per document and can't be reviewed at a glance.
PII. AGENTS.md requires SlayerLab/NERGAL. The source can merge with release: None after another scrubber, but it won't be admitted to a release until the shard is re-run through NERGAL. State in the datasheet which scrubber ran. Talk pages are user-generated, so this matters more than for official sources.
Inclusion argument. State what this adds that wikipedia doesn't, and how it relates to plwiki_user_talk (#40); text-level overlap check against both.

Address review: samples, scrubber + PII spot check, inclusion argument, text overlapc95ab889

PiotrSty

Fabryka AI org 11 days ago

Review addressed — rebased onto main @ 34e7ea3 (conflicts in .gitattributes and src/sources.py resolved: the two per-file LFS lines are gone in favour of the generic data/*/* rules, and the registry entry is re-applied with release: None). Point by point:

Sample inline. 3 real documents in the description, one fenced block each (key: value metadata, blank line, real line breaks, […] truncation, format as in #24): an article-talk source-request thread with signatures (ns=1), an article naming/translation dispute (ns=1), and a project-wide classification debate (ns=5, hand-picked so both namespaces are represented).
PII. Datasheet now names the scrubber: the pattern pass in src/build_plwiki_talk_pl.py (e-mail, labelled phone, IPv4, parsed IPv6 incl. compressed notation, labelled PESEL/NIP/REGON, 26-digit account candidates -> [REDACTED:*]) — not NERGAL. Residual-PII spot check added (60 records, seed 20260928): 25 [REDACTED:IP] placeholders remain as evidence of the applied redaction (incl. unregistered editors' signatures); the single long-digit-run hit is an academic DOI (10.1017/9781108668675.006), not PII; 0 hits on all other residual classes; 8 records read by hand — only public wiki attribution and editorial discussion. Agreed it matters more for user-generated text: no release before a NERGAL pass; merge with release: None is fine meanwhile.
Inclusion argument + text-level overlap. Argument is now in the datasheet and description: this slice is the editorial process around the articles (source requests, wording/naming disputes, merges/deletions, project consultations) - multi-voice disagreement that the finished prose in wikipedia does not contain. Namespace partition is exact (wikipedia = ns 0, this slice = ns 1+5, plwiki_user_talk (#40) = ns 3). Text-level overlap, measured the same way as in #40 (normalized exact-line/paragraph containment, ≥30/≥100 chars, 64-bit blake2b, verbatim only): vs wikipedia 0.053% of this slice's chars (418 of 567,171 lines; 1 paragraph) — the verbatim article-quoting fear is negligible at these unit sizes. Vs plwiki_user_talk 0.85% of chars (5,556 lines) / 0.39% of paragraph chars, with 4.6% of records touched — and the matched lines are almost all page-unique (4,941 of 5,138 on exactly one talk page): thread fragments mirrored between article talk and user talk, not template duplication (the reverse of #40's ns-3-side profile, where 47 shared notice lines dominated). Net: ~99% text-novel vs both; cross-shard verbatim dedup at integration moves ~0.9% of chars. Method + numbers: review_fix/overlap_check_pr36.py, overlap_pr36.json.

Verification: python3 -m pytest src/test_build_plwiki_talk_pl.py — 16 passed on the rebased branch.

ppuzio changed pull request status to merged 10 days ago

ppuzio deleted the refs/pr/36 ref 10 days ago

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment