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 expressions.

Field	Type	Description
form	string	Normalized surface form
lemma	string	Base lemma (where available in source)
upos	string	Universal Dependencies POS tag
feats	string	Morphological features in UD format (_ indicates no features; non-null)
glosses	list[dict]	Translations as {"lang": "...", "text": "..."} (available for 9,658 entries)
source	string	Originating source registry ID
licence	string	Upstream source license

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

The part-of-speech distribution reflects the underlying source datasets: 94% of entries are inflected verb forms drawn from verb inflection tables. Users should treat this dataset primarily as an extensive verbal morphological resource with an accompanying nominal and toponymic layer.

Composition

Source	Entries	License
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

Licensing Filter

From the raw 395,834 assembled entries, 28,942 entries were excluded by automated licensing filters to ensure a strictly permissive release:

• 8,554 entries licensed under ODbL were excluded to prevent share-alike requirements from restricting downstream model development.
• 20,388 entries with unresolved licensing metadata were excluded.

The published dataset is distributed under CC-BY-4.0.

pronunciations

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

Field	Type	Description
word	string	Normalized orthographic word
ipa	string	Primary IPA pronunciation
variants	list[dict]	Frequency-ranked alternative readings: [{"ipa": "...", "count": N}, ...]
repaired	bool	Whether character dropping in the upstream generator was corrected

This dataset provides orthographic-to-phonetic transliterations rather than continuous acoustic transcriptions. Characteristics across the 292,921 aligned tokens include:

• Deterministic mapping: Each word spelling maps to a single canonical IPA representation (0.0% ambiguity).
• Schwa representation: Phonetic ə maps directly to orthographic e across 100.00% of entries; this dictionary does not model context-sensitive epenthetic schwa insertion (for context-dependent schwa restoration from consonant skeletons, refer to agbalu/Juba-27M).
• Gemination: Length markers (ː) correspond directly to doubled consonant graphemes in 100.00% of entries.
• Word-isolated forms: Sandhi, cross-word spirantization, and assimilation across word boundaries are outside the scope of single-word dictionary entries.

This resource is intended as a lexical pronunciation dictionary and G2P training target.

Repaired Characters

The upstream generation pipeline dropped characters lacking explicit conversion rules. This release repairs 199 entries, flagged by repaired: true:

Character	Words	Original	Repaired
o	128	bob → ββ	βoβ
ţ (U+0163)	73	aţan → ææn	ætːæn
é, ï	3	rosé → rs	Unchanged (loanword foreign letters)

The letter ţ (U+0163) represents Dallet-tradition orthography for the spirantised / tense t, attested 21,058 times in the AƔBALU text corpus. In 43% of word types carrying ţ, a corresponding tt form exists, while 34% have a t form (compared to 0.7% for ṭ). Based on this distribution, missing pronunciations for ţ were mapped to the geminate stop ætːæn.

Eight foreign loanwords or malformed entries (3d, mp3, androïd, rosé, supermarché, muḥ€nd, xelleṣ̣, ṭeyyeb‟) containing digits or stray punctuation were preserved verbatim without speculative correction. Repairs were restricted strictly to entries where source rules truncated character strings.

Source: boffire/kabyle-g2p-training-data, CC0-1.0.

Usage

from datasets import load_dataset

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

Filtering by license tier:

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

Orthography

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

Text normalization standardizes character encoding across the dataset. In raw public Kabyle corpora, between 2.60% and 3.19% of rows exhibit Greek ε (U+03B5) substitutions for Latin ɛ (U+025B), which fragments tokenizer subwords if uncorrected. The normalization pipeline preserves all standard Berber Latin letters (ɣ, ɛ, ḥ, ḍ, ṣ, ṭ, ẓ, ṛ, č, ǧ, ţ).

Known Limitations

• Source imbalance: The lexicon is predominantly composed of verbal inflections rather than a balanced distribution of parts of speech.
• UPOS coverage: Part-of-speech tags are missing for 15,357 entries and reflect upstream source annotations where present.
• Loanwords in pronunciation data: Includes common loanwords and foreign names occurring in source text (amikruskop, armstrong, android).
• Glosses: Lexical definitions are primarily in French, reflecting original lexicographical sources.

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. Published entries derive from CC-BY-4.0 and MIT sources. The pronunciation configuration is CC0-1.0 upstream. Each row includes source licensing metadata in the licence column.