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 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.

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	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, normalised; see Orthography — emphatic ṛ is unmarked

analysis

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

Same 336,151 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 (33,065 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). The source text carries 0.00% homoglyph error and 0.00% legacy-font mojibake.

🔴 The emphatic ṛ is not represented, and this is a property of the source. Measured over all 270,026 training forms, ṛ U+1E5B occurs 0 times, while ḍ appears in 14.43% of forms, ḥ in 12.14%, ṣ in 3.36%, ẓ in 2.73% and ṭ in 4.37%. The upstream conjugation resource writes plain r throughout — its lemmas are ruḥ, ṣber and ɛreḍ, not ṛuḥ, ṣbeṛ and ɛṛeḍ.

The normaliser did not remove it; there was never anything to preserve. The ṛ/r contrast is therefore neutralised in this dataset, and it is not neutralised in agbalu/KabLex, which carries both ṛuḥ and ruḥ. A model trained here will not produce emphatic ṛ, and scoring its output against text that marks it will count every such form wrong.

Restoring it would mean deciding, for each r, whether the verb takes the emphatic — a lexical question this dataset has no evidence for. It is disclosed rather than guessed.

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	exact match	CER
inflection	copy the lemma unchanged	4.07%	30.73%

make bench TASK=inflect

The floor is not zero, and that is why it is measured rather than assumed. Kabyle's imperative singular is the citation form for most verbs, so a copy is correct in one cell of nearly every paradigm — 4.07% of 33,065 test forms. A headline exact match read against zero would be flattered by exactly that much.

No system has yet been scored on the analysis direction.

Because the splits are sealed by lemma, a model cannot reach the floor by memorising paradigms: no test verb appears in any training cell.

Known Limits

• No emphatic ṛ. See Orthography above — 0 occurrences in 270,026 forms, because the source resource does not mark it. This is the most consequential limit here.
• Deduplicated. The raw shards repeat rows verbatim: one verb's intensive stems can generate the same cell twice. 8,594 identical rows were dropped (344,745 → 336,151), so no form is silently reweighted during training. Earlier revisions of this dataset carried them.
• 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.