KabInflect

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

336,151 inflected verb form entries across 13,226 unique verb lemmas, partitioned into lemma-disjoint 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.

336,151 total entries across 13,226 verb lemmas.

Split	Lemmas	Inflected forms
train	10,580	270,026
dev	1,323	33,060
test	1,323	33,065
Total	13,226	336,151

Schema

Field	Type	Description
id	string	Unique record ID (kab_inflect_{split}_{idx})
lemma	string	Base verb citation form (infinitif)
feats	string	Universal Dependencies morphological features (e.g. Aspect=Perf|Gender=Masc|Number=Sing|Person=3)
tense_raw	string	Source tense descriptor (e.g. prétérit, aoriste intensif)
person_raw	string	Source person descriptor (e.g. 3s_m, 1p)
form	string	Normalized inflected surface form

analysis

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

336,151 entries indexed by inflected surface form, using the same lemma-disjoint split boundaries.

paradigms

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

Usage

from datasets import load_dataset

inflection = load_dataset("agbalu/KabInflect", "inflection", split="train")
print(inflection[0])
# {'id': 'kab_inflect_train_000000', 'lemma': 'ttemnunnuḍ',
#  'feats': 'Aspect=Imp|VerbForm=Part', 'tense_raw': 'participe aoriste intensif',
#  'person_raw': 'participe', 'form': 'ittemnunnuḍen'}

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

The inflection task maps (lemma, feats) → form, while morphological analysis performs the reverse mapping form → (lemma, feats).

Lemma-Disjoint Partitioning

To prevent data leakage, dataset splits are partitioned strictly at the verb lemma level rather than splitting individual surface forms randomly:

• Entries are grouped by base verb lemma (infinitif).
• Splitting is performed across lemmas using seed 42.
• 0% lemma overlap: All 1,323 verb lemmas in test (33,065 forms) are completely unseen in the training split, requiring models to generalize morphological inflection rules to novel lexemes rather than memorizing attested paradigms.

Orthography

All Kabyle text is normalized under the AƔBALU reference normaliser (1.3.0+rules1.0.0, 81 rules).

In the underlying source verb resource, the emphatic ṛ (U+1E5B) is not distinguished from plain r (occurring 0 times across 270,026 training forms, whereas other emphatics like ḍ, ḥ, ṣ, ṭ, ẓ are fully marked). Consequently, the r vs. ṛ phonemic distinction is neutralized in this dataset. Models trained exclusively on KabInflect will not generate emphatic ṛ; when evaluating on corpora where ṛ is marked, references should be normalized accordingly to prevent spurious orthographic mismatches.

Evaluation Protocol & Baseline Floor

The benchmark evaluates morphological generation (inflection) and parsing (analysis) on 1,323 held-out test verb lemmas (33,065 unseen test forms).

Task	System / Baseline	Exact match	CER
Inflection	Copy lemma unchanged	4.07%	30.73%

Because the imperative singular serves as the citation form for many Kabyle verbs, a trivial baseline copying the lemma unchanged achieves 4.07% exact match (and 30.73% CER) across the test split.

Known Limitations

• Unmarked ṛ: Emphatic ṛ is written as plain r throughout the upstream verbal database, neutralizing this contrast.
• Deduplication: 8,594 redundant duplicate entries present in raw source tables were removed during curation (reducing 344,745 to 336,151 rows).
• Verbal domain: Focuses on Kabyle verbal morphology; nominal state alternations are provided in agbalu/KabLex.
• Gloss language: Translations and definitions within paradigm tables are in French.

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.