A machine translation model for Kabyle (Taqbaylit, kab), fine-tuned from NLLB-200-distilled-1.3B on 544,729 human-authored pairs — everything in the AƔBALU parallel corpus that NLLB did not mine itself.

It beats its own base model in all four directions, and by the widest margin where it matters most: into Kabyle.

chrF++ and BLEU on FLORES+ devtest, against the orthographically corrected kab_Latn reference. Same harness for every row.

10.86 BLEU on eng→kab against NLLB's own published 6.2 for Kabyle, which this same harness reproduces at 6.02 on the 600M model — so the comparison is calibrated, not asserted.

The gains are asymmetric on purpose. Generating Kabyle is the unsolved half: kab→eng was already at 44.60 on the base model because English generation is solved, and it gains 1.65. eng→kab gains 4.81. The corpus is doing its work on the side where there was work to do.

Every result is scored twice — as-published, and with both sides passed through the reference normaliser — and the difference reported. FLORES+ kab_Latn is 16.2% homoglyph-corrupted and was never revised upstream, so a system that spells Kabyle correctly is penalised by the reference itself.

Normalising both sides no longer changes the score in any direction. The model already spells Kabyle canonically — the corpus-level orthographic repair showing up as model output.

Translation between Kabyle and English or French, in either direction. For a source language this model was not trained on, pivot through English: X→eng with stock NLLB, then eng→kab here, which is the direction with the largest gain.

Not suitable for: kab→X where X is neither English nor French — the base model's ability survives but gains nothing here, and pivoting through kab→eng (46.25 chrF++, the strongest direction) will do better. Also not suitable for safety-critical, legal or medical translation. No adequacy evaluation by human annotators has been performed on the training corpus or the output.

transformers and torch. The architecture is M2M100ForConditionalGeneration, one of the library's own, so no trust_remote_code is needed — but the vocabulary is not the base's.

This model has a trimmed vocabulary, so from_pretrained alone is not enough. The embedding was cut to the 52,209 tokens the fine-tuning corpus uses, while the tokenizer shipped beside it still speaks NLLB's full 256,206. Ids must be translated in both directions or the output is fluent nonsense. keep.json is the translation table, and it is in this repository.

Ids cross the boundary in four places, and a version that handles three of them returns plausible text in the wrong language: the source input_ids, the forced target-language token, the generated ids on the way back, and unk itself. Language codes are NLLB's — kab_Latn, eng_Latn, fra_Latn — and all of them survive the trim.

Special tokens are unaffected: ids 0–49 map to themselves, so <s>, <pad>, </s> and <unk> keep their usual values.

M2M100ForConditionalGeneration, the base's own, with one thing changed: the embedding.

The trim is the only architectural change and it is the one a caller has to know about. Dropping 203,997 unused rows removes 0.84 GB of embedding that this corpus never addresses, and it means the shipped tokenizer's ids are not the model's — see Usage above.

The corpus deliberately excludes NLLB's own mined bitext. 90.1% of the available Kabyle parallel data is NLLB-mined with unmeasured precision; training a NLLB derivative on it would be distilling the base model's own output. What remains — 552,534 non-mined pairs, 512,049 of them defect-free — is still 7.7× more public Kabyle bitext than NLLB's paper reports having seen (72,000 sentences, Table 12).

The fine-tuning corpus was checked for sibling contamination before any GPU time was spent: 98.31% of its judged Kabyle side is kab_Latn, and 14 lines of 38,486 carry a Berber sibling label. The 1.69% residue is untranslated English and French strings from localisation exports, not sibling text.

LoRA was ruled out by measurement in the literature, not by preference — arXiv 2404.04212 reports 18.63 BLEU against full fine-tuning's 30.25 on this class of task.

The vocabulary is trimmed, and this understates the scores above. The embedding table was scanned from the fine-tuning corpus and cut to 20.4% of NLLB's. FLORES+ then contains tokens the scan never saw — 0.07–0.22% of tokens, but 28–55 sentences per thousand — which are mapped to unk at generation. The baselines carry no such handicap, so the reported gains are a floor.

Adequacy is unmeasured, and will stay unmeasured. 90.1% of the available Kabyle parallel data is NLLB bitext-mined with unmeasured precision. This model's corpus excludes it, but the excluded portion's quality was never human-annotated either. The measured mechanical defect rate is a lower bound on errors, not a precision figure — do not quote it as one.

Sibling-language contamination is bounded, not cleared. Neither GlotLID nor NLLB's own lid218e can name Tarifit, Central Atlas Tamazight or Shawiya, so a kab_Latn label cannot exclude them. Measured on a balanced set, lid218e labels 87–95% of Tashelhit, Tarifit and Central Atlas Tamazight as Kabyle — and that identifier is what mined most public Kabyle bitext.

One benchmark, one domain. FLORES+ is Wikipedia-derived prose. Nothing here says how the model behaves on speech transcripts, dialogue, or the localisation strings that make up a sixth of the training corpus.

The tokenizer is deliberately untrimmed. Cutting it too would break every NLLB language code and the src_lang machinery that selects them; the mapping belongs in keep.json, where a caller can see it.

Taos Amrouche (1913–1976) sang exclusively in Kabyle and wrote in French. She published Jacinthe noire in 1947, the first novel by a Kabyle woman; from 1936 she collected and performed the Kabyle songs her mother had preserved; and she co-founded the Académie berbère in 1966. One of her albums is Chants sauvés de l'oubli — songs saved from oblivion.

A life spent carrying Kabyle into another language without ever surrendering it is what a translation model is for. Her mother, Fadhma Aït Mansour Amrouche, who from 1930 began writing down the songs and tales inherited from her ancestors, gives her name to the speech recognition model — the two halves of one family's work, split the way the two models are.

The naming is homage. It implies no endorsement, and neither Taos Amrouche nor her family is affiliated with this work.

Apache-2.0 for the weights. NLLB-200 itself is CC-BY-NC-4.0; check the base model's terms for your use case.

That grant does not relicense the text the weights were trained on. By licence, the 5.56M-pair parallel corpus this fine-tuning set was drawn from is unclear 4,871,469 pairs, permissive 376,389, non-commercial 297,058, share-alike 16,710. The unclear bulk, not the non-commercial slice, is the redistribution risk — unclear is the absence of a resolvable licence, not a permissive one. It is stated here rather than left for someone to discover.