Hugging Face's logo
Activity Feed

AI & ML interests

None defined yet.

Recent Activity
ainouche-abderahmane  updated a Space about 21 hours ago
agbalu/README
ainouche-abderahmane  updated a model about 23 hours ago
agbalu/Masinissa-Sentiment
ainouche-abderahmane  published a model about 23 hours ago
agbalu/Masinissa-Sentiment
View all activity

Organization Card
Community About org cards

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. Its entire filtered web crawl — FineWeb-2 and HPLT v2 together — is 13.8M words, and the corpus underneath these models, deduplicated across 42 sources, is 34.9M words.

The first artifact is not a model. It is the corpus.

The speech model is where that asymmetry pays: Kabyle has 571 validated hours of audio and roughly 70 million unique tokens of text, and transcription is what turns the abundant resource into the scarce one.

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 31.53 and 7.74 for the base model on the same harness
SiMohand-278M sentence embeddings and retrieval, 278M 97.0% Recall@1 and 0.9833 MRR finding the right Kabyle passage among 499 distractors — against 63.8% for the multilingual backbone it was fine-tuned from, which maps five Kabyle consonant characters to <unk> and collapses the language into one cone. Its 64-dimension Matryoshka slice, a 12× compression, scores the same 97.0%
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. 0.8880 macro F1 on three-class sentiment, against 0.7764 from a linear probe on the frozen encoder
Mammeri-Tok tokenizer Ten Unigram vocabularies, 8k–32k, two initialisation arms. The sweep settles two open questions about Kabyle segmentation
Juba-27M transliteration, 27M 94.22% sentence exact match converting Tifinagh to Kabyle Latin, against 1.16% for the character table that was the only prior tool — because Kabyle Tifinagh omits the vowel e, and where it belongs has to be predicted from the consonants around it
Fadhma-300M speech recognition, 300M 8.01% character error rate and 25.65% word error rate on 15,003 utterances from 888 speakers it has never heard — as far as we can establish, the first Kabyle speech recognition system published with an error rate measured on a speaker-disjoint split, and the first CTC result for the language at any size
Belaid-31M punctuation and casing, 31M 0.793 macro-F1 over marks on 5,160 held-out sentences, against 0.227 for the rule a system without a model uses. It reads what Fadhma writes and puts back the punctuation a CTC vocabulary cannot emit
Boulifa-48M orthography standardisation, 48M 97.39% character accuracy and 85.70% sentences exactly right converting informal French-keyboard and Arabizi Kabyle into canonical Kabyle Latin, against 89.70% character accuracy for leaving the input untouched — as far as we can establish, the first orthography standardisation model published for any Berber language. Resolves Arabizi digits, digraphs, clitic hyphens and preposition contractions in one pass, where each is ambiguous on its own
Feraoun-36M document OCR, 36M 2.85% character error rate and 70.20% line exact match over 1,000 held-out lines of printed Kabyle, reading Latin and Neo-Tifinagh from one 171-symbol character alphabet — the first OCR model trained for the language, and the first that keeps the sub-dot on ḍ ḥ ṛ ṣ ṭ ẓ instead of normalising it away
Feraoun-Pro document OCR, 51M 7.34% Latin and 0.91% Neo-Tifinagh character error rate, 81.77% line exact match over 30,000 held-out rendered lines. A second architecture for Feraoun-36M's task: one CTC forward pass per line instead of one decoder step per character, a canvas that follows the line to 1,024 px instead of squeezing it, and ţ and Ţ — Boulifa's transcription convention, which has never had a slot in an OCR alphabet before
Matoub-82M speech synthesis, 82M Speaks Kabyle at 24 kHz — twice the incumbent mms-tts-kab, and Apache-2.0 where that one is non-commercial — over a 42-symbol Kabyle phoneme inventory the base model's token table could not represent until three embedding rows were assigned and trained for ˤ ʕ ħ, which its own G2P would otherwise have dropped in silence. One voice published. The baseline it will be measured against, +3.56 Cycle-CER, was itself unscored by anyone until this project measured it
Dihya-5M language identification, 5M Separates six Latin-script Amazigh varieties (kab, shi, rif, taq, tzm, shy) and rejects non-Berber: 85.52% over the seven classes. On the six Berber classes, scored beside the public identifiers in one pass over the same rows, 84.44% against GlotLID's 41.78% and NLLB lid218e's 28.33% — and where those send 96.0% of Shawiya and 90.0% of Tarifit into Kabyle, this sends 19.3% and 6.7%. That absorption is what lid218e did to the 90.1% of AƔBALU-Parallel it mined. Tarifit, Central Atlas Tamazight and Shawiya have no label at all in either public identifier, so their recall here — 0.873, 0.660, 0.687 — is the first that exists. Returns a conformal set of labels where the text is genuinely ambiguous
Masinissa-Sentiment sentiment, 31M 90.27% accuracy and 0.9026 macro-F1 over negative, neutral and positive — the first sentiment classifier published for any Berber language, against 88.80% for an ordinary fine-tune of the same encoder and 77.53% for a linear probe on it, all three scored on the same 1,500 rows. Its labels are projected from the English side of a Tatoeba pair, and the transfer is measured rather than assumed: accuracy rises monotonically with the English classifier's own confidence — 87.85% on the 889 rows it was least sure about, 95.00% on the 40 it was surest of. Total homoglyph corruption moves it not at all, because the repair is inside the inference path. Returns a conformal set of labels on the 12.53% of sentences where the polarity is in the situation rather than the words

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
KabSentiment sentiment 15,000 human-written Kabyle sentences, labelled by projection from their English parallels and balanced exactly across three classes — the first Kabyle sentiment set that covers neutral, and the first whose text survives the language's own orthography. Its 12,000 training rows are what Masinissa-Sentiment is fitted on
KabInflect morphology 336,151 inflected verb form entries across 13,226 lemmas, partitioned into sealed splits (0 paradigm leakage), plus 6,198 full verb conjugation paradigms
KabTifinagh transliteration 497,944 parallel sentences matching Kabyle Tifinagh to canonical Kabyle Latin, plus 123,852 English and 205,637 French trilingual sentence alignments
KabPunct punctuation restoration 1,318,707 word-labelled sentences with per-token punctuation (NONE, COMMA, PERIOD, QUESTION, COLON) and casing (LOWER, UPPER_INIT) labels — the training and evaluation corpus for Belaid-31M, and as far as we can establish the first labelled punctuation restoration corpus for any Berber language
KabG2P phonetics / G2P 25,634 Kabyle word forms with IPA transcriptions recovered at 99.53% alignment rate from 292,921 aligned tokens — zero ambiguity, 42-symbol phonetic inventory, and per-entry spirantisation and vowel-backing annotations; the pronunciation dictionary for Matoub-82M and any downstream Kabyle TTS or ASR system
KabStandard orthography standardisation 497,944 parallel pairs mapping informal French-keyboard and Arabizi Kabyle to canonical Kabyle Latin — training and evaluation corpus for Boulifa-48M. Fully synthetic and reproducible at seed 42 from agbalu/KabTifinagh

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.

KabLex publishes 366,892 of 395,834 entries, licence-cut by code: excluding a share-alike source and one whose licence could not be established, with every row carrying its own source and licence so a stricter subset can be cut downstream.

KabSentiment is the only Kabyle sentiment dataset that covers neutral and that passes the language's orthography through without destruction — the prior binary corpus carries 0.00% retention of ɣ, against 45% in real Kabyle text. It is also a training set, not only a benchmark: Masinissa-Sentiment is fitted on its 12,000 train rows and scored once on the 1,500 it holds out.

KabInflect splits by verb lemma, not by verb form. Kabyle verb morphology is templatic, so a split by row puts two forms of one verb on either side of it and measures memorisation; splitting by lemma means no test verb's paradigm was seen in any cell. Copying the lemma unchanged scores 4.07%, which is the floor every result is read against.

KabTifinagh is 497,944 sentence pairs in both scripts. Kabyle Neo-Tifinagh writes no e, so converting back to Latin is not a lookup — the vowel has to be predicted. A character table gets 1.16% of sentences exactly right; Juba-27M, trained on this, gets 94.22%.

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 all 2,009 FLORES+ kab_Latn sentences, which are also the text SIB-200 reuses. FLORES+ is built from Wikipedia, which is in the corpus, so this was checked rather than assumed, and the detector is positive-controlled against injected leaks — including one that differs only by the homoglyph corruption.

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
sentence embeddings recalls Kabyle exactly Si Mohand Si Mohand ou Mhand (1848–1905), master of the Kabyle isefra, who wrote none of them down — the verses survive because they were heard once and recalled intact
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
transliteration translates scripts and restores text Juba King Juba II (48 BCE – 23 CE) — the ancient Numidian scholar-king, linguist, and author whose life's work was the study and preservation of North African scripts and language
speech recognition hears Kabyle, writes it down Fadhma Fadhma Aït Mansour Amrouche, Taos's mother, who from 1930 began writing down the songs and tales inherited from her ancestors
punctuation and casing writes Kabyle as prose on a page Belaïd Belaïd At Ali (1909–1950), founder of written Kabyle prose. Asked to set down oral tales, he composed instead — Les Cahiers de Belaïd, the first Kabyle prose in Latin characters, and the first text whose author had to decide where a Kabyle sentence ends
voice speaks Kabyle Matoub Lounès Matoub (1956–1998) — recorded 36 albums in Kabyle at a time when the Algerian state was suppressing Berber language and culture, making the language audible to an entire generation. Assassinated on 25 June 1998, ten days before the Arabisation law he had spent years opposing took effect. His voice is inseparable from the survival of Kabyle as a spoken language in collective memory
orthography standardisation writes Kabyle correctly from any input Boulifa Si Amar ou Saïd Boulifa (1865–1931) — the first grammarian to systematically codify Kabyle Latin orthography. His Recueil de poésies kabyles (1904) set 555 pages of Kabyle in one consistent Latin notation, solving in print what this model solves from a keyboard
document OCR reads Kabyle from printed paper and books Feraoun Mouloud Feraoun (1913–1962), novelist and teacher who taught reading and transcribed the oral poetry of Si Mohand onto the printed page
language identification tells Kabyle from its siblings Dihya the Amazigh leader who fought the Umayyad conquest from the Aurès, died c. 703 CE — the Shawiya homeland, and Shawiya is the variety this model exists to stop being absorbed into Kabyle
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.