Datasets:

---

SlayerLab

/

polish-dynaword

Like 24

Follow

Fabryka AI 84

Add Polish TED public-procurement notices

Adds ted_pl: 15,544 documents and 42,475,098 measured cl100k_base proxy tokens from official Polish-language TED procurement-notice records (2023-02-15 to 2024-05-15). The corpus uses 20,588 official Search API records and 1,255 official XML fallbacks.

Source dataset: https://huggingface.co/datasets/PiotrSty/ted-polish-procurement-notices/tree/c2075874d36035e6cbc29b252155ac1da299cd13

What the texts are

Genre: public-procurement notices — tender titles, object descriptions, lots, technical specifications, selection/award criteria, deadlines, review-appeal instructions (often quoting the Public Procurement Law). Register: administrative/technical Polish — dense formal prose, numbered clauses, product and standard designations, statutory citations. The contracting authority is the author field (100% coverage). Time span: publication dates 2023-02-15 – 2024-05-15. Rebuilt from structured API fields — not OCR'd, not machine-translated.

Data samples (with metadata)

id: ted_pl_261467-2024
source: ted_pl
added: 2026-09-08
created: 2024-05-02
token_count: 1500
license: LicenseRef-TED-Notice-Reuse
author: Gmina Miasto Rzeszów - Urząd Miasta Rzeszowa
notice_id: 261467-2024
url: https://ted.europa.eu/en/notice/261467-2024/xml

Opracowanie dokumentacji projektowej dla zadania pn.: „Budowa parku wodnego w Rzeszowie”

Przedmiotem zamówienia jest opracowanie kompletnej dokumentacji projektowej wraz z uzyskaniem pozwolenia na budowę dla obiektu rekreacji wodnej zwanego dalej „Park Wodny w Rzeszowie”, „Obiekt”, „Park Wodny”. Obiekt ma składać się z części krytej (w tym centrum edukacyjnego i saunarium) i części zewnętrznej (sezonowy basen odkryty) wraz z infrastrukturą towarzyszącą oraz zagospodarowaniem terenu wokół Parku Wodnego. Powierzchnia lustra wody krytej części Parku Wodnego co najmniej 1600 m2 (w tym całoroczny basen wypływowy o powierzchni lustra wody co najmniej 50 m2. Część zewnętrzna Parku Wodnego powinna

id: ted_pl_35480-2024
source: ted_pl
added: 2026-09-08
created: 2024-01-18
token_count: 1499
license: LicenseRef-TED-Notice-Reuse
author: Narodowy Fundusz Zdrowia Centrala
notice_id: 35480-2024
url: https://ted.europa.eu/en/notice/35480-2024/xml

Zakup usługi wsparcia technicznego i merytorycznego oprogramowania do replikacji i optymalizacji serwerów dziedzinowych IBM i oraz warsztaty szkoleniowe

Zakup usługi wsparcia technicznego i merytorycznego oprogramowania do replikacji i optymalizacji serwerów dziedzinowych IBM i oraz warsztaty szkoleniowe zgodnie z Załącznikiem nr 1 do SWZ.

Liczba dodatkowych osób dedykowanych do realizacji zamówienia posiadających umiejętności niezbędne dla uzyskania co najmniej jednego z następujących certyfikatów: MIMIX v9.0 Implementation Certification lub certyfikat MIMIX v10.0 Implementation Certification, potwierdzone tym dokumentem lub innym równoważnym dokumentem, który potwierdza umiejętności wyma

id: ted_pl_271535-2024
source: ted_pl
added: 2026-09-08
created: 2024-05-08
token_count: 1499
license: LicenseRef-TED-Notice-Reuse
author: Tramwaje Warszawskie sp. z o.o.
notice_id: 271535-2024
url: https://ted.europa.eu/en/notice/271535-2024/xml

Dostawa nagrzewnic tramwajowych wraz z częściami zamiennymi do nich, stanowiących wyposażenie tramwajów typu 105N, 120N, 120Na, 123N i pochodnych oraz typu 116Na i pochodnych (4 części).

Dostawa nagrzewnic tramwajowych wraz z częściami zamiennymi do nich, stanowiących wyposażenie tramwajów typu 105N, 120N, 120Na, 123N i pochodnych oraz typu 116Na i pochodnych (4 części), w tym: 1) w części nr 1 – dostawa nagrzewnic i części zamiennych do nich, stanowiących wyposażenie tramwajów typu 105N, 116Na, 123N i 120N; 2) w części nr 2 - dostawa nagrzewnic tramwajowych przedziału pasażerskiego, stanowiących wyposażenie tramwajów typu 120Na; 3) w części nr 3 - dostawa nagrzewnic tramwajowych oraz zabez

Rights and method

TED's official legal notice says that, unless otherwise noted, procurement notices may be freely reused for commercial or non-commercial purposes. It separately assigns CC0 to metadata; this contribution therefore records notice text as LicenseRef-TED-Notice-Reuse rather than silently relabelling it CC0. The legal instrument behind the reuse grant is Commission Decision 2011/833/EU of 21 December 2011 on the reuse of Commission documents: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32011D0833. Third-party-work and identifiable-person exceptions are scoped in "Legal-exception and name review" below.

The pinned PleIAs mirror is used only to enumerate Polish-labelled identifiers because its Polish preview contains replacement-character encoding damage. Every released text is rebuilt from a checksum-addressed official TED Search API record or, where narrative fields are absent, an official XML fallback. Structured contact/address fields are omitted; email, phone and labelled national-ID patterns are replaced. Frequent exact boilerplate and exact/near duplicates are removed.

Filtering (gates and counts)

• Acquire: 20,588 official Search API records (1,255 via official XML fallback); any record with a U+FFFD replacement character after cleaning is rejected (0 in this run).
• Text: narrative fields only (*-proc / *-lot / *-part / *-glo title/description/justification families); structured Organization/Company/PartyName/PostalAddress/Contact/PartyLegalEntity/Person blocks excluded by construction.
• Boilerplate: chunks occurring in >= max(100, 1% of documents) dropped.
• Gates: contracting authority present (−0); text >= 300 characters after cleaning (−1,014); langid Polish — 7-language set pl/en/de/cs/sk/uk/ru on the first 20,000 chars, reject at confidence >= 0.98 (−0); exact dedup on normalized text (−2,350); near dedup — 5-word shingles, MinHash LSH 128 perms / 0.86 candidates, exact Jaccard >= 0.9 (−1,680).
• Total dropped: 5,044 of 20,588 → 15,544 kept.

PII scrubber and spot check

Scrubber run: pattern redaction in src/build_ted_pl.py (e-mail -> [PII], phone -> [Telefon], labelled PESEL/NIP/REGON -> [PII]) — not SlayerLab/NERGAL. Spot check (60 of 15,544 records, seed 20260928) found one real residual contact: ted_pl_188212-2024 — "skontaktować się z A.Decem - 815384157, e-mail: [PII]" — a personal name next to an unlabelled 9-digit phone the phone pattern missed (the e-mail beside it WAS redacted). Other hits benign: 52 institutional delivery addresses, 3 statutory PESEL/NIP mentions, 0 e-mail leftovers. No release before a NERGAL pass; merge with release: None is fine meanwhile.

Legal-exception and name review (scoped)

Bounded to the 60-record sample (per-notice coverage remains a release gate): free-text name cues were statutory phrasing only ("imię i nazwisko" in the Public Procurement Law's list of what an appeal must contain) — 0 concrete personal names; third-party markers were product names (Windows Server, Microsoft 365/Canva licences), standard numbers (ISO 9001, PN-EN 14065) and legal citations/titles ("Ustawa o lasach", Dz.U.) — no reproduced third-party texts. The residual contact above is the known risk.

Overlap with uzp_orzeczenia_pl and inclusion

uzp_orzeczenia_pl is the same policy domain but a different document class — ex-post appellate rulings (KIO/SA/SN); this slice is ex-ante tender notices. Measured text-level overlap (normalized exact-line/paragraph containment, ≥30/≥100 chars, verbatim only): 1,082 of 191,353 lines (94,840 chars = 0.085% of this slice's chars), 0 paragraphs; reverse direction 0.01% of uzp_orzeczenia_pl chars — the shared lines are statutory procurement-law citations common to both registers. Net: textually near-disjoint and complementary stages of the procurement lifecycle; ordinary cross-source dedup at integration is enough.

Slayer ontology and remaining gates

The manifest separates content-addressed source/dataset Versions, Protocols, executed Runs, Evidence, falsifiable Claims, Actors and typed lineage. Per-notice legal-exception review, full free-text name review, cross-source DynaWord deduplication, benchmark-overlap testing and a deliberate training-mix cap/ablation remain pending (release gates). This PR proposes a source; it does not claim a merged or stable release.

Test plan

• python3 -m pytest src/test_ted_pl_contribution.py — 2 passed (rebased onto main @ 34e7ea3)
• Residual-PII / name / third-party spot check (above, one real residual contact)
• Text-level overlap vs uzp_orzeczenia_pl measured (above)
• License basis cited (TED legal notice + Commission Decision 2011/833/EU)
• Reviewer: LicenseRef-TED-Notice-Reuse treatment accepted
• Reviewer: scoped name/exception review accepted

Review against AGENTS.md ("What the PR must show", "Normalize the text before you build"). Needed before merge:

1. Rebase onto current main. Against main the diff currently shows european_hplt_v3_pl being re-added (removed from main for unverified license) and uzp_orzeczenia_pl / rock_pollub_pl being deleted. git fetch of this PR's ref currently fails on our side; a fresh push after the rebase should fix it.
2. Drop src/scrub_pii.py / src/scrub_entities.py. In-repo scrubbers were removed in #38; PII goes through SlayerLab/NERGAL. Merge with release: None is fine meanwhile, but no release before a NERGAL pass.
3. Keep the PR to this source. Drop edits to README / CHANGELOG / VERSION in make_docs and to shared files (AGENTS.md, clean_hplt_v3.py, the sejm fetchers). Registry entry stays, with release: None; the release bookkeeping is done once by the maintainer after merges.
4. License basis. Cite the legal instrument behind TED reuse (Commission Decision 2011/833/EU), not only the TED legal-notice page.
5. Samples are inline, good — add the metadata columns (id, url, date) to each block.
6. Overlap and inclusion. Address overlap with uzp_orzeczenia_pl (same procurement domain) and state what notices add to the corpus.
7. Filtering. Give min-length / language-ID / dedup details and counts dropped.
8. The description lists legal-exception and name review as pending — please resolve or scope them before this is reviewed for merge.

Correction to my review comment above: the git fetch failure was on the Hub side (have/ack negotiation), not a problem with this PR. It fetches fine with git -c fetch.negotiationAlgorithm=noop fetch. No re-push needed for that; the rebase ask still stands.

Review addressed — pushed as 7dc8fa3 (rebase) + 167a10c (review fixes) on refs/pr/32, rebased onto main @ 34e7ea3. (And acknowledged on the fetch note — fetch.negotiationAlgorithm=noop works from here.) Point by point:

1. Rebased onto current main. The two-dot diff is now scoped to this source: no european_hplt_v3_pl re-add, no uzp_orzeczenia_pl / rock_pollub_pl deletions — 12 files, +18,104, 0 deletions. The stale whole-file rewrites of src/sources.py and artifacts/source_findings.md are resolved to main's versions (the findings file is byte-identical to main in this PR; the new entry lands in the fix commit).
2. Scrubbers gone — src/scrub_pii.py / src/scrub_entities.py were never in this PR's commits (branch predates #38); the rebase removes them from the diff. PII story in the datasheet: pattern redaction in src/build_ted_pl.py (112-class contact patterns -> [PII]/[Telefon]) and no release before a NERGAL pass. Registry entry kept with release: None (added added: "2026-09-08" — the entry carried neither added nor release).
3. Scoped to this source. No README / CHANGELOG / VERSION / make_docs / AGENTS.md / clean_hplt_v3.py / sejm-fetcher edits remain (diff = data/ted_pl/*, src/build_ted_pl.py, src/test_ted_pl_contribution.py, the registry entry, the ontology manifest, the findings entry).
4. License basis. The description and datasheet now cite the legal instrument: Commission Decision 2011/833/EU of 21 December 2011 on the reuse of Commission documents (https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32011D0833), with the TED legal notice as the publisher's implementing statement; text stays LicenseRef-TED-Notice-Reuse, metadata CC0. The registry traceable field names the Decision too.
5. Samples with metadata. The 3 inline blocks now carry id, source, added, created, token_count, license, author, notice_id, url (the pinned official XML URL per notice) before the text.
6. Overlap and inclusion. Measured text-level overlap with uzp_orzeczenia_pl (normalized exact-line/paragraph containment, ≥30/≥100 chars, verbatim only): 1,082 of 191,353 lines — 94,840 chars = 0.085% of this slice's chars, 0 paragraphs; reverse direction 0.01% of uzp chars; the shared lines are statutory procurement-law citations. Inclusion argument: uzp_orzeczenia_pl is ex-post appellate rulings (KIO/SA/SN), this is ex-ante tender notices — complementary stages of the procurement lifecycle, near-disjoint textually.
7. Filtering details (now in the datasheet and description): boilerplate chunks ≥ max(100, 1% docs) dropped; gates — authority present (−0), min 300 chars after cleaning (−1,014), langid Polish (7-language set pl/en/de/cs/sk/uk/ru on first 20k chars, reject at conf ≥ 0.98, −0), U+FFFD rejection (−0), exact normalized dedup (−2,350), near dedup (5-word shingles, MinHash LSH 128 perms/0.86 candidates, exact Jaccard ≥ 0.9, −1,680). Total 5,044 of 20,588 dropped → 15,544 kept.
8. Legal-exception and name review — scoped with evidence. 60-record spot check (seed 20260928): 0 concrete personal names (the imię i nazwisko cues are the Public Procurement Law's statutory phrasing about appeals), third-party markers = product names, standard numbers and legal citations — no reproduced third-party texts. One real residual contact found and reported: ted_pl_188212-2024 ("A.Decem - 815384157" — name + unlabelled phone the patterns missed). Scope statement: per-notice coverage stays a release gate (NERGAL + per-notice legal review); merge-scope is resolved by this bounded review.

Verification: python3 -m pytest src/test_ted_pl_contribution.py — 2 passed on the rebased branch.