KabInflect

A morphological inflection and analysis benchmark for Kabyle (Taqbaylit, kab, Latin script), from the AƔBALU project.

344,745 inflected verb form entries across 13,226 unique verb lemmas, partitioned into paradigmatically sealed splits (0 paradigm leakage), plus 6,198 complete verb conjugation tables.

from datasets import load_dataset

inflect = load_dataset("agbalu/KabInflect", "inflection")
analysis = load_dataset("agbalu/KabInflect", "analysis")
paradigms = load_dataset("agbalu/KabInflect", "paradigms", split="train")

inflection

Morphological Inflection (Seq2Seq Morphological Generation): (lemma, feats) → form.

344,745 total entries across 13,226 verb lemmas.

split	lemmas	inflected forms
train	10,580	276,717
dev	1,323	34,109
test	1,323	33,919
total	13,226	344,745

Schema

field	type	description
id	string	kab_inflect_{split}_{idx}
lemma	string	Base verb infinitif (canonical)
feats	string	Universal Dependencies (UD) FEATS (e.g. Aspect=Perf|Gender=Masc|Number=Sing|Person=3)
tense_raw	string	Legacy source tense descriptor (e.g. prétérit, aoriste intensif)
person_raw	string	Legacy source person descriptor (e.g. 3s_m, 1p)
form	string	Inflected surface form in canonical Kabyle orthography

analysis

Morphological Analysis and Lemmatisation: form → (lemma, feats).

Same 344,745 entries, indexed by inflected surface form.

paradigms

6,198 full verb paradigms with French translations, irregularity indicators (is_irregular, is_derived), pattern verbs, and principal aspect stems (imperative, aorist, preterite, negative_preterite, aorist_participle, preterite_participle, negative_preterite_participle, intensive_forms).

Paradigmatically Sealed Partitioning (Zero Paradigm Leakage)

Upstream datasets randomly split individual inflected forms across train and test. That design flaw causes extreme data leakage: forms of the exact same verb (e.g. awḍeɣ, tewḍeḍ, yeweḍ) appear in both splits, so a test score measures simple paradigm memorisation rather than linguistic learning.

agbalu/KabInflect fixes this:

• Entries are grouped strictly by verb lemma (infinitif).
• Splitting is performed at the lemma level with seed 42.
• 0% lemma overlap: all 1,323 verbs in test (33,919 forms) are completely unseen during training. A system must learn the underlying rules of Kabyle morphology to succeed.

Orthography

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

Evaluation Protocol & Baseline Floor

The benchmark evaluates morphological generation (inflection) and parsing (analysis) on 1,323 held-out test verb lemmas (33,919 unseen test forms).

Task	Baseline / System	Metric	Score	Notes
Inflection	Identity Lemma Floor	Exact Match (EM)	16.82%	Naive floor predicting form = lemma
Inflection	Identity Lemma Floor	Character Error Rate (CER)	34.21%	Levenshtein edit distance ratio
Analysis	Unsupervised Lexicon Lookup	EM Accuracy	78.41%	Exact match via agbalu/KabLex lookup

Because splits are sealed by verb lemma (0% paradigm overlap), models must learn generalisable Kabyle morphological rules to surpass the baseline floor on unseen paradigms.

Known Limits

• Verb-focused. Covers Kabyle verbal morphology (the largest morphological paradigm family in the language); nominal annexed-state paradigms are covered separately in agbalu/KabLex.
• Gloss language. Paradigm translations are French, reflecting the source dictionary.
• Standard orthography. Uses standard Kabyle Latin script; does not cover dialectal spellings or Tifinagh (see agbalu/KabTifinagh for script conversion).

Citation

@misc{agbalu_kabinflect,
  title  = {KabInflect: a Kabyle morphological inflection and analysis benchmark},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/agbalu/KabInflect}
}

Please also cite boffire/kabyle-verbs for the upstream data collection.

Licence

CC-BY-4.0. Free to share, modify, and build upon with proper attribution.