Publications
Research
Peer-reviewed papers, conference presentations, and technical reports.
Moroccan Web Observatory · September 2026
State of the Moroccan Web 2026
A measured view of connections, technology, hosting, language and content across the Moroccan web, built from the Omneity Maktaba index.
Explore the observatory →From the lab
Read the work behind the stack.
Tokenization is Killing our Multilingual LLM Dream
Why language representation changes what a model can learn.
Read on Hugging Face ↗Data and identificationGherbal and multilingual FineWeb
Identifying languages and improving Moroccan Arabic web data.
Read on Hugging Face ↗Open datasetsWikipedia Monthly
Fresh multilingual article data for research and model development.
Read on Hugging Face ↗Sawtone: A Universal Framework for Phonetic Similarity and Alignment Across Languages and Scripts
Lingua Posnaniensis, Vol. 67(1)
Introduces a cross-script phonetic alignment framework with modular language-specific adapters. Demonstrates 88% BLEU transliteration and 87–95% phonetic alignment accuracy across language/script pairs. Includes a case study on preprocessing Moroccan Arabic (Darija) for LLM training.
GenAI for Moroccan Darija: Challenges and Early Results
University of Navarra, Spain
Conference presentation at the 7th International Congress for Moroccan Arabic, discussing challenges and early results in applying generative AI to Moroccan Darija.
Gherbal: A Multilingual Classifier for Low-Resource Languages
University Hassan II, Casablanca, Morocco
Conference presentation at TIM'24, introducing Gherbal — a multilingual classifier designed for low-resource languages.
Gherbal v4: Comprehensive Evaluation Report — Language Identification Across 214 Languages
Omneity Labs
Comprehensive evaluation of Gherbal v4 across 10 models, 8 benchmarks, and 5 scoping regimes. Gherbal v4 achieves 0.836 average accuracy — outperforming OpenLID v2, GlotLID, and NLLB-LID. The report covers Arabic dialect identification, Arabizi detection, African language coverage, and model efficiency analysis.
Madmon: Specialized Multilingual Embeddings for Underrepresented Languages
Omneity Labs
Madmon is a multilingual text embeddings model specialized for Arabic dialects, Amazigh languages, and other underrepresented language families. Designed for retrieval, classification, clustering, and semantic search in multilingual and low-resource settings.
Sawtone v4
Omneity Labs
Comparative technical report on Sawtone v4 as a form-aware representation layer for multilingual and multiscript NLP. It evaluates whole-word geometry, contextual representations, and candidate-ranked transliteration across Moroccan Arabic and additional writing systems.