KabStandard

A 497,944-pair parallel dataset for Kabyle orthography standardisation — mapping informal, French-keyboard, and Arabizi Kabyle text to canonical Kabyle Latin orthography. Derived from the Latin side of agbalu/KabTifinagh using a deterministic probabilistic corruption engine that simulates keyboard conventions used on mobile devices and social platforms.

Used to train agbalu/Boulifa-48M, which reaches 97.39% character accuracy and 85.70% sentence exact match on held-out test pairs, compared to an 89.70% character accuracy baseline for copying the input unchanged.

from datasets import load_dataset

ds = load_dataset("agbalu/KabStandard")
# DatasetDict({'train': Dataset(448149), 'dev': Dataset(24897), 'test': Dataset(24898)})

Splits

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

Split	Pairs
train	448,149
dev	24,897
test	24,898
Total	497,944

Schema

Field	Type	Description
source	string	Informal input (French-keyboard, Arabizi, or identity)
target	string	Canonical Kabyle Latin (normalized)

Usage

from datasets import load_dataset

train = load_dataset("agbalu/KabStandard", split="train")
print(train[0])
# {'source': 'Ssubbetd negh ad naligh.', 'target': 'Ṣṣubbet-d neɣ ad n-aliɣ.'}

The task maps source → target. Because corruption is applied probabilistically at the character level, a baseline copying the input unchanged retains 89.70% character accuracy. Evaluation should compare model performance against this baseline floor.

Construction

Source sentences are drawn from the text_latn column of agbalu/KabTifinagh (all splits combined) and normalized under the AƔBALU reference normaliser (1.3.0+rules1.0.0). Each sentence generates exactly one pair at seed 42, making the dataset fully reproducible.

• Identity pairs (15%): IDENTITY_RATE = 0.15. 15% of sentences are preserved unchanged (source == target), training models to avoid modifying already canonical orthography.
• Corrupted pairs (85%): The remaining 85% receive stochastic transformations applied independently per character:

Phoneme substitutions (PROB_SUBSTITUTION = 0.90)

Canonical	Informal Variants	Probabilities
ɣ / Ɣ	gh / g / 3 / 8	0.75 / 0.10 / 0.08 / 0.07
x / X	kh / k / 5	0.85 / 0.10 / 0.05
c / C	ch / c / sh	0.75 / 0.20 / 0.05
č / Č	tch / ch / tc	0.70 / 0.20 / 0.10
ğ / Ğ	dj / j / g	0.80 / 0.15 / 0.05
ḍ / Ḍ	dh / d	0.75 / 0.25
ṭ / Ṭ	th / t	0.70 / 0.30
ṣ / Ṣ	s / ss	0.75 / 0.25
ẓ / Ẓ	z / zz	0.80 / 0.20
ṛ / Ṛ	r / rr	0.90 / 0.10
ḥ / Ḥ	h / 7 / hh	0.70 / 0.25 / 0.05
ɛ / Ɛ	e / a / 3 / '	0.35 / 0.30 / 0.25 / 0.10

Vowel digraph (PROB_DIGRAPH_OU = 0.45)

u → ou (French convention for /u/) with probability 0.45; U → Ou with probability 0.45.

Clitic hyphen omission (PROB_CLITIC_DROP = 0.50)

For sentences containing -, hyphens are either replaced with spaces (d-yeffeɣ → d yeffegh) or removed (d-yeffeɣ → dyeffegh), each with probability 0.50.

Preposition contraction (PROB_PREP_SHORTEN = 0.25)

deg → g , seg → s with probability 0.25 at word boundaries.

Examples

source: "achimi ur d-thekhedmedh ara tamazight g l'ecole?"
target: "acimi ur d-tḥexedmeḍ ara tamaziɣt deg lɛecule?"

source: "3emmi l7adj yerza-d 5ir d lbaraka s wuzzal"
target: "Ɛemmi lḥadj yerza-d xir d lbaraka s wuzzal"

source: "Azul fell-awen, amek i telliḍ taṣebḥit-a?"
target: "Azul fell-awen, amek i telliḍ taṣebḥit-a?"

Evaluation

Evaluated across 1,000 held-out test pairs drawn with seed 4711 under greedy autoregressive decoding:

System	Character Accuracy	Character Error Rate	Exact Match
Boulifa-48M	97.39%	2.61%	85.70%
Copy input unchanged (baseline)	89.70%	10.30%	—

The baseline floor reflects that most characters in a typical corrupted sentence remain unchanged, resulting in an 89.70% character accuracy floor for an identity copy. Boulifa-48M removes approximately 75% of character-level corruption, achieving 97.39% accuracy.

Evaluation measures recovery of canonical orthography from synthetically corrupted text matching the defined probabilistic distribution; real-world informal typing exhibits additional idiosyncratic orthographic variance.

Known Limitations

• Synthetic generation: All source strings are synthetically generated from rules; while designed to reflect common keyboard practices, they do not constitute natural keystroke logs.
• Single corruption per sentence: Each canonical sentence generates one paired corrupted instance.
• Target normalization: Targets correspond to output from the reference normaliser.

Reproduction

The dataset is regenerated deterministically at seed 42 from agbalu/KabTifinagh using the standardized orthography rules and probabilistic corruption engine.

Citation

@misc{agbalu_kabstandard,
  title  = {KabStandard: a synthetic parallel corpus for Kabyle orthography standardisation},
  author = {AGBALU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/agbalu/KabStandard}
}

Derived from agbalu/KabTifinagh.

Licence

Apache-2.0. Derived from agbalu/KabTifinagh (CC-BY-2.0).

Part of AƔBALU, a Kabyle corpus and model collection.