README.md

---
license: cc-by-4.0
language:
  - kab
language_bcp47:
  - kab-Latn
size_categories:
  - 100K<n<1M
task_categories:
  - token-classification
task_ids:
  - lemmatization
  - part-of-speech
pretty_name: KabLex — Kabyle lexicon and pronunciation dictionary
tags:
  - kabyle
  - taqbaylit
  - berber
  - tamazight
  - low-resource
  - lexicon
  - morphology
  - grapheme-to-phoneme
configs:
  - config_name: lexicon
    data_files:
      - split: train
        path: lexicon/train.jsonl
  - config_name: pronunciations
    data_files:
      - split: train
        path: pronunciations/train.jsonl
---

# KabLex

A lexical layer for Kabyle (Taqbaylit, `kab`, Latin script), from the
[AƔBALU](https://huggingface.co/agbalu) 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.

```python
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 pronunciation, aligned from 59,185 sentence-level pairs
(99.53% alignment rate) over a 41-phoneme inventory.

| field | |
|---|---|
| `word` | the orthographic word, normalised |
| `ipa` | its most frequent pronunciation |
| `variants` | list of `{ipa, count}` — every attested pronunciation with its frequency |

**Ambiguity is 0.0%**: no word in this set has two pronunciations attested at comparable
frequency. That is a property of the source, which is single-annotator, and not evidence
that Kabyle has no heterophonic homographs.

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

## 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

- **Not a balanced dictionary.** See the POS distribution above.
- **`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** — `android`, `amikruskop`
  ("microscope"), `armstrong` — because it is aligned from running text and those occur in it.
  Three entries are genuinely non-lexical (`3d`, `mp3`, `aah`), 0.012% of the config.
  🚫 **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.
- Two upstream defects are known and *not* silently patched here: 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

```bibtex
@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.