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

What it claims

Falcon-Emirati-7B is TII's Emirati-dialect fine-tune of the 7B Falcon-H1-Arabic model. Its main differentiator is register: in open-ended generation it answers in Emirati dialect, while comparable Arabic models answer correctly but in Modern Standard Arabic (MSA). On LLM-judged dialect fidelity it scores 0.52 (partial credit). The four competitors score 0.05 (ALLaM-7B-Instruct-preview), 0.03 (gemma-3-27b-it), 0.02 (Jais-2-8B-Chat) and about 0.00 (Fanar-2-27B-Instruct).

How it works

• Base model: Falcon-H1-Arabic 7B. Its blocks run Mamba (SSM) and attention in parallel and fuse the outputs. The family comes in 3B, 7B and 34B sizes with 128K–256K context, and was already pretrained on MSA, several Arabic dialects, English and multilingual data. TII chose 7B because 34B's training and serving cost wasn't justified for a dialect chat model, and 3B lacked "headroom".
• Data, three sources: 1. Crawled native Emirati websites and forums. 2. MSA text about Emirati culture, heritage and stereotypes, used for grounding rather than dialect. 3. Synthetic dialect data whose generator was constrained by Emirati glossaries, dictionaries and style rules.
• Ablations: TII varied dialect-data volume and the training stage it was injected at, the ratio of authentic to synthetic data (to avoid overfitting to synthetic patterns), and the amount of MSA cultural text. The page reports none of the ablation outcomes, and it does not say which stage (continued pretraining, SFT or preference optimization) the final recipe uses.
• Model selection: native-speaker review plus the Alyah benchmark.

Evidence

• Alyah: 1,173 manually collected, native-speaker multiple-choice items. Falcon-Emirati-7B scores 84.83%, which TII says is above all compared instruction-tuned models. Competitor numbers appear only in a chart. Falcon-H1-Arabic models are excluded from the comparison.
• Open-ended generation: the same 1,173 questions, judged by Gemini 3.7 Flash on correctness and dialect separately.
• Falcon leads on correctness.
• Fanar-2-27B scores lowest on correctness (0.27) and abstains 26.2% of the time; every other model abstains under 5%.
• The dialect-fidelity lead holds in every Alyah category.
• Blind pairwise wins (Gemini 3.7 Flash judge):
• vs Jais-2-8B: Poetry 0.69, Language & Dialect 0.62; loses Greetings 0.46–0.54.
• vs ALLaM-7B: Poetry 0.66, Language & Dialect 0.58; ties Greetings 0.50.
• vs Fanar-2-27B: wins every category, including Poetry 0.88, Religious & Social Sensitivity 0.80 and Greetings 0.70.
• ArabCulture-Dialogue, UAE split: 283 scenarios, three-option choice of the most culturally appropriate reply, tested in both Emirati and MSA with varying location information, averaged. Results: Falcon-Emirati 85.57%, ALLaM-7B 83.39%, Jais-2-8B 73.79%, Fanar-2-27B 71.50%. The 2.2-point margin over ALLaM is small, and these are TII's own runs.

How to use it

The only access route the page gives is TII's chat platform: https://chat.falconllm.tii.ae/?model=Falcon-Emirati-7B. It mentions no downloadable weights, licence, API or inference requirements.

Limits

• Self-built benchmark: Alyah comes from the same team ("we and the community").
• No base-model comparison: the base Falcon-H1-Arabic is excluded from comparisons, so the gain from dialect fine-tuning over its own base is never shown.
• LLM-only judging: all open-ended and pairwise judging uses a single LLM judge. The page reports no agreement with native speakers.
• Dialect fidelity is partial: 0.52 means the model still often does not answer in dialect.
• Greetings are weak: this is where Emirati and MSA overlap most, and the model loses or ties against Jais and ALLaM there.
• Stated failure modes: rare expressions, highly localized references, and data-sparse edge cases. TII also notes that native speakers can disagree on what is correct. It advises use-case evaluation before any sensitive, official or high-stakes use.
• No ablation results, and no training compute or cost figures.