A 26.9M-parameter character-level Conv-Transformer for Kabyle (Taqbaylit, kab), trained from scratch to convert Neo-Tifinagh (ⵜⴰⵇⴱⴰⵢⵍⵉⵜ) into standard Kabyle Latin (Taqbaylit) — and, in doing so, to restore the vowel e that Tifinagh does not write.

The vowel is the whole task. Every consonant maps one-to-one between the two scripts, so a lookup table converts them in an afternoon. Kabyle Neo-Tifinagh omits the schwa, and where it belongs is a function of the surrounding consonants. That is a sequence problem, and it is why a character table reaches 1.2% sentence exact match on the test split where this model reaches 94.2%.

Test split of agbalu/KabTifinagh: 49,795 sentences, disjoint from train and dev, never seen during training or model selection. Free-running greedy decoding — the model is fed its own output, which is what a caller gets.

Both rows are measured on the same 5,000 sentences in the same pass, by agbalu.bench.tifinagh. The character table's figures over the full 49,795 are 1.02% / 13.49% / 1.72%, so the sample is not flattering it.

Reproduce either without a GPU:

Three things worth reading carefully.

The baseline is a table, not another model. No neural model had been trained for Kabyle script conversion or schwa restoration before this one. The comparison is therefore against the only tool that existed, and the gap is not incremental: the table gets 1.2% of sentences exactly right because it gets almost every schwa wrong, and one wrong character fails a sentence.

Schwa F1 measures placement, not count. A hypothesis with the right number of e in the wrong positions scores 100% under a count and 0% here. The metric removes every e from both strings, compares the consonant skeletons, and matches the vowels by their index into that skeleton; a hypothesis whose skeleton differs has no position to be judged at, so its vowels are charged to both error columns rather than dropped. 251 of 5,000 sentences fall in that class.

Character accuracy overstates this task and sentence exact match is the honest number. 0.33% CER means roughly one character in three hundred; 5.78% of sentences still contain at least one error. Most are hyphenated clitic chains — yefka-yas-d-t. — where the model must predict schwa placement and hyphen boundaries at once, and the hyphens are not in the Tifinagh source at all.

Converting Neo-Tifinagh Kabyle (kab_Tfng) into Kabyle Latin (kab_Latn): archival and educational corpora into a form the rest of the NLP stack can read, and schwa restoration for any Tifinagh source where the vowel was omitted by convention.

The reverse direction needs no model. Latin to Tifinagh is a character table, and it is lossless in that direction because Tifinagh simply does not write the schwa this model restores.

Not suitable for: translation between Kabyle and any other language (use Amrouche-1.3B); any decision about a person; or any language other than Kabyle. Tarifit, Tashelhit and Central Atlas Tamazight share part of the orthography and are not evaluated here — this project does not treat them as Kabyle, and neither should a caller.

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

Every example above was executed against the published directory before being written down, and both decodings are free-running — the model is fed its own output. num_beams=1 is the greedy decode; on these sentences it agrees with beam search.

Batch rows of equal length, or decode one at a time. The model attends over padding and attention_mask is accepted and ignored, because that is how the weights were evaluated; honouring a mask here would return different logits from the numbers above.

There is no key-value cache. Every step re-reads the whole prefix, which is what the published evaluation did and what makes the two agree. Inference is a few milliseconds per sentence on CPU; nothing here needs a GPU.

Output is lowercase. The model is defined over a case-folded alphabet and does not restore capitalisation; .capitalize() covers the sentence-initial case and proper nouns do not have a solution here. A snippet showing capitalised output would not reproduce.

The convolutional stem is not decoration. Tifinagh writes consonant clusters with the vowel deleted, so the evidence for where a schwa belongs is the adjacent two to five characters. Depthwise kernels of 3 and 5 hand the first encoder layer that window already computed, instead of spending attention capacity on adjacency.

Rotary rather than absolute positions, because the cue is the distance between consonants, not where in the sentence the cluster sits. RoPE is applied to self-attention only: cross-attention relates two sequences of different lengths, and a shared index there would assert an alignment that schwa restoration falsifies.

Tied embedding and output projection, because the input and output alphabets are the same alphabet. That is the correct constraint, not a saving.

agbalu/KabTifinagh, script_conversion config — 497,944 sentence pairs, split 398,355 / 49,794 / 49,795 at seed 42 with no sentence in two splits.

The Latin side is normalised under 1.3.0+rules1.0.0, the same normaliser as Masinissa-31M and Amrouche-1.3B. That is load-bearing rather than tidy: a corpus where ɛ is sometimes Greek epsilon teaches two consonant contexts where the language has one.

The 123,852 English and 205,637 French trilingual alignments in the same dataset were not used to train this model.

Hyphenated clitic chains are the hardest cases and account for most of the 5.78% sentence-level error. The hyphens are absent from the Tifinagh source and are predicted entirely from morphosyntactic context.

Capitalisation is not restored, as above — Neo-Tifinagh is unicameral, so the information is not in the input. agbalu/Belaid-31M supplies it downstream, at 0.933 F1 on non-initial proper nouns.

The training pairs are rule-derived, not human-transcribed. The Tifinagh side was produced by applying a deterministic Latin→Tifinagh mapping to Kabyle Latin text, and the model learns to invert it. Any systematic error in that mapping is inherited, and the error rates above are measured on the round trip — not against independently authored Tifinagh. This is the single largest caveat on the numbers.

Kabyle Neo-Tifinagh only. Tuareg Tifinagh, Moroccan variants and Libyco-Berber inscriptions are out of scope and untested.

One task, one corpus. OCR post-correction, keyboard input and archival digitisation are plausible uses and are not evaluated here.

No safety evaluation of any kind has been performed.

The training checkpoint this was exported from holds the weights, the config and the final dev accuracy — and nothing else. It carries no optimizer state, no scheduler state and no step-by-step validation curve, so training cannot be resumed from the published files. export.stats.json records the source's contents, which is what makes that checkable rather than a claim.

lm_head.weight is absent from model.safetensors on purpose: it is embed.weight under a second name, safetensors refuses to write shared storage twice, and the module re-ties it on load.

Juba II (r. 25 BCE – 23 CE) was a Numidian king who wrote in Greek on geography and Berber history, and whose court at Caesarea worked across scripts and languages at once. The model that carries his name does the inverse of his career: it recovers Kabyle Latin orthography from a script that encodes it without vowels.

This project had previously ruled Juba out for the translation model, on the grounds that a Roman client king writing in the empire's language is the wrong figure for a language-sovereignty project. That reasoning was about translation. Moving between scripts is what his life actually was, and it is what this model does.

The naming is homage; it implies no endorsement by anyone.

Apache-2.0 on the weights and the code. The training data derives from AƔBALU-Text v1; a permissive grant on weights makes no claim about the text they were trained on, so read agbalu/KabTifinagh's licence before redistributing derivatives.