# Polish Wikipedia user-talk pages (namespace 3) - CC BY-SA 4.0

This slice is rebuilt deterministically from the pinned official Wikimedia dump
`plwiki-20260901-pages-meta-current` (7 sha1-verified shards;
https://dumps.wikimedia.org/plwiki/20260901/). Scope is namespace 3
(user talk); mainspace (ns=0) is covered by the existing
`wikipedia` shard and article/project talk (ns=1,5) ships as the separate
plwiki_talk slice. This is conversational written Polish - editor-to-editor
messages, welcome notices, warnings, disputes, thanks and informal
coordination - a register otherwise nearly absent from the corpus.

- Non-redirect user-talk pages enumerated: 450,867
- Source slice retained after QA/dedup: 380,527
- Proposed records: 380,527
- Proposed tokens: 381,445,111 (`cl100k_base` proxy)
- Characters: 1,020,728,459
- License: CC BY-SA 4.0 per record (Wikimedia Terms of Use section 7: https://foundation.wikimedia.org/w/index.php?title=Policy:Terms_of_Use/pl&oldid=584706)
- Immutable source snapshot: https://huggingface.co/datasets/PiotrSty/plwiki-user-talk-pages/tree/8d12dbc8134dcfe777937e75e562e1b58033f20a

Per-record page id, title, namespace, last-revision timestamp and canonical URL
are preserved; page history is the contributor-attribution trail. Wikitext is
stripped with mwparserfromhell; lines occurring in >=1% of documents are removed
as cross-document boilerplate; emails, labelled phones, IP literals, labelled
national identifiers and 26-digit account candidates are redacted (pattern
redaction is not complete anonymization); exact and near deduplication applied
within source. The article-talk (ns=1) and project-talk (ns=5)
namespaces are excluded by scope decision; they ship as the separate
plwiki_talk slice.

## What the texts are (user-generated)

These are **user-generated** texts: messages written by volunteer Wikipedia
editors on each other's user-talk pages. Genre: editor-to-editor conversation —
help threads, article-source requests, edit discussions and disputes, thanks,
adminship debates, and templated bot notices (welcome, warning, account-renaming
announcements). Register: informal written Polish — typos, abbreviations,
occasional vulgarity and vandalism residue; signed comments (`user name +
timestamp`) are the norm and are kept. Not OCR'd, not machine-translated.

## Time span of the content

The content is dated by its own signatures, not only by the dump: 379,443 of
380,527 records contain wiki-signature timestamps spanning **2002-12 to 2026-09**
(the bulk falls in 2005-2026; the ~50 pre-2002 date strings found in the scan
are quoted historical dates - plwiki launched 2001-09). Per-record `created` is
the dump last-revision date (2002-12-11 - 2026-09-01). Measured by
signature/date extraction over all texts (per-year counts in the review thread).

## Scrubber and residual-PII spot check

Scrubber that ran: **pattern redaction `redact_pii()` in
`src/build_plwiki_usertalk_pl.py`** - e-mail, labelled phone (`telefon`/`tel.`/
`phone`), IPv4, parsed IPv6 (incl. compressed notation), labelled
PESEL/NIP/REGON, 26-digit account candidates -> `[REDACTED:*]`. This is **not**
[SlayerLab/NERGAL](https://huggingface.co/SlayerLab/NERGAL); the shard must be
re-run through NERGAL before it is admitted to a release.

Residual-PII spot check (deterministic sample of 60 records, seed 20260928):
0 hits on residual classes not covered above (bare 9-15-digit runs, 26-digit
runs, IBAN-like strings, unlabelled IPs, e-mail-like strings incl. obfuscated
"at"/"małpa" forms, `ul.`/`al.`/`os.` street cues, GG/ICQ/Skype handles, PESEL
mentions); 3 `[REDACTED:*]` placeholders remain as evidence of the applied
redaction. Manual reading of 8 sampled records found only public wiki
attribution (editor user names and timestamps in signatures - the CC BY-SA
credit trail) and no private personal data. Residual risk is low but non-zero
(60 of 380,527 records, 0.016%; free text can carry data no pattern catches),
which is why the NERGAL pass stays a release gate.

## Text overlap with the wikipedia / plwiki_talk shards (review audit)

Measured post-build by normalized exact-line and exact-paragraph containment
(64-bit blake2b; lines >= 30 chars, paragraphs >= 100 chars; exact-verbatim only):

- vs `wikipedia` (ns 0): 469 of 4,218,161 user-talk lines match (61,106 chars =
  **0.006%** of user-talk chars); 0 paragraphs. Verbatim article quoting in talk
  pages is negligible at this unit size.
- vs `plwiki_talk` (ns 1,5): headline line-char containment is 28.7% (286,157,711
  chars), but it concentrates in **47 distinct long notice/template lines** that
  occur on 500+ user-talk pages each - 284,115,864 occurrence chars = **99.3% of
  the matched chars**. These lines are longer than the 160-char line-boilerplate
  cap, so the build-time rule did not drop them. All other matched lines (4,725
  distinct) account for 2,041,847 occurrence chars = **0.2%** of user-talk chars;
  thread copies and genuine quotes (lines on at most 4 pages) are 1,253,089 chars =
  **0.13%**.

Integration gate: at target integration, drop or dedup the shared long notice
lines against `plwiki_talk`; the residual cross-namespace thread duplication
(~0.1-0.2%) can ride along with the ordinary cross-source dedup pass. Method and
numbers are in the review thread (`overlap_check.py`, `overlap_frequency.py`).

## Audit trail

The per-record trail ships in this repo: `plwiki_user_talk.attribution.jsonl`
(380,527 lines) and `plwiki_user_talk.decisions.jsonl` (450,867 lines) are LFS
files under `.gitattributes` (`data/*/*.attribution.jsonl`,
`data/*/*.decisions.jsonl`), so the PR diff shows pointer text while the Hub
resolves the full content at the file view. Both are reproduced byte-for-byte by
`src/build_plwiki_usertalk_pl.py` from the pinned dump; `plwiki_user_talk.checksums.json`
pins their digests.

Limitations: user-talk pages are informal (typos, bot notices, templated warnings,
edit disputes, occasional vandalism);
comments may quote article text (measured at review: 0.006% of chars match the
`wikipedia` shard verbatim at line level; the larger overlap is shared long
notice/template text with `plwiki_talk` - see the overlap audit section);
Record model is a whole talk page; comments are not split.