SlayerLab/polish-dynaword-mix-extended-500M

Polish DynaWord Mix — Extended (~500M tokens)

Maintained by Arkadiusz Słota · SlayerLab

A curated, openly-licensed Polish text corpus for language-model pretraining and research baselines. Built on top of the open polish-dynaword lineage and extended to ~500 million tokens (32k BPE) with additional curated, license-compatible sources and a documented cleaning + PII-scrubbing pipeline.

Why this exists: most large Polish web corpora are legal/parliamentary-heavy and carry uncurated PII from raw crawls. This dataset trades raw scale for traceable licensing, reduced stylistic bias, and verified PII removal — a corpus you can train on and publish results from without a legal or privacy footgun.

Dataset summary

Property	Value
Language	Polish (pl)
Tokens	~500,000,000 (32k BPE); validation ~10.2M
Sources	curated open-license mix (see Composition)
License	CC-BY-SA-4.0 (inherited) + attribution to parent sources
PII	email + phone scrubbed to [Email] / [Telefon] placeholders, independently verified
Format	Parquet (zstd), UTF-8 — mix section: one document per row; top-up section: one line per row (see Data format)

Data format (Parquet v2)

The corpus is published as Parquet (zstd-compressed), replacing the earlier plain-text train.txt / validation.txt. Row semantics follow the original corpus structure:

• mix section (polish-dynaword-mix, ~58.5M tokens): documents separated by blank lines → one row = one document (48,201 train / 2,536 validation documents).
• top-up section (DynaWord curated, ~441.5M tokens): the source text carried no document separators, so one row = one line (14,364,083 train / 152,646 validation lines).

Total rows: 14,412,284 train / 155,182 validation (column text).

Byte-fidelity: the parquet is a lossless repackaging of the previously published text — '\n\n'.join(mix_rows) + '\n'.join(topup_rows) reproduces the original train.txt / validation.txt byte-for-byte. Verified roundtrip SHA-256: train.parquet bdc2d3a0…401d (== previously pinned train.txt), validation.parquet d87026a1…68a5 (== previously pinned validation.txt). PII statements and token counts below are unchanged from the v1 card.

Composition

Exact token counts (32k BPE):

Source	Train tokens	Validation tokens
polish-dynaword-mix (source dataset), PII-scrubbed	58,542,031	3,417,717
DynaWord curated top-up	441,457,969	6,827,009
Total	500,000,000	10,244,726

Curated open Polish sources: Wikipedia, Wikisource, Wolne Lektury, Wikinews/Wikibooks/Wikivoyage; EUR-Lex/parliamentary capped for style balance.

Source proportions follow a quality-oriented sampling policy (temperature_alpha=0.5, legal/parliamentary capped at ≤15%) to reduce legalese over-representation while preserving fully traceable open licensing.

Curation & cleaning

• Deduplication: exact + normalized-text; validation held out with zero train↔val leak.
• Boilerplate removal: navigation, cookie banners, repeated headers/footers, template remnants.
• Structural hygiene: UTF-8 validated, control characters removed, sentence/paragraph structure preserved.
• PII scrubbing: emails and phone numbers (Polish + international) replaced with neutral placeholders [Email] / [Telefon] — not fake values (which would poison training and risk collisions with real numbers) and not typos (which leave identifiability).
• Scrubbing method (reproducible): normalize-first over separators (space, dot, all Unicode dashes incl. figure-dash, slash, non-breaking/thin spaces) rather than format-enumeration; strong/weak label detection with digit-length gating; structural false-positive guards (dates, time-ranges, historical year-ranges, software versions, currency/units, registry IDs — NIP/REGON/KRS/PESEL/IBAN). Deterministic, versioned, and test-covered; every clean artifact is SHA-256-pinned and independently re-verified.

PII statement

PII scrubbing: contact PII (email; phone incl. foreign +CC and space/dot/dash/en-dash/em-dash/figure-dash/slash separators, concatenated and multi-per-label formats) replaced with [Email]/[Telefon] placeholders. Independently verified by a second reviewer using a superset lens wider than the scrubber (check ⊋ scrub) with false-positive-guard self-audit: email-residual = 0, real-phone-residual = 0 on corpus train sha256 bdc2d3a012a5d37f721442519436dc5261cce6261d2e6359d8f1ea57bc8e401d / val sha256 d87026a1e938cec8ea0531da6ddcdd6b989384e4c28ad3f1723926e5c73a68a5, cross-confirmed by literal-grep (two-method agreement). Regex residual-tail (institutional/registry numbers, prices, software versions, historical year-ranges) confirmed non-PII by byte-level triage.

Totals: 27,639 phone placeholders, 1,124 email placeholders. Verification method and reproducible scrubber: see Curation & cleaning.

Licensing & attribution

Distributed under CC-BY-SA-4.0, inheriting the share-alike terms of the parent SlayerLab/polish-dynaword sources. Attribution to the original per-source providers is preserved; downstream use must retain attribution and share-alike.

Intended use & limitations

• Intended: pretraining/finetuning small–medium Polish LMs, reproducible baselines, tokenizer and data-mix research.
• Limitations: small-to-medium scale (not web-scale); placeholder tokens replace contact details; residual rare edge-cases may remain in the long tail (documented in the PII statement); not intended for tasks requiring verbatim contact data.

Citation

@misc{polish_dynaword_mix_extended,
  title     = {Polish DynaWord Mix — Extended (~500M tokens)},
  author    = {Arkadiusz S{\l}ota},
  publisher = {SlayerLab},
  year      = {2026},
  note      = {CC-BY-SA-4.0; extended and PII-scrubbed from the polish-dynaword lineage}
}