Datasets:

SlayerLab
/

polish-dynaword

@@ -1,59 +0,0 @@
-# european_hplt_v3_pl
-Filtered Polish subset of **HPLT v3.0** (`pol_Latn`), top quality bins (**WDS = 10 down to 5**), merged into one file.
-## Dataset description
-- **Source (upstream):** https://hplt-project.org/datasets/v3.0 (`pol_Latn`, WDS shards `10_1`, `9_1`–`9_4`, `8_1`–`8_4`, `7_1`–`7_4`, `6_1`–`6_2`, `5_1`–`5_2`)
-- **Domain:** web (modern natural Polish — informational, narrative, opinion, how-to, discussion)
-- **Language:** Polish (pl)
-- **License:** `CC0-1.0` (packaging; see licensing note)
-- **Created (range):** web crawls 2012-2024 (Internet Archive + Common Crawl)
-- **Added:** 2026-07-14 (WDS 10+9); expanded 2026-08-07 (WDS 8-5 + additional WDS-9 shards)
-## Licensing - traceable basis
-HPLT licenses the **packaging** of these text data under Creative Commons CC0 1.0 ("no rights reserved"),
-per the HPLT v3.0 Terms of Use (hplt-project.org/datasets/v3.0) and the HF card `HPLT/HPLT3.0`.
-HPLT does not own the underlying web text; a notice-and-takedown policy applies and downstream users
-remain responsible for compliance (EU CDSM Directive 2019/790, GDPR). This is a web-crawl source: it is
-offered at parity with the already-listed `european-hplt-v1` (CC0) basis and still needs maintainer
-web/legal review (`accept_after_web_review`), not auto-accept.
-## Provenance
-Fetched directly from the HPLT distribution (`data.hplt-project.org/three/sorted/pol_Latn/{shard}.jsonl.zst`),
-NOT via SpeakLeash. HPLT v3 is globally deduplicated upstream and sorted into Web-Docs-Scorer (WDS) quality
-bins. Per-document URL, WDS score, web-register, MT-probability and PII spans
-are available in the raw HPLT record and were used for filtering (not retained in the released canonical schema).
-## Statistics
-| documents | tokens (tiktoken cl100k proxy) | license |
-|---:|---:|---|
-| 1,603,974 | 2,555,525,652 | 100% `CC0-1.0` |
-Composition by WDS bin (documents; single merged parquet):
-| WDS bin | shards | documents | added |
-|---:|---|---:|---|
-| 10 | `10_1` | 55,590 | 2026-07-14 |
-| 9 | `9_1` (+`9_2`,`9_3`,`9_4`) | 55,000 + 298,543 | 07-14 / 08-07 |
-| 8 | `8_1`–`8_4` | 398,795 | 2026-08-07 |
-| 7 | `7_1`–`7_4` | 398,431 | 2026-08-07 |
-| 6 | `6_1`,`6_2` | 199,038 | 2026-08-07 |
-| 5 | `5_1`,`5_2` | 198,577 | 2026-08-07 |
-Per-shard char/token/drop breakdown in `*.stats.json`. Higher WDS = higher web-doc quality; lower bins add volume with modestly lower mean quality.
-## Filters applied (src/clean_hplt_v3.py + run_pipeline.py)
-Per-shard keep ~28-40% (higher for lower bins). Drop reasons:
-`register` (not ID/OP/HI/IN/NA), `machine_translated` (MT>=0.20), `domain`
-(wikipedia/wikisource/wikimedia + file-hosts excluded), `adult_spam`, `legalish` (drops legal/parliamentary
-phrasing - serves the anti-legalese goal), `boilerplate`, `too_long` (>120k chars = book-length copyright/OCR
-risk), `mojibake` (>=3 CJK/replacement chars), `filehost`, `domain_cap`. MT-probability of kept docs: mean ~0.08. Mojibake residual: 0.
-## PII phone-scrub (v12b) & cross-source dedup
-- **Phones → `[Telefon]`** (semantic-tag, not fake-number — avoids model-poisoning). Iterative scrub v6→v12b closed all systematic phone-recall classes (address-block-suppression, `dzwoń`-label family, leading-0 landline, `+48`-paren, newline-over-span, foreign-UK-bare). Residual **0.0135%** = genuine clearance-tail (FP-not-phone / foreign-out-of-scope / bare-no-leading-0-ambiguous / label-distance / institutional-published / rare-edge). Email/national-ID → `[PII]`.
-- **Verification:** TWO independent lenses (deep-verify + broad recall-scan), NOT pipeline self-report — a self-report (idempotency) blind-spot had masked a systematic recall-gap; the independent gate caught it pre-release (SAFETY-INVARIANT: check-coverage ≥ scrub-coverage).
-- **Cross-source dedup vs the existing WDS-10+9 file:** expansion deduplicated against the base shards — **exact-dup = 0, near-dup overlap = 0.0000% (byte-verified)**. This resolves the "cross-source near/containment dedup remains a downstream step" caveat from the original release.
-## Status & caveats (web-crawl - NOT "release-clean")
-- **Copyright is not fully solved by filtering.** In-copyright text recurs in web-crawl at ALL lengths (news reprints, blog reposts, excerpts). Inherent to web sources; the reason HPLT ships CC0-on-packaging + notice-and-takedown.
-- Surface-metric parity with accepted sources: mojibake 0%, boilerplate ~0.2%. ~9% commercial/SEO-adjacent pages (web-inherent).
-- Lower WDS bins (8-5) add volume; a domain-cap keeps per-domain skew bounded. The full `pol_Latn` corpus is ~270B raw tokens across WDS bins 10->5.