Organization Card

AƔBALU

Aɣbalu (ⴰⵖⴱⴰⵍⵓ) is Kabyle for the source — a natural spring where water originates.

This organization develops open-source NLP and speech resources for Kabyle (Taqbaylit, kab, Latin script), a Northern Berber language spoken by 5–7 million people in Kabylia, Algeria. While public web crawls for Kabyle are relatively scarce (FineWeb-2 and HPLT v2 total approximately 13.8M words), the language possesses extensive speech resources, including over 571 validated hours in Mozilla Common Voice. The AƔBALU project builds foundational datasets, models, and benchmarks to expand Kabyle language technology from curated corpora to high-capacity foundation models.

Models

Model	Task & Capacity	Description
Jugurtha-2-9B	Sovereign foundation & reasoning LLM, 8.95B	Second-generation sovereign foundation model for Kabyle. Achieves 80.81% on SIB-200 topic classification, 49.45 chrF++ / 27.86 BLEU on FLORES+ kab→eng, and 30.80 chrF++ / 7.62 BLEU on eng→kab with zero orthography gap (0.00). Native chain-of-thought in Taqbaylit (`<think>` blocks). Also available in GGUF and MLX
Jugurtha-4B	Foundation & reasoning LLM, 4.2B	Sovereign Kabyle generative model with native step-by-step reasoning (`<think>` blocks). Continued pretraining on 800M tokens and SFT on curated Kabyle dialogues. 100.0% NIAH at 32,768 context, 80.89% GSM8K, 62.86% MMLU-Pro, 54.16% IFEval, 73.74% SIB-200 topic classification on Kabyle. Available in GGUF and MLX
Amrouche-2-1.3B	Machine translation, 1.37B	Six-direction translation model (`kab ↔ eng`, `kab ↔ fra`, `kab ↔ arb`) preserving the full 256,206-token NLLB vocabulary and outperforming NLLB-200-1.3B across all six FLORES+ `devtest` directions (47.78 chrF++ / 25.81 BLEU on kab→eng, 44.64 / 21.68 on kab→fra, 35.26 / 9.90 on eng→kab, 34.98 / 10.19 on kab→arb, 33.53 / 7.65 on fra→kab, 33.13 / 7.54 on arb→kab; 52.36 macro chrF++ on held-out gold test). Also available in CTranslate2 int8 (1.38 GB, 7.1× faster CPU inference)
Amrouche-2-600M	Machine translation, 615M	Lightweight, high-throughput six-direction translation model (34.88 macro chrF++ / 11.09 BLEU on FLORES+ `devtest`, 48.58 macro chrF++ on held-out gold test, +45% relative BLEU gain on eng→kab) at 1.23 GB BF16. Also available in CTranslate2 int8 (622.6 MB, 155 ms single-sentence latency)
Amrouche-2-1.3B-CT2-int8	Edge translation, 1.37B (int8)	Standalone CTranslate2 int8 release of Amrouche-2-1.3B. Delivers 205.6 ms single-sentence CPU latency, 29.22 sent/s 16-thread server throughput, and 1.38 GB footprint with only -0.14 chrF++ quantization delta
Amrouche-2-600M-CT2-int8	Edge translation, 615M (int8)	Standalone CTranslate2 int8 release of Amrouche-2-600M. Ultra-compact 622.6 MB model size, 155.0 ms single-sentence CPU latency, 47.33 sent/s server throughput, and 4.9× speedup over PyTorch with only -0.15 chrF++ quantization delta
Amrouche-1.3B	Machine translation, 1.3B	Vocabulary-trimmed translation model outperforming NLLB-200-1.3B across all four directions (36.34 chrF++ / 10.86 BLEU on English→Kabyle vs. 31.53 / 7.74 for the baseline)
Fadhma-300M	Speech recognition, 300M	Acoustic CTC model achieving 8.01% character error rate and 25.65% word error rate on 15,003 utterances across 888 unseen speakers in Common Voice Kabyle
Belaid-75M	Discourse punctuation & truecasing, 75M	Restores punctuation and truecasing across continuous multi-sentence discourse using a 1,024-token context window over Masinissa-73M (0.630 period F1, 0.623 comma F1, 94.0% word casing accuracy, 0.755 acronym F1)
Belaid-31M	Utterance punctuation & casing, 31M	Sentence-level punctuation and casing restoration for ASR transcripts (0.793 macro-F1 and 86.3% exact sentence match across 5,160 held-out utterances)
Matoub-4B	Speech synthesis, 4.2B	Neural codec language model with an 8-codebook 24 kHz audio tokenizer fine-tuned on Common Voice Kabyle. Supports zero-shot voice cloning and cross-lingual reference conditioning. Also available in GGUF
Matoub-82M	Speech synthesis, 82M	24 kHz neural TTS model with a 42-phoneme Kabyle inventory including dedicated embeddings for `ˤ`, `ʕ`, and `ħ`. Shipped with single voice checkpoint
Boulifa-48M	Orthography standardisation, 48M	Converts informal French-keyboard input and Arabizi into canonical Kabyle Latin (97.39% character accuracy, 85.70% sentence exact match)
Juba-27M	Transliteration, 27M	Transliterates vowel-less Neo-Tifinagh to canonical Kabyle Latin with learned schwa placement (94.22% sentence exact match, 96.06% schwa F1)
Feraoun-Pro	Document OCR, 51M	Vision-encoder CTC document OCR with variable-width canvas up to 1,024 px and 173-character alphabet including `ţ` / `Ţ` (7.34% Latin CER, 0.91% Neo-Tifinagh CER, 81.77% exact line match)
Feraoun-36M	Document OCR, 36M	Vision transformer OCR reading Latin and Neo-Tifinagh across a 171-character alphabet (2.85% CER and 70.20% line exact match on held-out printed lines)
SiMohand-278M	Sentence embeddings & retrieval, 278M	Kabyle passage retrieval model fine-tuned with Matryoshka representations (97.0% Recall@1, 0.9833 MRR; 64-dimensional slice retains 97.0% Recall@1)
Masinissa-73M	Flagship encoder, 73M	High-capacity masked language model leading all models on `KabEncBench` (1,953.7 Bradley–Terry Elo; 94.08% surface POS accuracy, 80.59% UPOS macro-F1)
Masinissa-31M	Lightweight encoder, 31M	Efficient bidirectional encoder (1,708.4 Bradley–Terry Elo on `KabEncBench`, #2 overall; 90.51% gold POS surface accuracy)
Masinissa-Sentiment	Sentiment classification, 31M	Three-class sentiment classifier with split-conformal prediction (90.27% accuracy, 0.9026 macro-F1, 94.00% conformal coverage at 1.13 mean set size)
Dihya-22M	Pan-Berber & Maghrebi LID, 22M	Language identification across 22 classes and 5 scripts, covering Berber varieties in Latin and Neo-Tifinagh, Maghrebi Arabic dialects, MSA, and world languages (91.90% accuracy, 0.9180 macro-F1)
Dihya-5M	Lightweight language identification, 5M	Distinguishes six Latin-script Amazigh varieties and rejects non-Berber text (85.52% overall accuracy, 84.44% on Berber classes)
Mammeri-Tok	Tokenizer collection	Ten SentencePiece Unigram vocabularies (8k–32k) across standard and lexicon-seeded initialization arms

Interactive Spaces

Space	Focus	Description
KabEncBench-Leaderboard	Benchmark leaderboard	Pairwise Bradley–Terry Elo tournament (2,000 bootstraps), radar chart comparisons, and win-rate matrices across 9 foundation encoders on Kabyle
Jugurtha-Chat	Conversational AI interface	Interactive chat interface powered by `Jugurtha-2-9B`, supporting native chain-of-thought (`<think>` blocks) and streaming generation

Datasets

Dataset	Domain	Description
KabEncBench	Multi-task evaluation	100,609 annotated instances across 6 core tasks (Syntax UPOS, Aspect Morphology, NER, Sentiment, SIB-200 Topic, Punctuation) and 3 diagnostic probes
KabBench	Evaluation benchmark	Audited and orthographically corrected Kabyle MT reference (correcting homoglyph corruption in 326 sentences), plus a balanced six-language Amazigh discrimination set
KabLex	Lexicon & pronunciation	366,892 permissive lexical entries over 17,090 lemmas with morphological features, and 25,642 word–pronunciation pairs in IPA
KabInflect	Morphology	336,151 inflected verb forms across 13,226 lemmas partitioned into lemma-disjoint splits, plus 6,198 complete verb conjugation paradigms
KabTifinagh	Transliteration	497,944 parallel sentences pairing vowel-less Neo-Tifinagh with canonical Kabyle Latin, plus English and French trilingual alignments
KabStandard	Orthography	497,944 parallel sentence pairs mapping informal keyboard and Arabizi text to canonical Kabyle Latin
KabPunct	Punctuation restoration	1,318,707 word-labelled sentences with per-token punctuation and casing labels
KabG2P	Phonetics & G2P	25,634 Kabyle word forms with validated IPA transcriptions across a 42-symbol phonetic inventory
KabSentiment	Sentiment	15,000 human-written Kabyle sentences balanced across three sentiment classes, labelled via cross-lingual teacher projection
KabLiterary	Literary corpus	28,907 paragraphs (657,713 words) of translated world literature and oral poetry in canonical Kabyle Latin with book-boundary splits

Technical Principles

Orthographic consistency: Public Kabyle text frequently exhibits systematic homoglyph substitutions (e.g. Greek ε U+03B5 for Latin ɛ U+025B, and γ for ɣ) in 2.6–3.2% of lines in raw web data. All AƔBALU datasets and models pass through a versioned 81-rule normaliser that standardises character representations while preserving contrastive spirantised and emphatic phonemes (ɣ, ɛ, ḥ, ḍ, ṣ, ṭ, ẓ, ṛ, č, ǧ, ţ).
Provenance tracking: Every sentence in the corpus carries traceable source identifiers, licensing metadata, and retrieval timestamps. The complete license distribution of training corpora is disclosed on each model card.
Audited benchmarks: Standard benchmarks with known upstream orthographic defects (such as the 16.2% homoglyph corruption in FLORES+ kab_Latn) are evaluated both in original and normalized forms to isolate model translation quality from benchmark encoding errors.
Decontamination controls: Training corpora are strictly decontaminated against benchmark evaluation splits (such as FLORES+ and SIB-200), verified through positive-control leak injection.
Standardized representation benchmarking: On KabEncBench, representation quality is evaluated with frozen backbones, layer-weighted ScalarMix pooling, and linear probes across identical hyperparameters, using 2,000 paired cluster-bootstrap tournament rounds to establish statistical significance.

The names

Each model is named in homage to a historical or cultural figure whose life and work reflect the function of that model.

Model	Task	Homage	Contribution
Tokenizer	Subword vocabulary	Mammeri	Mouloud Mammeri wrote the first Berber grammar written in Kabyle (1976), inventing the metalanguage to describe the language in its own words. The 1980 cancellation of his lecture on Kabyle poetry began the Berber Spring
Encoder	Bidirectional representation	Masinissa	First king of a united Numidia (r. 202–148 BCE), who brought eastern and western territories into a single state
Sentence embeddings	Exact passage retrieval	Si Mohand	Si Mohand ou Mhand (1848–1905), master of the Kabyle isefra, whose oral poetry survived intact through oral recitation and memory
Translation	Cross-lingual transfer	Amrouche	Taos Amrouche sang exclusively in Kabyle and wrote in French, bridging oral Kabyle traditions and literature
Transliteration	Script conversion	Juba	King Juba II (48 BCE – 23 CE), ancient Numidian scholar-king and author who studied and documented North African languages and scripts
Speech recognition	Speech-to-text	Fadhma	Fadhma Aït Mansour Amrouche, Taos's mother, who began transcribing ancestral Kabyle songs and oral histories in the 1930s
Punctuation & casing	Discourse structuring	Belaïd	Belaïd At Ali (1909–1950), author of Les Cahiers de Belaïd, the first continuous narrative prose written in Kabyle Latin orthography
Speech synthesis	Voice generation	Matoub	Lounès Matoub (1956–1998), celebrated singer and poet who recorded 36 albums in Kabyle, defending the language and culture throughout his life
Orthography standardisation	Text normalization	Boulifa	Si Amar ou Saïd Boulifa (1865–1931), author of Recueil de poésies kabyles (1904), who established early systematic conventions for Kabyle Latin orthography
Document OCR	Document recognition	Feraoun	Mouloud Feraoun (1913–1962), novelist and educator who transcribed oral poetry onto the printed page
Language identification	Sibling discrimination	Dihya	Queen Dihya, Amazigh historical leader from the Aurès region (c. 703 CE)
Generative foundation	Generative reasoning	Jugurtha	King Jugurtha of Numidia (c. 160–104 BCE), grandson of Masinissa, who defended Numidian independence

The model lineage reflects two historical and family relationships: Masinissa and Jugurtha represent the grandfather who unified Numidia and the grandson who defended its sovereignty, linking the foundational encoder and the generative model. Fadhma and Taos Amrouche are mother and daughter, reflecting the connection between transcribing spoken oral heritage into text (speech recognition) and translating across languages.

The naming is homage. It implies no endorsement by anyone, and none of these figures or their families are affiliated with this work.

Licence

Model weights and code are distributed under Apache-2.0 (unless otherwise noted, such as Matoub-4B's non-commercial research license). Training data licensing distributions are disclosed on each respective model and dataset card.