---
title: "Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance"
description: "Falcon-Emirati-7B is a dialect-specialized language model built on Falcon-H1-Arabic to understand and generate Emirati Arabic, a Gulf dialect with distinct vocabulary and cultural context that Modern Standard Arabic models miss. The model uses a hybrid architecture combining Mamba state-space models and Transformer attention running in parallel inside every block, with fused outputs for efficient long-context processing. Training required a dedicated data pipeline of authentic Emirati web content, MSA material about Emirati culture, and synthetic data guided by strict glossaries and style rules to cover gaps in available dialectal text. On the Alyah benchmark, Falcon-Emirati-7B scored 84.83%, outperforming larger multilingual models, and achieved 85.57% on the UAE cultural understanding task in ArabCulture-Dialogue. The 7B parameter size was chosen as the best balance of quality against training and inference cost for dialect adaptation."
---

# Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance

[Hugging Face](https://yomu.fyi/company/hugging-face) · Shaikha Alsuwaidi, Omar saif alkaabi, Maitha Alhammadi, Ahmed Alzubaidi, Mohammed Alyafeai, Leen AlQadi, Basma Boussaha, Hakim Hacid · Oct 6, 2026

**Type:** Announcement

## Summary

Falcon-Emirati-7B is a dialect-specialized language model built on Falcon-H1-Arabic to understand and generate Emirati Arabic, a Gulf dialect with distinct vocabulary and cultural context that Modern Standard Arabic models miss. The model uses a hybrid architecture combining Mamba state-space models and Transformer attention running in parallel inside every block, with fused outputs for efficient long-context processing. Training required a dedicated data pipeline of authentic Emirati web content, MSA material about Emirati culture, and synthetic data guided by strict glossaries and style rules to cover gaps in available dialectal text. On the Alyah benchmark, Falcon-Emirati-7B scored 84.83%, outperforming larger multilingual models, and achieved 85.57% on the UAE cultural understanding task in ArabCulture-Dialogue. The 7B parameter size was chosen as the best balance of quality against training and inference cost for dialect adaptation.

## Context

Modern Standard Arabic (MSA) is the standard written form, but daily conversation in the UAE uses Emirati Arabic, a Gulf dialect with its own vocabulary, rhythm, and cultural context. General Arabic models trained on MSA fail to capture the nuance, idioms, and cultural references that native speakers understand. Falcon-Emirati-7B aims to close that gap by specializing a base Arabic model for the Emirati dialect.

## Approach / What changed

Falcon-Emirati-7B is built on the 7B variant of Falcon-H1-Arabic, which uses a hybrid Mamba-Transformer architecture. The adaptation used a three-source data pipeline: authentic Emirati web content, MSA material about Emirati culture, and synthetic data generated with strict glossaries and style rules. Training strategy was determined experimentally through ablations on data mix and training stages, evaluated by native speakers and the Alyah benchmark.

## Takeaways

- Falcon-Emirati-7B scores 84.83% on the Alyah benchmark, outperforming larger multilingual models, showing that size alone does not guarantee dialect competence.
- The model achieved 85.57% on the UAE cultural understanding task in ArabCulture-Dialogue, demonstrating its ability to recognize culturally appropriate responses.
- The 7B parameter size was chosen as the sweet spot for dialect adaptation, balancing quality against training and inference cost, as the 34B model would be too expensive and the 3B model lacks headroom.

## FAQ

### How was Falcon-Emirati-7B trained?

It was built on Falcon-H1-Arabic 7B using a dedicated data pipeline of authentic Emirati web content, MSA cultural material, and synthetic data guided by strict glossaries. Training involved experimental ablations on data mix and stages, with native speaker review and automatic evaluation on the Alyah benchmark.

### What makes Emirati dialect adaptation difficult?

Emirati is mostly spoken with limited written data, meaning is often non-literal with idioms and proverbs, and there is no established recipe for how much dialectal data is needed or which training stage matters most. The team relied on trial and error to find the right approach.

### How does Falcon-Emirati-7B compare to larger models?

On the Alyah benchmark, Falcon-Emirati-7B scored 84.83%, ahead of every other Arabic and multilingual model tested, including models many times its size. It also outperformed ALLaM-7B, Jais-2-8B, and Fanar-2-27B on the UAE cultural understanding task.

**Tags:** [Arabic](https://yomu.fyi/topic/arabic), [Benchmark](https://yomu.fyi/topic/benchmark), [LLMs](https://yomu.fyi/topic/llm)

- Source: [Hugging Face](https://huggingface.co/blog/tiiuae/falcon-emirati)
- Source URL: https://huggingface.co/blog/tiiuae/falcon-emirati
- Ingested by Yomu: 2026-10-06T09:00:07.990Z

[Read original post](https://huggingface.co/blog/tiiuae/falcon-emirati)