Add source: european_hplt_v3_pl (HPLT v3.0 Polish, top WDS bins, CC0)

#5
by Maggio33 - opened

First dataset contribution — review requested 🙏

This is my first dataset PR to polish-dynaword. Please review carefully — happy to adjust
anything (thresholds, schema, naming, provenance wording).

What: european_hplt_v3_pl — filtered Polish subset of HPLT v3.0 (pol_Latn), the two top
Web-Docs-Scorer quality bins (WDS = 10 + 9), 2 shards.

documents tokens (cl100k proxy)
bin-10 (european_hplt_v3_pl.parquet) 55,590 261,468,813
bin-9 (european_hplt_v3_pl-part-b9.parquet) 55,000 250,914,965
total 110,590 512,383,778

Why: the v0.3 candidate config lists only HPLT v1 (european-hplt-v1, ~1.98B PL tokens).
HPLT v3.0 is a newer, much larger CC0 release (pol_Latn ≈ 270B raw tokens) with improved
extraction (Trafilatura 2.0), global dedup, and WDS quality bins. This contributes the cleaned top
bins; the same pipeline scales to bins 9→5.

How it was made (src/clean_hplt_v3.py, thresholds inherited from src/filter_european_hplt.py):

  1. Stream the raw HPLT v3 pol_Latn shards 10_1 + 9_1 (jsonl.zst) from data.hplt-project.org (NOT SpeakLeash).
  2. Filter per document: lang=pol_Latn & prob≥0.80, MT-prob<0.20, register ∈ {ID,OP,HI,IN,NA},
    ≥400 chars & ≥80 words, drop boilerplate/adult/legalish (anti-legalese goal), length cap 120k chars
    (removes book-length copyright/OCR risk), mojibake drop (upstream HPLT UTF-8→GBK mis-decode, e.g. ż→偶),
    exclude wikipedia/wikisource/wikimedia + file-host domains, per-domain cap 250.
  3. Emit canonical schema (id/text/source/added/created/token_count/license/author), token_count via tiktoken cl100k.
  4. src/test_hplt_v3_contract.py passes (schema + open license + non-empty text + token>0 + stats consistency);
    mojibake residual 0.

Quality (surface metrics vs accepted sources): MT-prob kept ~0.08, boilerplate ~0.2%, diacritic-ratio ~0.066,
TTR@500 ~0.64 (within accepted range, ≈ parliamentary); bin-9 ≈ bin-10 on every metric.

Honest caveats (please weigh in review):

  • Web-crawl → accept_after_web_review, not release-clean. CC0 is on the packaging; underlying web text
    may include in-copyright material that recurs at all lengths — the length cap only removes obvious books.
  • Near-dup NOT removed (per Dynaword "heavy filtering downstream"): ~11.8% 8-gram overlap with plwiki measured;
    wikipedia/wikisource domains excluded to mitigate, but cross-source containment dedup vs the full corpus is left downstream.
  • src/sources.py entry added so the release build recognizes the source and lifts the CC0 license from the registry.
Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment