agbalu/Amrouche-1.3B
Translation • 1B • Updated
AƔBALU — Kabyle (Taqbaylit) NLP
non-profit
AI & ML interests
None defined yet.
Recent Activity
Organization Card
AƔBALU
Aɣbalu (ⴰⵖⴱⴰⵍⵓ) is Kabyle for the source — a spring, the place water comes from.
This organization builds NLP resources for Kabyle (Taqbaylit, kab, Latin script), a
Northern Berber language of Kabylia, Algeria, spoken by 5–7 million people. Kabyle is
unusual among low-resource languages: it has a disproportionately large speech resource
and a thin, noisy text resource. The entire clean Kabyle web crawl is smaller than a single
Tatoeba export.
The first artifact is not a model. It is the corpus.
Models
| Amrouche-1.3B | translation, 1.3B | Beats NLLB-200-1.3B in all four directions, and by the widest margin into Kabyle: 36.34 chrF++ and 10.86 BLEU on English→Kabyle, against NLLB's own published 6.2 BLEU for the language |
| Masinissa-31M | encoder, 31M | 90.51% on gold UD part-of-speech tags — against 83.42% for a most-frequent-tag baseline and 63.69% for the previously published Kabyle tagger |
| Mammeri-Tok | tokenizer | Ten Unigram vocabularies, 8k–32k, two initialisation arms. The sweep settles two open questions about Kabyle segmentation |
Still to come: speech recognition, voice, and a generative model. Each is named for a figure whose life was that model's work — see below.
Datasets
| KabBench | evaluation | A repaired Kabyle MT reference — 326 of its 2,009 sentences were corrupt in the public original and have never been fixed upstream — plus a balanced six-language set for telling Kabyle from its Berber siblings |
| KabLex | lexicon | 366,892 lexical entries over 17,090 lemmas, with morphological features, and 25,642 word–pronunciation pairs in IPA |
Every Kabyle BLEU score ever published is measured against a broken reference. 16.2% of
the Kabyle side of the standard MT benchmark carries homoglyph corruption, so a system that
spells Kabyle correctly is penalised for it. KabBench is that reference, repaired.
Both are licence-cut by code rather than by hand: KabLex publishes 366,892 of 395,834
entries, excluding a share-alike source and one whose licence could not be established, and
every row carries its own source and licence so a stricter subset can be cut downstream.
What makes these different
The orthography is repaired, not ignored. Public Kabyle text carries systematic homoglyph
corruption: Greek ε U+03B5 standing in for Latin ɛ U+025B, γ for ɣ, in 2.6–3.2% of
rows of the largest sources. Published Kabyle tokenizers have baked this into their merges at
a measured cost of +17.8% to +21.3% tokens on correctly spelled text. Every artifact here
passes a versioned 81-rule normaliser first, and every artifact is stamped with its version.
That repair shows up downstream as measurement rather than as a claim, twice. Under deliberate homoglyph corruption of its input, Masinissa-31M still scores 87.42% — above every baseline's clean number. And Amrouche-1.3B's score is identical whether or not both sides are normalised before scoring — a gap of 0.00 into Kabyle, where the untuned base model gains 0.17–0.19 — so the model has learned to spell Kabyle canonically.
Provenance is a precondition, not a footnote. Every sentence in the training corpus carries a source id, a licence, and a retrieval date. A record whose origin cannot be named does not enter. The licence composition of the training text is published on every model card, including the uncomfortable part: 34.9% of it has no licence anyone could resolve. A permissive grant on weights makes no claim about the text underneath them, and saying so is cheaper than pretending otherwise.
Evaluation reports what it measures. FLORES+ kab_Latn, the benchmark every published
Kabyle result is scored against, is 16.2% orthographically corrupt and has never been
revised upstream — which means 2.71 BLEU on its devtest is unreachable by a system that
spells Kabyle correctly. Results here are scored twice, as-published and with both sides
normalised, and the difference is reported. A published accuracy is a claim about a dataset,
not about a task: one Kabyle model card's 94.8% is agreement with the lexicon projection that
generated its own labels; on gold annotation it is 63.69%.
Decontamination is measured, with a positive control. Zero overlap between the training corpus and FLORES+ or SIB-200. Both derive from Wikipedia, which is in the corpus, so this was checked rather than assumed.
The names
Each model is named for someone whose life was the work that model does, in two families.
| model | does | name | |
|---|---|---|---|
| tokenizer | decides how Kabyle is written down | Mammeri | Mouloud Mammeri wrote the first Berber grammar written in Kabyle (1976), inventing the metalanguage to do it. The 1980 cancellation of his lecture on Kabyle poetry began the Berber Spring |
| encoder | understands; one representation | Masinissa | first king of a united Numidia, r. 202–148 BCE — he brought the eastern and western tribes into one kingdom, as the encoder brings 42 sources into one representation |
| translation | carries Kabyle out and back | Amrouche | Taos Amrouche sang exclusively in Kabyle and wrote in French — a life spent carrying one into the other |
| speech recognition | hears Kabyle, writes it down | Fadhma | her mother, Fadhma Aït Mansour Amrouche, who from 1930 began writing down the songs and tales inherited from her ancestors |
| voice | speaks Kabyle | Matoub | Matoub Lounès — the voice |
| generative | generates; sovereign | Jugurtha | Masinissa's grandson, who fought Rome 111–104 BCE and was never taken in battle |
Two lineages carry the design. Masinissa and Jugurtha are grandfather and grandson, so the smallest model and the largest are the founding of the kingdom and the refusal to surrender it. Fadhma and Taos Amrouche are mother and daughter, and they split the oral-tradition work exactly as their two models do: the mother heard the songs and wrote them down, which is speech recognition, and the daughter carried them into French while never singing in anything but Kabyle, which is translation. One of Taos's albums is called Chants sauvés de l'oubli — songs saved from oblivion.
Mammeri and Matoub stand alone, because what each of them did nobody shared: the grammar, and the voice.
The naming is homage. It implies no endorsement by anyone, and none of these figures or their families are affiliated with this work.
Licence
Weights are Apache-2.0. The text they were trained on is not, and cannot be relicensed by that grant — each model card publishes the full composition so you can judge for yourself.