boffire/nllb_en_kab

NLLB English-Kabyle Parallel Corpus (Filtered & Cleaned)

Parallel English–Kabyle sentence pairs derived from the OPUS-NLLB corpus, filtered with GlotLid v3 and cleaned through a multi-stage Kabyle-specific pipeline.

Dataset Structure

• nllb_en_kab.parquet: Parquet file with two columns:
  • english: English sentence
  • kabyle: Kabyle sentence

Statistics

Metric	Count
Total sentence pairs	2,786,012
Non-null English	2,786,012
Non-null Kabyle	2,786,012

Pipeline

1. Source: OPUS-NLLB v1 (Moses format)
2. Language filtering: GlotLid v3 (__label__kab / __label__kab_Latn, threshold ≥ 0.95) — reduced 4,123,481 → 2,809,651 pairs
3. Deduplication: Exact duplicate (en, kab) pairs removed — 221 duplicates dropped
4. Forbidden-character filter: Multi-script blocklist (CJK, Arabic, Devanagari, etc.) — 23,418 pairs removed
5. Character fixing: Normalized mis-encoded / lookalike glyphs:
   • Greek/Cyrillic lookalikes: ε → ɛ, γ → ɣ, Γ → Ɣ, Σ → Ɛ, Ԑ → Ɛ
   • Romanian cedilla/comma variants: ţ → ṭ, Ț → Ṭ, ţţ → tt, etc.
   • Encoding fixes: î → ṭ, â → ɛ, ê → ḥ, ô → ṛ, ř → l, ğ → ǧ, ş → ṣ, ḏ → ḍ, ṯ → ṭ, ɤ → ɣ
   • Phrase-level: aţ- ţ → ad t
6. Validation: Checked against Kabyle CLDR alphabet + loan letters (o, O, v, V); final non-conformant rate: 1.78% (French diacritics in loanwords)

Usage

from datasets import load_dataset

dataset = load_dataset("boffire/nllb_en_kab", split="train")
print(dataset)

License

CC-BY-NC 4.0

Citation

@dataset{nllb_en_kab,
  title = {NLLB English-Kabyle Parallel Corpus},
  author = {Athmane MOKRAOUI},
  year = {2026},
  url = {https://huggingface.co/datasets/boffire/nllb_en_kab}
}