Falcon-Emirati-7B answers in Emirati dialect where larger models default to formal Arabic

Modern Standard Arabic is the Arabic of news and textbooks, but in the UAE daily life happens in Emirati Arabic, a Gulf dialect with its own vocabulary, idioms and poetry. The Falcon team has released Falcon-Emirati-7B, built on its Falcon-H1-Arabic family, to understand and answer in that dialect as a native speaker would. Falcon-H1-Arabic pairs Mamba-style state-space layers with transformer attention in each block, comes in 3B, 7B and 34B sizes and was trained on formal and dialectal Arabic; the team chose 7B as the best balance of quality and cost.

The headline result concerns register. On the team’s open-ended tests, Falcon-Emirati scored 0.52 on dialect fidelity, meaning how often it actually answered in Emirati, while competing models scored roughly 0.00 to 0.05. Those models often knew the answer but replied in formal Arabic.

A model that answers a grandparent’s question in textbook Arabic is not exactly wrong, but it is not speaking to them. The team found that some of the largest multilingual models scored below smaller dialect-aware ones, a corrective to the assumption that bigger is better for every community. The effort fits a wider push for AI in more languages (every language).

The method acknowledges the difficulty. Emirati is mostly spoken, so written text is scarce. The team combined dialect text crawled from Emirati sites and forums, formal-Arabic material about Emirati culture, and a large volume of synthetic dialect data constrained by glossaries and style rules. It describes much of the process as trial and error, which is typical for low-resource dialect adaptation.

The team built the Alyah benchmark with the community, where the model scores 84.83%, and on a separate cultural-appropriateness test of 283 UAE scenarios it scored 85.57%, ahead of three other models. Open-ended scoring used another company’s model as judge, which is standard but, like home-built benchmarks, benefits from outside checks. The post notes limits on rare expressions, advises caution for sensitive or official uses and asks Emirati speakers for feedback.

Generating large volumes of dialect text with a model, even under strict rules, can quietly standardize a living dialect around what the generator thinks it sounds like. If dialect models built this way spread, the communities they serve need a real role in testing them. Independent native-speaker evaluations and similar models for other Arabic varieties would show whether the recipe generalizes. More in our research coverage, generative AI coverage and AI coverage.