Juba-27M

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%.

Results

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.

system	sentence exact match	character error rate	schwa placement F1
Juba-27M	94.22%	0.33%	93.51%
deterministic character table	1.16%	13.55%	1.61%

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:

make tifinagh TASK=evaluate LIMIT=5000   # the model
make bench TASK=tifinagh                 # the table, over the whole split

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.

Intended use

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.

Usage

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.

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

REPO = "agbalu/Juba-27M"
tokenizer = AutoTokenizer.from_pretrained(REPO, trust_remote_code=True)
model = AutoModelForSeq2SeqLM.from_pretrained(REPO, trust_remote_code=True).eval()


def transliterate(text, num_beams=4):
    encoded = tokenizer(text, return_tensors="pt")
    generated = model.generate(**encoded, num_beams=num_beams, max_length=256)
    return tokenizer.decode(generated[0], skip_special_tokens=True)


transliterate("ⵜⵛⴼⵉⴹ ⴼⵍⵍⵉ ?")
# 'tecfiḍ fell-i ?'
transliterate("ⴰⵣⵓⵍ ⴼⵍⵍⴰⵡⵏ, ⴰⵎⴽ ⵜⵜⵉⵍⵉⴹ ?")
# 'azul fell-awen, amek ttiliḍ ?'

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.

Architecture

Parameters	26,901,888
Encoder / decoder layers	6 / 6
Hidden / feed-forward	384 / 1,152 (SwiGLU)
Attention heads / head size	6 / 64
Positions	rotary (RoPE), on self-attention only
Conv stem	1D depthwise separable, kernels 3 and 5, before the encoder
Vocabulary	128 character slots; 96 used, 0 out-of-vocabulary on Kabyle
Tied weights	input embedding is the output projection
Label smoothing	ε = 0.05

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.

Training data

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.

Training recipe

Objective	character cross-entropy, label smoothing ε = 0.05
Optimiser	AdamW, lr 5e-4, β (0.9, 0.95), weight decay 0.1
Batch	micro-batch 64 × gradient accumulation 2 = 128
Schedule	cosine annealing to η_min 1e-5 over 15,500 steps
Steps	15,500 optimizer steps, ~5 epochs
Gradient clipping	max norm 1.0
Hardware	one NVIDIA A10G (Modal), detached
Seed	42

make modal-tifinagh TASK=train

Limitations

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.

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.

Files

file	contents
model.safetensors	145 tensors, 26,901,888 parameters
config.json	the architecture, and the auto_map that points at the code below
configuration_juba.py, modeling_juba.py	the architecture in code, importing only torch and transformers
tokenizer.json, tokenizer_config.json	the 96-character alphabet, id for id with the one the model was trained on
export.stats.json	what the source checkpoint held, and what the export dropped

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.

The name

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.

Citation

@software{agbalu_juba_2026,
  title  = {Juba-27M: character-level script conversion and schwa restoration for Kabyle},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/agbalu/Juba-27M},
  note   = {Trained on agbalu/KabTifinagh; normaliser 1.3.0+rules1.0.0}
}

Licence

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.