dataseek/ptbr-wiki

PT-BR Wikipedia (cleaned, encyclopedia only)

Part of the MagTina350m pretrain corpus release by Dataseek under the Magestic.ai brand. This is one of nine silver-layer datasets that fed `dataseek/magtina350m-base`.

Summary

1.08 M Brazilian-Portuguese Wikipedia articles, cleaned of wiki-markup, templates, user-talk welcome banners ("Bem-vindo, X!"), VfD voting discussions, eliminação notification templates, JS/CSS gadget pages and talk-page conversations (detected by inline HHhMMmin timestamps). 35 % of the raw silver was discarded as MediaWiki-internal content. Used at ~2 epochs in MagTina350m pretrain.

Source and collection method

Source: PT-BR Wikipedia dump (mid-2026 snapshot) → mwparserfromhell → template/reference strip → NFKD normalisation → langid='pt' gate (catches mistaken en/es entries) → length floor 200 chars → v3 cleanup pass (2026-05-12) — see `scripts/release/cleanup_and_export_wiki.py` — that drops welcome banners, VfD ballots, talk-page timestamp signatures, eliminação notification templates and gadget JS/CSS pages.

ETL script (in the MagTina1B repository): `scripts/etl/recover_wikipt_cleaned.py` (public release of the ETL scripts is on the roadmap; until then the data card below documents the recipe in full).

Filters and deduplication

The following filters were applied before this dataset reached its silver (release-ready) state:

• —Strip wiki templates, refs, infoboxes (initial ETL)
• —FastText langid='pt'
• —v3 cleanup: 22 pattern filters covering welcome banners, talk pages, VfD votes, eliminação notifications, gadget JS/CSS, redirects
• —Inline HHhMMmin de DD de MMM de YYYY timestamp filter (catches all talk-page conversation content not already covered)
• —len(text) ≥ 200 chars (silver gate, preserved through cleanup)

Global URL-normalised deduplication was applied across all web-derived corpora (webpages, news, blogs) so the same article does not appear twice across those three datasets.

Schema

Columns dropped at export (kept private as ETL internals): _part

Size statistics

Sampled at ~2.8 epochs — the silver corpus has 581.25 M unique tokens, of which 1.601 B were drawn during MagTina350m's single pretrain pass.

Used in MagTina350m pretrain: 1.601 B tokens (9.2 % of MagTina350m's 17.39 B-token pretrain budget).

How to load

from datasets import load_dataset

ds = load_dataset("dataseek/ptbr-wiki", split="train", streaming=True)
for row in ds.take(5):
    print(row["text"][:200])

Streaming is recommended for the larger configs. For the smaller datasets (ptbr-dou, ptbr-books-publicos) eager loading is fine.

Licensing

CC-BY-SA 4.0 — inherited from Wikipedia. Any derivative work that incorporates this dataset must also be licensed CC-BY-SA-compatible.

Upstream attribution: Wikimedia Foundation — https://dumps.wikimedia.org/ptwiki/ ; individual contributors retain copyright per CC-BY-SA 4.0.

Citation

If you use this dataset, please cite both the upstream source and MagTina350m:

@misc{magtina350m_pretrain_2026,
  title  = {MagTina350m pretrain corpus — PT-BR Wikipedia (cleaned, encyclopedia only)},
  author = {Frasson, Ricardo and {Dataseek Team}},
  year   = 2026,
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/dataseek/ptbr-wiki}
}

Please also honour the upstream license terms — for CC-BY-derived data, attribution to the upstream creators is mandatory; for CC-BY-SA, downstream derivatives must remain CC-BY-SA-compatible.

Intended use

• —Pre-training, continued pre-training, or domain-adapting of Brazilian Portuguese language models.
• —PT-BR NLP research where statistically representative public-web / academic / legal / encyclopedic data is needed.
• —Reproducing or improving on the MagTina350m result.

Known limitations and PII statement

• —Text was NOT PII-scrubbed. URLs, emails, phone numbers and personal names that occurred in the source data may still be present. We strip zero-width characters and normalise Unicode but we do not run an NER pass.
• —Crawled data carries upstream biases of CommonCrawl, Wikipedia, news outlets and academic institutions present in the source. We have not audited these.
• —No safety filtering beyond langid and basic alpha-ratio gates. Hate-speech, spam and adult content present in the source remain unless caught incidentally.
• —Provenance preserved at row level. Every row has either a url, source or doc_id column that points back to upstream — this is intentional, so consumers can re-license, redact or filter.

Related releases

• —Model: `dataseek/magtina350m-base` (354.6 M params, pretrained on this corpus + 8 sibling datasets)
• —Instruct model: `dataseek/magtina350m-instruct`
• —Sibling datasets: see dataseek/ptbr-* for all nine corpora

License

cc-by-sa-4.0