boffire/tatoeba-en-kab

Tatoeba English-Kabyle Parallel Corpus

A cleaned and aligned English-Kabyle parallel corpus extracted from Tatoeba, with both direct en↔kab links and indirect kab→fra→en links.

Statistics

Split	Pairs
train	240,056
dev	2,449
test	2,449
Total	244,954

Source

• Tatoeba direct en↔kab links
• Indirect kab→fra→en links (Kabyle linked to French, French linked to English)

Cleaning Pipeline

1. Character standardization: Fixed common character encoding errors:

   • Greek epsilon ε (U+03B5) → Latin open e ɛ (U+025B)
   • Greek capital sigma Σ (U+03A3) → Latin capital open E Ɛ (U+0190)
   • Greek capital gamma Γ (U+0393) → Latin capital gamma Ɣ (U+0194)
   • Greek small gamma γ (U+03B3) → Latin small gamma ɣ (U+0263)
   • Cyrillic reversed ze Ԑ (U+0510) → Latin capital open E Ɛ
   • Romanian/Turkish t-cedilla ţ → Kabyle geminate t
   • Turkish g-breve ğ → Kabyle g-caron ǧ
   • Normalized no-break spaces, zero-width spaces, non-breaking hyphens

2. GlotLID v3 filtering: Language identification using GlotLID v3 (fasttext model model_v3.bin).

   • Retained only sentences predicted as kab_Latn with high confidence
   • Retention rate: 245,247 / 245,747 = 99.80%
   • Average confidence: 0.9990
   • Dropped: 543 sentences predicted as other languages (eng_Latn, tur_Latn, ita_Latn, deu_Latn, cat_Latn, shi_Latn, etc.)

3. Exact deduplication: Removed 4,440 duplicate (en, kab) pairs

4. Quality filtering:

   • Length filter: dropped pairs where either side < 3 chars or > 300 chars
   • Ratio filter: dropped pairs with length ratio > 5×
   • Removed empty, punctuation-only, and identical-sides pairs

Format

from datasets import load_dataset

ds = load_dataset("boffire/tatoeba-en-kab")

# Example:
# ds["train"][0] →
# {
#   "translation": {
#     "en": "When did you graduate?",
#     "kab": "Melmi i d-tewwimt igerdasen-nkent?"
#   }
# }

CLDR Kabyle Alphabet Compliance

The corpus is filtered to use only characters from the Unicode CLDR kab locale main exemplar set:

• Lowercase: a b c č d ḍ e ɛ f g ǧ ɣ h ḥ i j k l m n p q r ṛ s ṣ t ṭ u w x y z ẓ
• Uppercase: A B C Č D Ḍ E Ɛ F G Ǧ Ɣ H Ḥ I J K L M N P Q R Ṛ S Ṣ T Ṭ U W X Y Z Ẓ
• Auxiliary (loanwords): o v

License

CC0 1.0 Universal (Public Domain) — same as Tatoeba source data.

Citation

If you use this dataset, please cite Tatoeba:

@misc{tatoeba,
  title = {Tatoeba: Collection of sentences and translations},
  howpublished = {\url{https://tatoeba.org}},
}

Changelog

• 2026-07-08: Initial release (v1.0)
  • 244,954 aligned EN-KAB pairs
  • Extracted from Tatoeba weekly exports
  • Cleaned with GlotLID v3 filtering and character standardization