Datasets:

---

SlayerLab

/

polish-dynaword

Like 24

Follow

Fabryka AI 84

Summary

Adds a pinned, normalized snapshot of Polish court judgments from PiotrSty/saos-polish-court-judgments as the new saos source.

This PR does not claim to contain the complete current SAOS database. The pinned source has 296,692 rows, while the official SAOS API returned 540,726 items on 2026-09-04. That limitation is represented in the datasheet and as addressable Evidence, a falsifiable Claim, and an evidence-backed attestation.

The final build retains 296,689 documents, containing 5,897,001,250 characters and 2,265,448,504 cl100k_base tokens. Three records below 200 characters were removed; no exact normalized duplicates or language-gate failures were found. Courts: COMMON 229,370 / SUPREME 38,058 / NATIONAL_APPEAL_CHAMBER (KIO) 19,758 / CONSTITUTIONAL_TRIBUNAL 9,503 rows.

Scope cleanup (review items 1–3)

Replaced by a single clean commit on main @ 34e7ea3: only this source's files (data/saos/*, src/fetch_saos.py, src/test_saos_contract.py, the SAOS artifacts in artifacts/) plus the registry entry with release: None and a source_findings.md entry. The stale-branch drift is gone — no european_hplt_v3_pl re-add, no uzp_orzeczenia_pl / rock_pollub_pl deletions, no scrub_pii.py / scrub_entities.py, no README / CHANGELOG / VERSION / make_docs / AGENTS.md / sejm-fetcher edits.

Provenance and legal status

• Official origin: https://www.saos.org.pl/api/dump/judgments (API docs: https://www.saos.org.pl/help/index.php/dokumentacja-api/api-pobierania-danych)
• Pinned source: PiotrSty/saos-polish-court-judgments@736cba4f92999e29d3e49409b0713c88fc683714
• License label: public-domain (official documents) / LicenseRef-Polish-Official-Documents
• Legal basis documented: Polish Copyright Act art. 4(2) and the Polish Open Data Act.
• Formal legal review: not performed — scope and responsible party below.

Legal review — what is needed and who does it (review item 5)

Needed before release, on the merged corpus as used for training:

1. Copyright-status confirmation of the art. 4(2) exclusion as applied to full judgment texts, including third-party quoted material inside judgments (party submissions, cited reports, quoted correspondence) which may carry separate rights;
2. Privacy/GDPR assessment of named private persons in published judgments (parties, attorneys, witnesses) for TDM reuse — including the CJEU/Polish case-law overlay on court-publication practice;
3. Provider-side terms check (SAOS / Ministry of Justice redistribution terms).

Responsible: the corpus maintainer's legal reviewer with a Polish legal background (bar/counsel or the institution's legal team). The engineering side prepares only the evidence pack: per-record SAOS URLs + court metadata (attribution sidecar), the saos.license-evidence.json snapshot, and sample spot-checks — none of which substitutes for the opinion. Until it lands: mergeable with release: None, excluded from stable releases.

Overlap with uzp_orzeczenia_pl (review item 4)

uzp_orzeczenia_pl holds KIO / appellate / supreme procurement rulings; this snapshot is general jurisdiction, and its NATIONAL_APPEAL_CHAMBER slice (19,758 rows) is exactly the KIO family. Measured two ways:

1. Verbatim text (normalized exact-line/paragraph containment, ≥30/≥100 chars, 64-bit blake2b, exact): 0 matches in both directions — no record appears verbatim in the other corpus (5.90B vs 1.07B chars checked; note saos records are single-line texts, so this test is whole-record verbatim matching).
2. Case-number cross-check — the precise question: all 19,758 KIO rows carry case_numbers, and 4,536 of them (22.9% of the KIO slice; 1.5% of the whole snapshot) match a signature in uzp_orzeczenia_pl — e.g. KIO/UZP 5/08, KIO 2741/10. The same decisions are published in both corpora via different extraction paths (SAOS database vs KIO PDFs), so their texts differ verbatim but are the same rulings — a near-duplicate risk at integration.

Integration gate (concrete): drop or pairwise-collapse the 4,536 saos KIO records whose normalized case_numbers intersect uzp_orzeczenia_pl's signature (both keys sit in the attribution sidecars — a mechanical filter), or run near-dedup restricted to that subset. Method + tooling in the review thread.

Processing (original)

• NFKC Unicode normalization, normalized spaces/quotes/dashes and preserved paragraphs.
• Article and section numbering retained as substantive legal content.
• Minimum 200 characters and DynaWord Polish-diacritic ratio gate.
• Exact normalized-text deduplication.
• Pattern scrubbing for email, PESEL candidates, labelled NIP/REGON, labelled phones, and bank accounts (14 / 253 / 60 / 66 strings replaced). Not NERGAL — a NERGAL pass is a release gate (judgments name private persons; boundary is a legal question, not a regex question).
• Implausible dates preserved raw in the attribution sidecar and left empty in the training schema.
• Per-record SAOS URL, court type, judgment type, case numbers and judge attribution preserved in the sidecar.

Registry

Registry entry kept with release: None (now with added: "2026-09-04"); src/sources.py diff is the entry only.

Test plan

• python3 -m pytest src/test_saos_contract.py — passed on the clean branch
• Text-level overlap vs uzp_orzeczenia_pl measured (above)
• Legal-review scope + responsible party stated (above, datasheet)
• Reviewer: scope cleanup accepted
• Maintainer: legal review scheduled before release