PiotrSty/tatoeba-turkish

Tatoeba Turkish sentences (tatoeba_tr) - community example sentences

Turkish sentences from the weekly Tatoeba sentences_detailed export (downloaded 2026-10-05, ISO 639-3 tur rows). Sentence-level corpus: single community-contributed example sentences with per-record author attribution (username + dates).

• Source rows (tur): 749,368 of 13,597,683 export rows
• Retained records: 627,603 (83.8%)
• Measured cl100k_base proxy tokens: 9,593,737
• Characters: 23,534,897
• Time span (date added): 2009-11-29 - 2026-10-03
• Source snapshot: sentences_detailed.tar.bz2 (sha256 0af302eebda1d7c5773c505b7b6fa5bb41690eb8bcb31cf213a07caf04a88ba3, weekly export current through 2026-10-11; Tatoeba exports refresh Saturdays 06:30 UTC)

License - evidence at the origin

• Verbatim Tatoeba downloads page (https://tatoeba.org/en/downloads): "These files are released under CC BY 2.0 FR. A part of our sentences are also available under CC0 1.0."
• Per-record license field: CC-BY-2.0-FR; per-record attribution in artifacts/attribution.jsonl.gz: sentence id, contributing username, date added, date modified, source URL (https://tatoeba.org/en/sentences/view/<id>) and text_sha256. Contributor usernames are preserved as attribution.

Samples

Verbatim records from artifacts/sample.jsonl.

Sample 1

id: tatoeba_tr_2110912
source: tatoeba_tr
created: 2012-12-29
token_count: 8

Bizimle gelebilirsin.

Sample 2

id: tatoeba_tr_3425633
source: tatoeba_tr
created: 2014-08-13
token_count: 14

İngilizce öğrenmek ister misin?

Sample 3

id: tatoeba_tr_4038857
source: tatoeba_tr
created: 2015-04-05
token_count: 13

Seni ikna etmeye çalışmayacağım.

Filtering

Sentence-level record model (one record per sentence). Rejections (121,765 records, one final decision per record in artifacts/decisions.jsonl.gz):

{
  "low_letter_ratio": 14,
  "non_turkish_text": 38930,
  "too_little_text": 82821
}

Minimum final text is 20 characters (short fragments and non-sentence rows drop); language gate langid == tr after the upstream ISO 639-3 tur filter (38,930 rows re-classified as non-Turkish - short or ambiguous sentences). PII redaction (labelled patterns: e-mail, labelled phone, labelled IP, labelled national identifier, IBAN): 2 e-mail matches, 0 residual documents across all patterns. Exact dedup on 128-bit truncated sha256; near dedup (Jaccard >= 0.90 on capped 25-shingle sets) is inert for sentence-length records and documented as such. No boilerplate lines exceeded the 1% document-frequency threshold (max bin df 3).

Known limitations

• Sentence-level corpus: single example sentences, often translationese; not documents or conversations.
• Community-contributed content of uneven quality and mixed register.
• Contributor usernames remain as attribution (per the license).
• Near-dedup does not fire on sentence-length rows.

See artifacts/ for attribution, per-record decisions, QA counts, samples and checksums. verify() passes: 627,603 rows, unique IDs, token counts recomputed and text_sha256 re-checked against the emitted text. Tests: 10 (pytest).