Datasets:

---

SlayerLab

/

polish-dynaword

Like 24

Follow

Fabryka AI 84

Add Polish Wikipedia user-talk pages (ns 3) - user-generated conversational register

Adds plwiki_user_talk: 380,527 records and 381,445,111 measured cl100k_base proxy tokens from the pinned official Wikimedia dump plwiki-20260901-pages-meta-current. This is user-generated text: messages volunteers leave on each other's user-talk pages.

Source dataset: https://huggingface.co/datasets/PiotrSty/plwiki-user-talk-pages/tree/8d12dbc8134dcfe777937e75e562e1b58033f20a

What the texts are

Genre: editor-to-editor conversation — help threads, article-source requests, edit discussions and disputes, thanks, adminship debates, and templated bot notices (welcomes, warnings, account-renaming announcements). Register: informal written Polish, typos and abbreviations included, comments signed user name + timestamp (kept). User-generated, not OCR'd, not machine-translated. Record model: one whole talk page; comments are not split.

Time span of the content (measured, not just the dump date): 2002-12 – 2026-09 by wiki-signature timestamps, present in 379,443 of 380,527 records (bulk 2005–2026; the ~50 pre-2002 date strings found are quoted historical dates — plwiki launched 2001-09). Per-record created is the dump last-revision date (2002-12-11 – 2026-09-01).

Sample (verbatim, from the built shard)

Real editor-to-editor conversation (two short threads on one page):

id: plwiki_user_talk_2850435
source: plwiki_user_talk
added: 2026-09-27
created: 2020-02-14
token_count: 214
license: CC-BY-SA-4.0
author: plwiki contributors

Prośba o źródła w artykule Związek Chłopski (II RP)

--Pablo000 dysk 22:21, 9 sty 2013 (CET)

Statystyki Czarnogóry

Witam. Fajnie, że edytujesz statystyki. Jest to dość trudny i zaniedbany nieco temat na Wikipedii. Nie mogę zrozumieć, skąd wziąłeś dane do tej edycji. Ja w wynikach spisu widzę co innego (np. liczba ludności 620 029) i na takie wartości poprawiam. Jak coś źle zrozumiałem, to proszę daj mi znać, o co chodzi w tych rozbieżnościach. Pozdrawiam --Wiklol (Re:) 10:03, 1 cze 2013 (CEST)

Templated bot notice — the most common document type on ns 3 pages:

id: plwiki_user_talk_4715143
source: plwiki_user_talk
added: 2026-09-27
created: 2024-03-26
token_count: 583
license: CC-BY-SA-4.0
author: plwiki contributors

Wikipedię można edytować na dwa sposoby. Edytor wizualny jest proststzym, bardziej intuicyjnym narzędziem. Z kolei edytor kodu źródłowego daje dostęp do bardziej zaawansowanych funkcji. Można łatwo przełączać się między nimi. Najłatwiej jest nauczyć się edytowania w praktyce, dlatego też po wykonaniu wybranym sposobem kilku edycji testowych w swoim brudnopisie, śmiało edytuj strony!

W naszych szeregach mamy przewodników, którzy pomogą Ci w Twoich pierwszych edycjach. Wikipedyści odpowiedzą na pytania dotyczące edytowania Wikipedii. Możesz zadać je tutaj. Pomocnych wikipedystów znajdziesz też na naszym czacie.
Jeżeli chcesz samodzielnie szukać stron pomocy, zacznij od tej strony (znajdziesz ją zawsze w menu po lewej: „” ⇒ „”). Możesz też ustawić własne preferencje m.in. dotyczące wyglądu […]

Long-lived community page (multi-year thread with informal voices and old-style signatures):

id: plwiki_user_talk_3130457
source: plwiki_user_talk
added: 2026-09-27
created: 2024-06-30
token_count: 13401
license: CC-BY-SA-4.0
author: plwiki contributors

Zostaw mi wiadomość. Jeśli z jakichś względów jest to niemożliwe – napisz maila.
180px|Hello!Wydania archiwalne: 1 • 2 • 3 • 4 • 5 • 6 • 7 • 8 • 9 • 10 • 11 • 12 • 13 • 14 • 15 • 16 • 17 • 18 • 19 • 20 • 21 __TOC__

Ad:Wikipedia:Tablica_ogłoszeń/Ogłoszenia

[#bodyContent AdWikipedia:Tablica_ogłoszeń/Ogłoszenia]

Zła to wiadomość! Byłeś jednym z niewielu adminów, do których miałem zaufanie... Pozdrawiam, Przemek Jahr @ 11:29, 5 mar 2014 (CET)
Zaraz tu pewnie będzie lista solidarnych typu „wróć!” albo „o nie!”. Więc trochę Tobie, trochę im napiszę: szkoda, że kwestie poza-wiki w Twoim przypadku mają jakiś związek z uprawnieniami na wiki, ale coś postanowił, to postanowiłeś, więc powodzenia. I tak wiem, że wrócisz – administrator Wikipedii to stan umysłu ;) Tar Lócesilion|queta! 11:37, 5 mar […]

Before/after of the wikitext stripping and boilerplate removal

Full set of excerpts in the review thread below; two shown here. BEFORE is raw wikitext from the pinned dump, AFTER is the shipped record (strip_wikitext → strip_boilerplate → normalize → redact_pii, all in src/build_plwiki_usertalk_pl.py).

Template + styled-signature stripping (plwiki_user_talk_181941):

BEFORE

{{witaj}}
[[Wikipedysta:AndrzejzHelu|<span style="color:darkblu;">Andrzej z Helu</span>]] [[Dyskusja wikipedysty:AndrzejzHelu|<sup><b style="color:#c00;">LIST</b></sup>]] 20:15, 25 cze 2005 (CEST)

== Podpis ==

Witaj, 

W ramach sprzątania przestrzeni Szablon: wszystkie podpisy Wikipedystów są przenoszone do ich przestrzeni, czyli przestrzeni Wikipedysta:. Twój podpis, tj. {{s|Martin}} {{#if: {{{3|}}} |i {{s|{{{3}}}}}|}} {{#if: {{{4|}}} |i {{s|{{{4}}}}}|}} został przeniesiony do [[Wikipedysta:Martin/podpis]]. […]

AFTER

Andrzej z Helu LIST 20:15, 25 cze 2005 (CEST)

Podpis

Witaj,

W ramach sprzątania przestrzeni Szablon: wszystkie podpisy Wikipedystów są przenoszone do ich przestrzeni, czyli przestrzeni Wikipedysta:. Twój podpis, tj. został przeniesiony do Wikipedysta:Martin/podpis. […]

Boilerplate-line removal (plwiki_user_talk_3562; dropped from this record: "Szanowny użytkowniku!", "Nazwa Twojego konta zostanie zmieniona", "Przepraszamy za wszelkie niedogodności.", "Renamed"):

BEFORE

== Linki zewnętrzne ==
Nie jestem pewien czy moje zdanie jest słuszne, ale wydaje mi się
powinniśmy umieszczać linków do polskich serwerów.<br>
Powód zawartość powinna być na Wikipedii.<br>
Odciąga to ludzi od pracy nad Wikipedią.<br>
Jeśliby pociągnąć to do extremum to Wikipedia stałaby się
w znacznej większości zbiorem linków do innych stron z drobnym
komentarzem.<br>--[[Wikipedysta:Kpjas|Kpjas]]

AFTER

Nie jestem pewien czy moje zdanie jest słuszne, ale wydaje mi się powinniśmy umieszczać linków do polskich serwerów.
Powód zawartość powinna być na Wikipedii.
Odciąga to ludzi od pracy nad Wikipedią.
Jeśliby pociągnąć to do extremum to Wikipedia stałaby się w znacznej większości zbiorem linków do innych stron z drobnym komentarzem.--Kpjas

83 frequent line forms are removed line-wise (>=1% document frequency, <160 chars, pinned in artifacts/boilerplate_lines.json); 818 high-df 5-word shingles are excluded from the near-dedup sketches. Literal markup inside <nowiki> tutorial quotes is preserved as content on purpose.

Scrubber and residual-PII spot check

Scrubber run: pattern redaction redact_pii() in the builder — e-mail, labelled phone (telefon/tel./phone), IPv4, parsed IPv6 (incl. compressed), labelled PESEL/NIP/REGON, 26-digit account candidates → [REDACTED:*]. This is not SlayerLab/NERGAL.

Residual-PII spot check (60 records, seed 20260928): 0 hits on residual classes (bare 9–15-digit runs, 26-digit runs, IBAN-like, unlabelled IPs, e-mail-like incl. obfuscated forms, ul./al./os. street cues, GG/ICQ/Skype, PESEL mentions); 3 [REDACTED:*] placeholders remain. Manual reading of 8 records: only public wiki attribution (user names + timestamps in signatures — the CC BY-SA credit trail), no private personal data. Residual risk low but non-zero (0.016% sample) — this is user-generated conversation, exactly the case the policy exists for: the shard must be re-run through NERGAL before release. It can merge with release: None meanwhile.

Text overlap with wikipedia / plwiki_talk (measured, not just IDs)

Normalized exact-line / exact-paragraph containment (64-bit blake2b; lines ≥30 chars, paragraphs ≥100 chars; exact-verbatim only):

• vs wikipedia (ns 0): 469 of 4,218,161 lines match (61,106 chars = 0.006% of user-talk chars); 0 paragraphs. The feared verbatim article quoting is negligible at this unit size.
• vs plwiki_talk (ns 1,5): 909,953 line matches (286,217,144 chars = 28.7% of user-talk chars); 363,049 paragraph matches (179,191,013 chars = 18.8%). 262,674 of 380,527 records (69.0%) contain ≥1 matched line.

Interpretation and integration gate: the headline plwiki_talk containment concentrates in 47 distinct long notice/template lines occurring on 500+ user-talk pages each — 284,115,864 of 286,157,711 matched occurrence chars (99.3%); they are longer than the 160-char line-boilerplate cap, so the build-time rule kept them. All other matched lines (4,725 distinct) are 2,041,847 occurrence chars (0.2% of user-talk chars), and real thread copies/quotes (lines on ≤4 pages) are 1,253,089 chars (0.13%). Proposal: at target integration, drop or dedup the shared long notice lines against plwiki_talk (or explicitly accept the duplicated notice text); wikipedia needs no special treatment beyond the ordinary cross-source pass. Numbers and method: review_fix/overlap_check.py, overlap_frequency.py + reports in the review thread.

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: ns=3 (user talk); ns=0 is covered by the existing wikipedia shard and ns=1,5 ship as the separate plwiki_talk slice. Dump shards sha1-verified against dumpstatus.json. Audit trail ships in-repo: attribution.jsonl (380,527 lines) and decisions.jsonl (450,867 lines) are LFS files (the PR diff shows pointers; the Hub file view resolves the full content).

Overlap and Slayer ontology

Text-level overlap against wikipedia and plwiki_talk is now measured (above); the remaining gates are cross-shard dedup at integration and benchmark checks. The manifest carries content-addressed Versions, Protocol, Run, Evidence, falsifiable Claims, Actors and typed lineage. Abuse screening and controlled ablations remain pending. This PR proposes a source, not a merged or stable DynaWord release.

Test plan

• python3 -m pytest src/test_build_plwiki_usertalk_pl.py — 17 passed (with src/plwiki_usertalk_requirements.txt installed)
• Manual license review — Wikimedia ToU §7 linked and verified at the origin
• Manual PII review — residual spot check above
• Manual quality review — samples read; caveats: bot-notice share is large, informal register with typos and occasional vulgarity, whole-page records
• Ready for merge with release: None; NERGAL pass + cross-shard dedup before release

Proposed source: plwiki_user_talk — 380,527 retained records, 381,445,111 cl100k tokens (1.02B chars).

• Source: Polish Wikipedia user talk pages (namespace 3, Dyskusja wikipedysty) from the pinned official dump plwiki-20260901 (7 pages-meta-current shards, all SHA-1 verified)
• Repo: PiotrSty/plwiki-user-talk-pages @ v1.0.0 (commit 8d12dbc8134dcfe777937e75e562e1b58033f20a)
• License: CC BY-SA 4.0 — attribution preserved per record via page URL + history (artifacts/attribution.jsonl)
• Scope complement: wikipedia shard = ns0, plwiki_talk = ns1+5, this source = ns3 only; 0 source-prefixed ID collisions vs current target (1,171,897 rows)
• QA: wikitext strip, three-window langid vote, exact + near dedup (Jaccard>=0.90, capped-shingle prefix index, 818 boilerplate shingles >2% df excluded — bot greeting/warning templates), PII pattern redaction (email/phone/IP/national-ID/account candidates; residual counts = 0)
• Rejections: 56,140 too short, 10,296 near-dup, 2,734 exact dup, 1,084 non-Polish, 84 low letter ratio
• Caveats noted in artifacts: whole-page records (comments not split), created = dump last-revision date, text-level overlap vs target not proven by ID disjointness, manual license/PII/quality review pending — not claimed as complete anonymization

@kacperwikiel

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

1. 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.
2. 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. This is user-generated conversation, so it's the case the policy exists for; include a residual-PII spot check, not only a pattern count of 0.
3. Before/after excerpts of the wikitext stripping and boilerplate removal.
4. Time span of the talk-page content (not only the dump date), and state explicitly that it's user-generated.
5. Overlap. The audit compares page IDs only; talk pages quote article text verbatim. Check text-level overlap against wikipedia and plwiki_talk (#36).
6. Audit trail. attribution.jsonl / decisions.jsonl are 3 lines each for 380k records — if the per-record trail lives only in your personal repo, it isn't reproducible from here.
7. The description lists manual license/PII/quality review as pending — please mark the PR ready once that's done.

Review addressed — pushed as daea1bc (rebase) + b9e938c (review fixes) on refs/pr/40, rebased onto main @ 34e7ea3 (the .gitattributes conflict is resolved to main's version: the two per-file LFS lines are gone, the generic data/*/* rules cover this source). Point by point:

1. Sample inline. 3 real documents are now pasted into the description (one fenced block each: key: value metadata, blank line, real line breaks, […] truncation, format as in #24) — one real editor conversation, one templated bot notice (the most common document type), one long-lived community thread. render_sample.py output, verbatim.
2. PII. Datasheet now states the scrubber that ran: pattern redaction redact_pii() in the builder (not NERGAL) — and the NERGAL gate. Residual-PII spot check added (60 records, seed 20260928): 0 hits on residual classes (bare digit runs, 26-digit/IBAN-like, unlabelled IPs, e-mail-like incl. obfuscated "at"/"małpa" forms, ul./al. street cues, GG/ICQ/Skype, PESEL mentions), 3 [REDACTED:*] placeholders remain; manual reading of 8 records found only public wiki attribution (user names + timestamps in signatures) and no private personal data. Residual risk low but non-zero (0.016% sample) — agreed, no release before a NERGAL pass; merge with release: None is fine meanwhile.
3. Before/after. Full excerpts in the description (template + styled-signature stripping; boilerplate-line removal with the exact dropped lines) and 4 pairs in review_fix/before_after.md — BEFORE is raw wikitext read straight from the pinned dump XML (the extracted/ intermediate is post-strip_wikitext), AFTER is the shipped record. Literal markup inside <nowiki> tutorial quotes is preserved as content on purpose.
4. Time span + user-generated. Measured from the content itself: wiki-signature timestamps in 379,443 of 380,527 records span 2002-12 – 2026-09 (bulk 2005–2026; ~50 pre-2002 date strings are quoted historical dates). Per-record created = dump last-revision date (2002-12-11 – 2026-09-01). The datasheet and description now say explicitly: user-generated editor-to-editor conversation.
5. Overlap — measured at text level now. Normalized exact-line/paragraph containment (≥30/≥100 chars, 64-bit blake2b, verbatim only): vs wikipedia (ns 0) 0.006% of user-talk chars (469 lines; 0 paragraphs) — the verbatim article-quoting fear is negligible at this unit size. vs plwiki_talk (ns 1,5) the headline is 28.7% of line chars / 18.8% of paragraph chars, but a frequency split shows 99.3% of the matched chars are 47 distinct long notice/template lines present on 500+ user-talk pages each (longer than the 160-char line-boilerplate cap); real thread copies and quotes (lines on ≤4 pages) are just 0.13% of user-talk chars. Integration gate proposed in the datasheet: drop/dedup those shared notice lines against plwiki_talk at target integration. Method + numbers: review_fix/overlap_check.py, overlap_frequency.py, overlap_report.{md,json}, overlap_frequency.json.
6. Audit trail. It is not pointer stubs: attribution.jsonl (380,527 lines, 500 MB) and decisions.jsonl (450,867 lines, 137 MB) are LFS files — the PR diff shows pointer text, the Hub file view resolves the full content (verified by downloading both from refs/pr/40). src/build_plwiki_usertalk_pl.py reproduces them byte-for-byte from the pinned dump; plwiki_user_talk.checksums.json pins the digests. The datasheet now explains this.
7. Manual review done — ready. License verified at the origin (Wikimedia ToU §7, linked), PII spot check clean (above), quality review done (samples read; caveats: large bot-notice share, informal register with typos/vulgarity, whole-page records). Marking ready for merge with release: None.

Verification: python3 -m pytest src/test_build_plwiki_usertalk_pl.py — 17 passed (after pip install -r src/plwiki_usertalk_requirements.txt).