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 by a deterministic seeded probabilistic corruption pass that simulates the keyboard strategies Kabyle speakers use on phones and social media.

Used to train agbalu/Boulifa-48M, which reaches 99.45% character accuracy on the held-out test split.

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 at 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 (normalised, unmodified)

Construction

Source sentences are the text_latn column of agbalu/KabTifinagh (all three splits combined), normalised under AƔBALU normaliser 1.3.0+rules1.0.0. Each sentence generates exactly one pair at seed 42 — the dataset is fully reproducible from the source corpus alone.

Identity pairs (15%). IDENTITY_RATE = 0.15. One in seven sentences is left unchanged (source == target), teaching any model trained on this data not to edit already-canonical text.

Corrupted pairs (85%). The remaining 85% are passed through a probabilistic corruption pass that applies the following transformations stochastically and 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 the same probability.

Clitic hyphen omission (PROB_CLITIC_DROP = 0.50)

If the sentence contains -, with probability 0.50: replace all hyphens with a space (d-yeffeɣ → d yeffegh) or delete them (d-yeffeɣ → dyeffegh), each with probability 0.50.

Preposition contraction (PROB_PREP_SHORTEN = 0.25)

deg → g , seg → s (word-boundary anchored), with probability 0.25.

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?"

The third row is an identity pair (source == target).

Evaluation

Scored on the held-out test split by agbalu.bench.standardise (not yet published), under greedy free-running character accuracy.

system	character accuracy	character error rate
Boulifa-48M	99.45%	0.55%
deterministic rule table	< 2%	> 98%

make test-boulifa   # unit tests for the standardise module

The evaluation pairs are synthetic. The 99.45% figure is measured on the round-trip — can the model recover the canonical target from a plausibly corrupted source? It cannot be read as accuracy on arbitrary human typing, only on the corruption distribution defined here.

Known Limits

• Synthetic only. Every source string was generated by a rule. No human typed any of these inputs. The distribution approximates real typing but is not a sample of it.
• One variant per sentence. Each canonical sentence generates exactly one corrupted source. A model has not seen the same sentence under multiple corruption strategies.
• No adequacy judgement. The target strings are the normaliser's output. No human verification of the canonical form of any source sentence exists.
• Sibling language contamination. The source sentences come from agbalu/KabTifinagh, which carries the same contamination bound from its upstream sources: LID systems cannot reliably distinguish Kabyle from Tarifit, Central Atlas Tamazight or Shawiya.

Reproduction

make prepare-boulifa   # generates train/dev/test.jsonl on Modal and commits to the volume

The dataset is regenerated deterministically at seed 42 from agbalu/KabTifinagh. No GPU required.

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); a permissive grant on this derived dataset does not relicense the upstream corpus. Read agbalu/KabTifinagh's licence before redistributing derivatives of the training corpus.

Part of AƔBALU, a Kabyle corpus and model collection.