A 31M-parameter masked language model for Kabyle (Taqbaylit, kab, Latin script), trained from scratch on the cleanest Kabyle corpus assembled to date.

Frozen, with only a linear probe fitted on top, it scores 90.51% on gold Universal Dependencies part-of-speech tags — against 83.42% for a most-frequent-tag baseline and 63.69% for the previously published Kabyle tagger. It is, as far as we can establish, the first Kabyle model to beat the statistical floor on gold annotation.

Under systematic homoglyph corruption of its input it still scores 87.42%, above every baseline's clean number.

Gold UD_Kabyle-ADPT test split: 845 sentences, 10,181 words, 8,563 whitespace tokens. All four systems scored by one harness, under four conditions.

Macro-F1 over 15 tags, surface/canonical: 0.7301 for the probe, 0.7051 for the baseline, 0.4957 for kabyle-pos-v2.

Three things worth reading carefully:

The encoder is frozen. Only the linear head is fitted. This measures what pretraining put into the representation, not what a task head can learn.

Both tokenisation settings are reported because they disagree. 29.5% of UD_Kabyle-ADPT words sit inside a multiword token, so "accuracy" is two different numbers depending on whether you score gold word segmentation or raw whitespace tokens. Reporting one is a choice that should be visible.

The corrupted condition is not noise injection for its own sake. Kabyle text in the wild carries systematic homoglyph substitution — Greek ε U+03B5 for Latin ɛ U+025B, γ for ɣ — in 2.6–3.2% of rows of the largest public sources. The corrupted condition applies exactly that substitution. The 3.1-point drop, from a baseline that loses 3.9, is the corpus-level orthographic repair showing up as downstream robustness.

Kabyle NLP where a sentence or token representation is needed: sequence labelling, token classification, sentence similarity, clustering, retrieval, and as an initialisation for task-specific fine-tuning. It fills [MASK] tokens directly.

Not suitable for: generation of any kind (it is a bidirectional encoder), translation, any language other than Kabyle, or any decision about a person. It has not been evaluated for bias, toxicity, or factuality, and its training corpus has not been audited for offensive content.

transformers and torch, nothing else. LTG-BERT is not one of the library's own architectures, so the modelling code travels in this repository and trust_remote_code=True is what loads it.

For representations rather than predictions, AutoModel returns the encoder alone:

Normalise the input first. The vocabulary was built over text where ɛ is U+025B and never Greek epsilon; homoglyph-corrupted text fragments into byte pieces and the representation degrades accordingly. This is not hypothetical — it is 2.6–3.2% of the seed corpora this model was trained from.

agbalu-tok-base-16k.model is the SentencePiece model the vocabulary was trained as, kept beside the converted tokenizer.json for anyone who needs the original. The two are id for id, which the release staging checks before publishing.

An LTG-BERT / GPT-BERT encoder, reimplemented from ltgoslo/gpt-bert — the BabyLM Challenge winner — with its published recipe.

Relative position buckets rather than RoPE, and a tied classifier, both following the reference implementation.

AƔBALU-Text v1 — 3,041,989 deduplicated Kabyle sentences from 42 sources, tokenised to 70,184,279 training tokens and 359,340 validation tokens (0.5% held out).

Every source carries a provenance record: source id, licence, retrieval date. Every sentence passed a reference normaliser at version 1.3.0+rules1.0.0, which repairs the homoglyph corruption described above while preserving the letters Kabyle actually uses (ɣ ɛ ḥ ḍ ṣ ṭ ẓ ṛ č ǧ ţ) and the emphatic and spirantised distinctions that carry meaning.

Decontamination. Measured against FLORES+ kab_Latn and SIB-200 at zero overlap, with a positive control confirming the detector fires when contamination is present. Both benchmarks derive from Wikipedia, which is in the corpus, so this was checked rather than assumed.

The weights are Apache-2.0. That grant does not relicense the text they were trained on, and a third of that text has no licence anyone could resolve. The composition is published so you can make your own judgement:

unclear is not a permissive category — it is the absence of a resolvable licence. It is 29.9% sources declaring other, plus 5.0% declaring cc, which is not a licence at all because no variant is given. One source contributing 711 sentences has unsettled copyright.

A permissive-only rebuild is possible by code from the 973,218 permissive sentences. If you need one, open an issue.

Masking is inverse — harder early, standard late — following the reference recipe, and independently arrived at by mmBERT. LAMB at 1.2e-2 does not transfer to AdamW; the learning rate is specific to the optimiser.

Final validation loss 3.1957 (perplexity 24.43), monotonically improving across all 18 evaluations:

It is over-trained on unique tokens, and this is the binding constraint. 67 epochs over 70.18M tokens, against a literature threshold of roughly 4 repetitions before returns collapse (Muennighoff et al. 2023). The run shows it directly: between steps 3,000 and 3,250 training loss fell 0.22 while validation loss fell 0.027. More steps will not help, and the model is not too large either — LTG-BERT runs 0.77 parameters per token against this model's 0.443. The constraint is unique Kabyle text, and there is very little more of it.

It cannot represent the annexed state. Kabyle marks a word-initial state alternation (axxam "house" → wexxam in the annexed state). The vocabulary memorises both forms whole at every size tested from 4k to 48k, in both initialisation arms — 0 of 15 test pairs share a stem. This is a measured refutation of a design hypothesis, not an oversight, and it means morphological structure of this kind has to come from the objective or an explicit morphological layer rather than from segmentation.

Sibling-language contamination is bounded but not cleared. Kabyle has closely related neighbours — Tashelhit, Tarifit, Central Atlas Tamazight, Tamasheq, Shawiya — and many "Tamazight"-labelled datasets silently mix them. Corpus sources were language-identified, but neither GlotLID nor NLLB's lid218e can name Tarifit, Central Atlas Tamazight or Shawiya, so a kab_Latn label cannot exclude them. Measured on a balanced set, NLLB's identifier labels 87–95% of Tashelhit, Tarifit and Central Atlas Tamazight as Kabyle. What is excluded is Tashelhit and Tamasheq, which both systems can name.

Not evaluated beyond POS. One downstream task, one treebank, 845 test sentences. Treat everything else as unmeasured.

No safety evaluation of any kind has been performed.

agbalu/Belaid-31M restores punctuation and capitalisation on Kabyle ASR output — two token-classification heads over this encoder, at 0.793 macro-F1 over marks against a rule baseline's 0.227.

The weights file is exactly 31,123,840 float32 parameters. Two things are deliberately not in it: the classifier decoder, which is tied to the embedding table and re-tied on load; and twelve identical 512×512 relative-position index tables, which the module rebuilds byte-identically in __init__ and which would otherwise add 25.2 MB — 17% of the download — of derived data. The exporter asserts both before writing.

The training checkpoint, with LAMB's optimizer moments and RNG state, is 398 MB and is not published. Ask if you need it to resume training.

The source checkpoint is best.pt, SHA-256 b37b77ca22f0ed8e52dfcc52307cd09e4536d71670917051eba824ebf0c7a743, and it loads into the architecture above with zero missing and zero unexpected keys.

The POS numbers reproduce exactly. Head initialisation is seeded per-layer with an explicit generator, so two fits of the same frozen checkpoint are bit-identical. Seeding only the batch order is not enough: it leaves the reported accuracy moving by up to 0.5 points per refit.

Masinissa (r. 202–148 BCE) was the first king of a united Numidia, who brought the eastern and western tribes into one kingdom. The encoder is the unifying representation of this project — 42 noisy sources rendered into one shared space — so it carries his name.

His grandson Jugurtha, who fought Rome from 111 to 104 BCE and was never taken in battle, is reserved for the generative model. The naming is homage; it implies no endorsement by anyone.

Apache-2.0 for the weights and code. Read the licence composition of the training text above before redistributing derivatives — a permissive grant on the weights makes no claim about the underlying text, 34.9% of which has no resolvable licence.