Principal Researcher · Technology Innovation Institute
I lead the Arabic & Dialects team at TII, building the next generation of Arabic foundation models: large language models, dialect-specialized models, Arabic OCR, and the benchmarks used to evaluate them.
01 — Selected work
Models my team and I have designed, trained and released as part of TII's Falcon family.
State-of-the-art Arabic LLMs on a hybrid Mamba-Transformer architecture, with attention and SSM running in parallel inside each block. Available in 3B, 7B and 34B, with context windows up to 256K tokens.
A 7B model adapted from Falcon3-7B. We extended the tokenizer with 32K Arabic tokens and initialized their embeddings from textually similar existing tokens. It outperformed Arabic models up to 4× its size.
A model that actually answers in Emirati dialect instead of falling back to MSA. It scores 84.8% on Alyah and 85.6% on UAE cultural dialogue scenarios.
A 270M early-fusion OCR model trained with SFT and then RL. It ranks #2 of 17 on 12K real Arabic documents, has the best table score (TEDS 59.95%), and is #1 on forms, receipts and official documents.
Benchmarks and leaderboards that the Arabic NLP community uses to compare models.
A leaderboard that validates benchmark quality first: 109 subsets from 14 benchmarks, 52K+ samples, 46 models.
The community reference for Arabic LLMs, built with Hugging Face. Version 1 launched in 2024 and version 2 in 2025.
3,609 Arabic STEM and code tasks: native exam questions, synthetic reasoning, and Arabic HumanEval+ / MBPP+.
1,173 questions written by native Emirati speakers, covering dialect, poetry, heritage, etiquette and figurative meaning.
A NeurIPS competition on benchmarks that give useful signal early in LLM pre-training.
A map of Arabic benchmarks and evaluation methods, and of the gaps that remain.
02 — The team
I lead a team of 6–7 researchers and engineers in Abu Dhabi. We own Arabic across the Falcon family, from pre-training data and tokenization to post-training, dialect adaptation, multimodal document understanding and evaluation. Our work is a major contribution to the Arabic NLP community and to AI more broadly.
Continued pre-training, synthetic data generation, vocabulary extension and hybrid architectures for Arabic.
Emirati and Gulf dialects, culturally grounded data, and native-speaker annotation (Hamsa, Alyah).
Cursive script, diacritics, right-to-left layouts, tables and forms.
Benchmarks that are checked for quality first, plus public leaderboards.
03 — News
Released Falcon-OCR-Arabic, a 270M model ranked #2 of 17 on Arabic documents.
Released Falcon-Emirati-7B, a model specialized in Emirati dialect and culture.
Hamsa, a manually annotated Emirati Arabic corpus, accepted at Interspeech 2026.
Launched QIMMA قِمّة, a quality-first Arabic LLM leaderboard.
Released Alyah, a benchmark of Emirati dialect capabilities.
Released Falcon-H1-Arabic (3B / 7B / 34B), state of the art on OALL.
Two papers accepted to the Third Arabic NLP Conference: 3LM, and adapting Falcon3-7B to Arabic.
Co-organized the E2LM competition at NeurIPS 2025.
Released Falcon-Arabic, the first Arabic model in the Falcon series.
Launched the Open Arabic LLM Leaderboard v2 with Hugging Face.
Launched the Open Arabic LLM Leaderboard.
04 — Publications
The full, up-to-date list is on Google Scholar.
05 — Experience
Abu Dhabi, UAE
Principal Researcher · 2025 — now
Lead Researcher · 2023 — 2025 (joined as Senior Researcher)
I lead the Arabic & Dialects team within the Falcon LLM program (Falcon-Arabic, Falcon-H1-Arabic, Falcon-Emirati, Falcon-OCR-Arabic) and contributed to Falcon2-11B. We train on thousands of GPUs across AWS, GCP, Nebius and on-premise clusters, orchestrated with Kubernetes and Slurm.
Paris, France
Senior Data Scientist, building question answering, document clustering and triage models for legal contracts. I migrated our models to Hugging Face and BERT, automated training with Airflow, and deployed on Kubernetes with Terraform and CI/CD.
Nantes, France
PhD researcher in neural conversational systems, focused on response selection for retrieval-based dialogue. Supervised by Emmanuel Morin, Christine Jacquin and Nicolas Hernandez.
Paris, France
Research intern.
06 — Contact
I'm open to research collaborations, talks and partnerships on Arabic and multilingual AI.
Send me an email