KabLex

A lexical layer for Kabyle (Taqbaylit, kab, Latin script), from the AƔBALU project.

366,892 lexical entries merged from three permissively licensed sources and normalised to one orthography, plus 25,642 word–pronunciation pairs aligned from sentence-level grapheme-to-phoneme data.

from datasets import load_dataset

lex = load_dataset("agbalu/KabLex", "lexicon", split="train")
ipa = load_dataset("agbalu/KabLex", "pronunciations", split="train")

lexicon

366,892 entries — 234,936 distinct surface forms over 17,090 lemmas, 2,408 of them multiword.

field
form	the surface form, normalised
lemma	its lemma, where the source gives one
upos	Universal POS tag, where the source gives one
feats	morphological or domain features, UD-style. "_" is UD's placeholder for "no features", on 25,858 entries; the other 341,034 carry a real value, and the column is never null
glosses	list of {lang, text} — 9,658 entries carry at least one
source	the registry id this entry came from
licence	that source's licence

POS	entries
VERB	345,057
unlabelled	15,357
NOUN	5,603
PROPN	287
PART	249
ADJ, ADV, INTJ, others	246

The distribution is a property of the sources, not of Kabyle. 94% of the entries are verb forms because the largest contributing source is an inflected verb list. Treat this as a wide verbal paradigm resource with a thin nominal layer, not as a balanced dictionary.

Composition

source	entries	licence
hf.boffire.kabyle-verbs	342,257	CC-BY-4.0
hf.boffire.hunspell-kab	20,967	MIT
hf.agurzil.tafsut-maths-lexicon	3,668	CC-BY-4.0

What was excluded, and why

The built lexicon holds 395,834 entries. 28,942 are not published here, and the cut is made by code from each source's licence rather than by hand:

• 8,554 entries from an ODbL toponym source. ODbL §4.4 obliges a publicly used derivative database to be offered under ODbL too, so including them would impose share-alike on this whole release. ODbL is share-alike, not permissive — a distinction that is easy to get wrong and that changes what a downstream user is allowed to do.
• 20,388 entries whose upstream licence could not be established. An unresolved licence is not a permissive one. That source is independently available on the Hub.

Anyone who wants the full set can rebuild it from the registry; the exclusions are mechanical and reproducible, not editorial.

pronunciations

25,642 distinct words with an IPA reading, aligned from 59,185 sentence-level pairs (99.53% alignment rate) over a 42-phoneme inventory.

field
word	the orthographic word, normalised
ipa	its reading
variants	list of {ipa, count} — every attested reading with its frequency
repaired	true where this project restored a character the source dropped (§ below)

🔴 This is a transliteration, not a phonetic transcription, and the distinction is load-bearing. Measured over all 292,921 aligned tokens:

• No word ever takes two readings — ambiguity is 0.0% — and that holds even when conditioned on the preceding word. The reading is a deterministic function of the spelling, which is a property of the generator, not evidence that Kabyle has no heterophonic homographs.
• count(ə) equals count(<e>) in 100.00% of entries. The schwa is orthographic e relabelled. This data does not restore schwa, and Kabyle schwa epenthesis is context-sensitive. For that, use agbalu/Juba-27M, which scores 93.51% positional schwa F1.
• ː corresponds to a doubled letter in 100.00% of the 10,465 entries carrying it.
• There is no cross-word phonology. Kabyle spirantization and assimilation cross word boundaries; a per-word table cannot express that.

Use it as a pronunciation dictionary and as a G2P training target. Do not use it as evidence about Kabyle phonetics.

Repaired characters

🔴 The upstream generator emits nothing for a character it has no rule for, silently. This release repairs 199 entries, flagged by the repaired field:

character	words	before	after
o	128	bob → ββ	βoβ
ţ U+0163	73	aţan → ææn	ætːæn
é, ï	3	rosé → rs	left as-is, see below

ţ is real Kabyle, not a foreign letter. It is the Dallet-tradition notation for the spirantised/tense t, attested 21,058 times in this project's corpus; of 1,407 word types carrying it, 43% have a tt counterpart and 34% a t one, against 0.7% for ṭ. It is repaired to the geminate on that evidence. 🟡 No audio confirms this mapping — ţ occurs zero times in Common Voice Kabyle.

Eight entries are left exactly as the source had them: 3d, mp3, androïd, rosé, supermarché, muḥ€nd, xelleṣ̣, ṭeyyeb‟. Each carries a digit, a currency sign, a stray combining mark or a curly quote — outside the writing system, so repairing them would mean guessing.

Repairs are made only where the source reading is shorter than the reference table's. Sound entries are untouched, because the table does not model i/u laxing and regenerating one would replace an attested allophone with an approximation.

Source: boffire/kabyle-g2p-training-data, CC0-1.0. The repair is this project's.

Usage

from datasets import load_dataset

lexicon = load_dataset("agbalu/KabLex", "lexicon", split="train")
lexicon[0]
# {'form': 'asider', 'lemma': None, 'upos': None, 'feats': 'Domain=Math',
#  'glosses': [{'lang': 'fra', 'text': 'abaissement'}],
#  'source': 'hf.agurzil.tafsut-maths-lexicon', 'licence': 'cc-by-4.0'}

pronunciations = load_dataset("agbalu/KabLex", "pronunciations", split="train")
pronunciations[0]
# {'word': 'a', 'ipa': 'æ', ...}

licence is a column because the tiers are not uniform. Cut on it before redistributing:

permissive = lexicon.filter(lambda row: row["licence"].startswith("cc-by-4"))

_ in feats is UD's own placeholder for "no features", and None in lemma or upos means the source supplied none. Neither is missing data to be cleaned.

Orthography

Everything is normalised to one documented orthography — normaliser 1.3.0+rules1.0.0, 81 rules, zero idempotence violations over 931,342 sentences.

This matters more for Kabyle than the phrase usually suggests. Two seed corpora in this project carry 2.60% and 3.19% homoglyph-corrupted rows — Greek ε U+03B5 written for Latin ɛ U+025B — and published Kabyle tokenizer vocabularies have that corruption baked into their merges, at a measured +17.8–21.3% token cost.

Kabyle legitimately uses ɣ ɛ ḥ ḍ ṣ ṭ ẓ ṛ č ǧ ţ, and the emphatic and spirantised distinctions carry meaning. Do not "clean" them away.

Known limits

• It is a union, not a balanced dictionary. The POS distribution above is what the three sources contain, not a designed sample. Read it before drawing a rate from this file.
• upos is absent for 15,357 entries and is the source's tag where present, not a reviewed annotation.
• The pronunciation set covers loanwords and proper nouns — amikruskop ("microscope"), armstrong, android — because it is aligned from running text and those occur in it. These are also where the source's dropped-character defect concentrated, so most of the 199 repaired entries are loanwords. Do not filter this config by word length or character inventory. The shortest entries are the most frequent words in the language: ad occurs 12,024 times, d 18,862, i 11,141, n 6,001.
• One upstream defect is patched here and flagged: the G2P source deletes characters it has no rule for, and the repaired field marks every entry this release restored.
• Two upstream defects are known and not patched: one source maps ţ→ṭ against corpus evidence, and another tags i as isem. Both are owed as upstream reports.
• Glosses are mostly French, reflecting the sources.
• Nothing here is human-reviewed at scale. It is a merged, normalised, provenance-carrying union of existing resources — which is exactly what did not exist before, and is not the same thing as a curated dictionary.

Citation

@misc{agbalu_kablex,
  title  = {KabLex: a Kabyle lexicon and pronunciation dictionary},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/agbalu/KabLex}
}

Please also cite the upstream sources listed under Composition.

Licence

CC-BY-4.0. Every published entry is CC-BY-4.0 or MIT; CC-BY-4.0 is the more restrictive of the two and governs the release. The pronunciation config is CC0-1.0 upstream. Each row carries its own licence field, so a stricter subset can be cut by code.