PiotrSty/law-turkish

Turkish statutes - article chunks (law_tr)

The 106 most widely used Turkish statutes (kanun), article-level text chunks, re-derived from CtnkyaABC/turkish-law-corpus with per-record provenance to the official source mevzuat.gov.tr (Turkish legislation portal). Register: formal legal/administrative Turkish - articles, subsections and consolidated amendment notes.

• Source rows: 16,001 (106 statutes)
• Retained records: 14,139 (88.4%)
• Measured cl100k_base proxy tokens: 5,363,447
• Characters: 13,752,930
• Pinned source snapshot: hf://datasets/CtnkyaABC/turkish-law-corpus (corpus/corpus.csv, 34,125,450 bytes, sha256 in artifacts/run.json upstream chain via artifacts/attribution.jsonl URLs)

License - evidence at the origin

• The texts are official Turkish legal texts (statutes published by the Ministry of Justice legislation portal https://www.mevzuat.gov.tr/); every record carries its official source PDF URL (e.g. https://www.mevzuat.gov.tr/MevzuatMetin/1.4.193.pdf) and statute number in artifacts/attribution.jsonl. Official-works treatment under FSEK (Act No. 5846) [INFERENCE: statutory reading, not counsel-verified]. Per-record license field: LicenseRef-PD-Official.
• The upstream packaging declares CC BY 4.0 (https://huggingface.co/datasets/CtnkyaABC/turkish-law-corpus) over its collection; that grant covers the compilation as delivered and is preserved here as provenance, not as a verified rights review of the statute texts.
• Chain-of-title gaps: the upstream collection code is not published (scraping from mevzuat.gov.tr described in its card only); statute consolidation dates are not present in the rows (created is empty); no cross-source dedup against other corpus shards. release: None pending the standard review.

Excluded from upstream

qa/qa.csv (5,011 question-answer pairs) is not included: its source column says generated - synthetic QA text, not source documents.

Samples

Verbatim records from artifacts/sample.jsonl, real line breaks preserved, truncated with [...].

Sample 1

id: law_tr_Karayollar-Trafik-Kanunu-madde-47-chunk-0
source: law_tr
token_count: 927
author: Karayolları Trafik Kanunu

Karayollarından faydalananlar aşağıdaki sıralamaya göre; a) Trafiği düzenleme ve denetimle görevli trafik zabıtası veya özel kıyafetli veya işaret taşıyan diğer yetkili kişilerin uyarı ve işaretlerine, b) Trafik ışıklarına, c) Trafik işaret levhaları, cihazları ve yer işaretlemeleri ile belirtilen veya gösterilen hususlara, d) T [...]

Sample 2

id: law_tr_Sosyal-Sigortalar-ve-Genel-Sa-l-k-Sigortas-Kanun
source: law_tr
token_count: 840
author: Sosyal Sigortalar ve Genel Sağlık Sigortası Kanunu

necek ilaçlar dikkate alınmaz. Kurum, birinci fıkranın (a) bendi gereği belirlediği katılım payını; birinci basamak sağlık hizmeti sunucularında yapılan muayenelerde almamaya ya da daha düşük tutarlarda belirlemeye veya tekrar birinci fıkranın (a) bendi için belirlenen tutara getirmeye, ikinci ve üçüncü basamak sağlık hizmet sun [...]

Sample 3

id: law_tr_Katma-De-er-Vergisi-Kanunu-gecici-madde-22-chunk
source: law_tr
token_count: 662
author: Katma Değer Vergisi Kanunu

112 16/2/2016 tarihli ve 6676 sayılı Kanunun 6 ncı maddesiyle, bu fıkrada yer alan "teknoloji geliştirme bölgesinde" ibaresinden sonra gelmek üzere "ve ihtisas teknoloji geliştirme bölgesinde" ibaresi eklenmiştir. [...]

Filtering

From upstream corpus.csv rows (cleaned article text in context): whitespace normalization -> labelled-only PII redaction (e-mail, labelled phone, labelled T.C. kimlik no/TCKN/VKN, TR-IBAN candidates: 0 matches across the corpus - statutes carry no personal data) -> langid == tr gate -> cross-document boilerplate line filter -> exact + near dedup (sha256 of normalized text; Jaccard >= 0.90 on capped 25 x 5-word shingle sets).

Rejections (sum exactly to 1,862; one decision per row in artifacts/decisions.jsonl): 1,703 too short (min 200 chars) / 144 low letter-ratio / 8 duplicate context_key rows / 6 non-Turkish / 1 near-duplicate.

Known limitations

• Article-level chunks as delivered upstream (statute_full / chunk strategies); consolidated amendment notes are retained as substantive legal content.
• created is empty: statute consolidation dates are not in the source rows.
• Table-like chunks kept (flagged per record in artifacts/attribution.jsonl, is_table_like).
• No cross-source dedup against other corpus shards.

See artifacts/ for attribution, per-record decisions, QA counts, samples and checksums. verify() passes (14,139 texts reconstructed from the extracted source, token counts recomputed); tests 5/5.