Falcon-Emirati-7B release: Arabic LLM specialized for the Emirati dialect

TL;DR

Falcon-Emirati-7B is a 7 B parameter Arabic language model fine‑tuned to understand and generate Emirati Arabic, delivering 84.83 % accuracy on the Alyah dialect benchmark and outperforming larger multilingual models in both correctness and dialect fidelity.

Model Overview

Falcon-Emirati-7B builds on the Falcon‑H1‑Arabic 7 B checkpoint, which combines State‑Space Models (Mamba) and Transformer attention in each block. This hybrid architecture provides linear‑time efficiency for long sequences while preserving attention‑based precision, a crucial property for morphologically rich Arabic. The base family spans 3 B, 7 B, and 34 B parameters with context windows up to 256 K tokens and training on a mix of Modern Standard Arabic (MSA), Gulf dialects, and multilingual data.

Why Dialect Adaptation Is Hard

  • Sparse written data – Emirati Arabic is primarily spoken, so online text is scarce compared with MSA.
  • Non‑literal meaning – Idioms, proverbs, and nabati poetry rely on cultural context that literal translation misses.
  • No established recipe – The optimal proportion of dialectal versus MSA data and the best training stage (continued pre‑training, SFT, or preference optimization) are unknown, requiring extensive trial‑and‑error.

Data Pipeline

  1. Authentic Emirati‑dialect web data – Curated crawls from Emirati forums and websites that contain native‑written dialect, capturing everyday phrasing and code‑switching with MSA.
  2. MSA material about Emirati culture – Articles on heritage, customs, and social norms that provide background knowledge without teaching dialectal form.
  3. Synthetic data guided by glossaries – Large‑scale generated text constrained by Emirati‑specific vocabularies and style rules to ensure authenticity across topics not covered by the crawled data.

Training Strategy

A series of ablations explored:

  • The proportion of dialectal versus synthetic data.
  • The injection point of dialectal data (early continued pre‑training vs. later instruction tuning).
  • Balancing cultural MSA content to maintain grounding while avoiding over‑fitting to synthetic patterns. Human native‑speaker review and automatic metrics guided the selection of the final recipe.

Evaluation Methodology

Manual Native‑Speaker Review

Native Emirati speakers rated outputs for naturalness, tone, and cultural appropriateness, capturing nuances that automatic scores miss.

Automatic Benchmark – Alyah

Alyah (الياه, “North Star”) is a 1,173‑sample, multiple‑choice benchmark created by the authors and the community to test Emirati‑dialect competence. It covers greetings, idioms, poetry, heritage knowledge, and language‑specific categories.

Results

  • Alyah accuracy: 84.83 % – the highest among all compared Arabic and multilingual models, including several larger models.
  • Dialect fidelity (LLM‑as‑judge): Partial‑credit score of 0.52, two orders of magnitude higher than the next best model (0.05).
  • Open‑ended generation: Falcon‑Emirati‑7B leads in correctness and consistently answers in Emirati Arabic, whereas competing models default to MSA.
  • Cultural understanding (ArabCulture‑Dialogue UAE subset): 85.57 % accuracy, outperforming ALLaM‑7B (83.39 %) and other baselines.

Detailed Comparisons

  • Multiple‑choice vs. other models – Size alone does not guarantee dialect competence; targeted dialect training is essential.
  • LLM‑as‑judge generation – Falcon‑Emirati‑7B is the only model that reliably switches to Emirati when prompted, with a win‑rate of 0.69–0.88 in poetry and cultural categories.
  • Category‑level analysis – The model excels across all Alyah categories; the only close competition appears in greetings, where Emirati and MSA overlap.

Implications

The results demonstrate that dialect‑specific fine‑tuning can yield substantial gains over scaling alone, especially for languages with rich oral traditions and cultural nuance. Falcon‑Emirati‑7B provides a practical tool for applications requiring authentic Emirati interaction—customer support, cultural content generation, and localized AI assistants—while highlighting the need for dedicated data pipelines and evaluation benchmarks for each dialect.

Responsible AI and Limitations

The model may inherit biases from its training data and can err on rare expressions or highly localized references. Dialectal nuance is sometimes subjective; native speakers may disagree on the “correct” answer. Users should evaluate the model for their specific use case and provide feedback to improve both the model and the Alyah benchmark.

How to Try Falcon‑Emirati‑7B

The model is accessible via the Falcon chat platform: https://chat.falconllm.tii.ae/?model=Falcon-Emirati-7B.

Citation

@misc{falcon_emirati_7b_2026,
  title = {Falcon-Emirati-7B: A Dialect-Specialized Arabic LLM for the Emirati Dialect},
  author = {Shaikha Alsuwaidi, Omar Alkaabi, Maitha Alhammadi, Hamza Alobeidli, Ahmed Alzubaidi, Mohammed Alyafeai, Leen AlQadi, Basma Boussaha, Hakim Hacid},
  organization = {Technology Innovation Institute},
  year = {2026}
}

Sources