abderahmane-ai/agbalu

Kabyle (Taqbaylit) NLP built from the corpus up — orthographic normalisation, parallel and speech corpora, a benchmark suite, and six models on the Hugging Face Hub: encoder, translation, ASR, punctuation restoration, script conversion, tokenizer. # AƔBALU *Aɣbalu* — Kabyle for **the source, the fountainhead**. Natural language processing for Kabyle (Taqbaylit, ISO 639-3 `kab`, BCP-47 `kab-Latn`), a Northern Berber language of Kabylia, Algeria with roughly 5–7 million speakers. The corpus is the primary artifact; the models are built on top of it. ## Why Kabyle has an unusual resource profile: it is **speech-rich and text-poor**. Common Voice v26.0 holds 571.29 validated hours of it — 10th of 294 locales — while the entire filtered Kabyle web crawl is 13.8M words and the whole deduplicated text corpus is 34.9M. The text that exists is also damaged in a specific, measurable way. Both seed corpora carry systematic homoglyph corruption: Greek `ε` U+03B5 standing in for Latin `ɛ` U+025B in 2.60% and 3.19% of rows. Kabyle legitimately uses `ɣ ɛ ḥ ḍ ṣ ṭ ẓ ṛ č ǧ ţ`, and the emphatic and spirantised distinctions carry meaning, so the repair has to be a versioned normalisation layer rather than a character filter. ## Published artifacts Twenty repositories on the Hugging Face Hub under `agbalu`, twelve models and eight datasets, plus the organization Space. Every model loads with `transformers` and `torch` alone, without this repository. | Model | | | |---|---|---| | `Masinissa-31M` | encoder | 90.51% on gold POS, against a most-frequent-tag baseline of 83.42% | | `SiMohand-278M` | sentence embeddings and retrieval | 97.0% Recall@1 against the backbone's 63.8, and the same 97.0 at a 12× compression | | `Amrouche-1.3B` | translation | beats NLLB-1.3B in all four directions; eng→kab 36.34 chrF++ | | `Fadhma-300M` | speech recognition | CER 8.01 / WER 25.65 over 888 unseen speakers | | `Belaid-31M` | punctuation and casing | macro-F1 0.793 | | `Boulifa-48M` | orthography standardisation | informal → canonical Kabyle Latin; 97.39% character accuracy against 89.70% for leaving the input alone | | `Feraoun-36M` | document OCR, Latin and Tifinagh | CER 2.85% and 70.20% line exact match over 1,000 held-out lines | | ` …