data/ted_pl/ted_pl.md

metadata

license: other
license_name: ted-notice-reuse-terms
language:
  - pl
task_categories:
  - text-generation
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-00000-of-00001.parquet

TED Polish procurement notices

Polish narrative text rebuilt from official TED procurement-notice records.

• Indexed notices: 20,588
• Search API records: 20,588
• XML fallbacks: 1,255
• Retained documents: 15,544
• Tokens: 42,475,098 (cl100k_base proxy)
• Dates: 2023-02-15 to 2024-05-15
• Contracting-authority attribution: 100.0%

The pinned PleIAs mirror is an identifier index only because its Polish preview contains replacement-character encoding damage. Released text comes from official TED Search API narrative fields or, when absent, official XML; any remaining U+FFFD record is rejected. Structured contact/address fields are omitted; emails, phone numbers and labelled national identifiers in narrative fields are replaced. Recurrent boilerplate and within-source duplicates are filtered.

See NOTICE.md and artifacts/ for the source index, attribution, selection decisions, QA, checksums, full samples and Slayer ontology manifest.

License basis

The TED legal notice permits free reuse of procurement notices, for commercial or non-commercial purposes, unless otherwise noted; it assigns CC0 to metadata separately. The legal instrument behind that 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). Notice text is recorded as LicenseRef-TED-Notice-Reuse rather than relabelled CC0. Exceptions to the reuse grant (third-party works, identifiable private individuals) are handled in "Legal-exception and name review" below.

What the texts are

• Genre: public-procurement notices - tender titles, object descriptions, lots, technical specifications, selection and 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. Contracting authority is the author field (100% coverage). Not OCR'd (rebuilt from structured API fields), not machine-translated.
• Time span: publication dates 2023-02-15 - 2024-05-15.

Filtering (gates and counts)

All steps in src/build_ted_pl.py against the pinned acquisition:

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

PII scrubber and spot check

Scrubber that ran: pattern redaction in src/build_ted_pl.py - e-mail -> [PII], phone -> [Telefon], labelled PESEL/NIP/REGON -> [PII]. This is not SlayerLab/NERGAL; the shard must be re-run through NERGAL before it is admitted to a release. Merge with release: None is fine meanwhile.

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 were benign: institutional delivery addresses (ul./al. - 52), statutory mentions of PESEL/NIP (3), no e-mail leftovers. Residual risk: moderate - unlabelled numbers beside names defeat label-anchored patterns.

Legal-exception and name review (scoped)

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

Overlap with uzp_orzeczenia_pl and inclusion argument

uzp_orzeczenia_pl covers the same policy domain but a different document class: appellate rulings on public procurement (KIO/SA/SN), ex-post. 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 boilerplate (procurement-law citations common to both registers). Net: the slices are textually near-disjoint and complementary in the procurement lifecycle; ordinary cross-source dedup at integration is enough.

Limitations

Free-text personal names are not comprehensively de-identified. The official reuse notice has exceptions for third-party works and identifiable private individuals, so per-notice/legal review remains appropriate. Cross-source DynaWord deduplication, benchmark-overlap checks and controlled training ablations remain pending. This administrative/technical register should be capped deliberately in a training mix.