KabTifinagh

A standardized bidirectional script transliteration and schwa (e) vowel restoration benchmark for Kabyle (Taqbaylit, kab, Latin & Tifinagh scripts), created by the AƔBALU project.

KabTifinagh normalises, repairs, deduplicates, and structures 497,944 parallel sentence entries matching Neo-Tifinagh (kab_Tfng) to canonical Kabyle Latin (kab_Latn), alongside 123,852 English and 205,637 French trilingual sentence alignments.

from datasets import load_dataset

script = load_dataset("agbalu/KabTifinagh", "script_conversion")
trilingual_en = load_dataset("agbalu/KabTifinagh", "trilingual_en")
trilingual_fr = load_dataset("agbalu/KabTifinagh", "trilingual_fr")

script_conversion

Bidirectional Transliteration and Vowel Restoration: text_tfng ↔ text_latn.

497,944 total parallel sentences, partitioned with seed 42 into 0-leakage splits.

split	parallel sentences
train	398,355
dev	49,794
test	49,795
total	497,944

Schema

field	type	description
id	string	Unique record ID (kab_tfng_{split}_{idx})
text_latn	string	Canonical Kabyle Latin sentence (100% normalized)
text_tfng	string	Neo-Tifinagh Kabyle sentence

trilingual_en & trilingual_fr

Parallel Kabyle Latin, Kabyle Tifinagh, and English / French translations for cross-lingual multimodal and multi-script AI research.

• trilingual_en: 123,852 sentence triples (id, text_latn, text_tfng, text_en).
• trilingual_fr: 205,637 sentence triples (id, text_latn, text_tfng, text_fr).

Trilingual Schema

field	type	description
id	string	Unique record ID (kab_tfng_en_{split}_{idx} / kab_tfng_fr_{split}_{idx})
text_latn	string	Canonical Kabyle Latin sentence
text_tfng	string	Neo-Tifinagh Kabyle sentence
text_en / text_fr	string	English or French translation

The Vowel Restoration & Script Conversion Challenge

Tifinagh orthography presents a key sequence-to-sequence learning task due to structural orthographic asymmetries between the two scripts:

1. Schwa (e) Omission: Standard Tifinagh orthography omits schwa vowels (e).
   • Latin: Tecfiḍ fell-i ? → Tifinagh: ⵜⵛⴼⵉⴹ ⴼⵍⵍⵉ ?
   • While forward mapping (Latin → Tifinagh) is straightforward, reverse transliteration (Tifinagh → Latin) requires predicting where omitted schwas belong based on Kabyle phonotactics and context.
2. Morphological Boundaries: Clitic hyphens (fell-i, yefka-yas) are represented contiguously in Tifinagh (ⴼⵍⵍⵉ).
3. Vowel & Consonant Mappings: Standard Tifinagh codepoints map o/u to ⵓ and b/p to ⴱ.

A deterministic rule-based character mapper provides a baseline floor, but sequence-to-sequence models (character-level transformers or fine-tuned LLMs) trained on agbalu/KabTifinagh learn contextual phonotactic rules to reconstruct missing schwas (e) and clitic boundaries with high fidelity.

Evaluation Protocol & Baseline Floor

Evaluated on the 49,795 held-out test split sentences with zero sentence leakage:

Task Direction	Baseline Method	Metric	Score	Description
Latin → Tifinagh (Forward)	Deterministic Character Lookup	Exact Match (EM)	1.02%	Standard rule-based conversion
Latin → Tifinagh (Forward)	Deterministic Character Lookup	Character Error Rate (CER)	14.83%	Levenshtein edit distance ratio
Tifinagh → Latin (Reverse)	Deterministic Character Lookup	Exact Match (EM)	0.00%	Deterministic baseline floor
Tifinagh → Latin (Reverse)	Deterministic Character Lookup	CER	17.42%	Edit distance ratio
Tifinagh → Latin (Reverse)	Deterministic Character Lookup	Schwa e Recall	1.23%	Baseline schwa reconstruction rate
Tifinagh → Latin (Reverse)	Deterministic Character Lookup	Schwa e F1	2.40%	Harmonic mean of precision & recall

Orthography & Engineering

All Kabyle Latin text is normalised under AƔBALU normaliser 1.3.0+rules1.0.0 (81 rules), preserving sub-dot emphatic distinctions (ɣ ḍ ḥ ṭ ṛ ẓ ṣ). The text carries 0.00% homoglyph error (all Greek ε homoglyphs repaired).

Known Limits

• Script Standard. Uses IRCAM Neo-Tifinagh codepoints; does not cover historical Tuareg Tifinagh variations.
• Prose Domain. Drawn primarily from general prose and Tatoeba parallel sentences.

Citation

@misc{agbalu_kabtifinagh,
  title  = {KabTifinagh: a Kabyle script transliteration and vowel restoration benchmark},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/agbalu/KabTifinagh}
}

Upstream data sourced from abdelhaqueidali/kab-latn-tfng.

Licence

CC-BY-2.0. Free to share, modify, and build upon with proper attribution.