The Falcon team has released Falcon-Emirati-7B, a dialect-specialized large language model built on the Falcon-H1-Arabic foundation to understand and generate Emirati Arabic with native-level nuance. The model addresses the gap between Modern Standard Arabic and the specific vocabulary, tone, and cultural context of the UAE's Gulf dialect.
- Built on the 7B variant of Falcon-H1-Arabic using a hybrid Mamba-Transformer architecture for efficient long-context processing.
- Trained on a curated mix of authentic Emirati web data, MSA content regarding Emirati culture, and synthetic data constrained by specific glossaries.
- Achieved an 84.83% accuracy score on the Alyah benchmark, outperforming larger multilingual models.
- Demonstrated superior dialect fidelity in open-ended generation, scoring 0.52 compared to near-zero scores for competitors like ALLaM and Fanar.
The release highlights that targeted dialect adaptation is necessary for cultural competence, as general Arabic models often default to Modern Standard Arabic even when asked to respond in the local dialect.