Falcon-Emirati-7B shows dialect fidelity needs curated data, not just scale
lmoroney · x · 2026-10-07
Laurence Moroney analyzes TII's Falcon-Emirati-7B, built on Falcon-H1-Arabic (parallel Mamba SSM + attention layers) and further trained on crawled Emirati dialect text, MSA writing about Emirati culture, and glossary-constrained synthetic data. It scores 84.83% on the 1,173-question Alyah benchmark and hits 0.52 dialect fidelity in LLM-judged open-ended answers, versus ≤0.05 for four compared models that answered correctly but in Modern Standard Arabic. Evaluations are TII's own; a useful data point for adapting models to dialects or niches.
More from Models
- Speed benchmarks: System One models compared on clinical decision tasks — MaziyarPanahi · 2026-10-07
- Offline Smart Mirror runs Qwen 2.5 VL 7B and EmbeddingGemma 2 on Qualcomm NPU for outfit advice — HowDevelop · 2026-10-07
- Opus 5.5 builds 'black slopboxes' that beat human code on speed and quality despite ugly aesthetics — burny_tech · 2026-10-07
- Reddit User Spots Unreleased Gemini Flash Options Twice in Two Weeks — alovoids · 2026-10-07
- LLMs lack 'idempotency': rechecking their own large codebases yields contradictions — StephanSturges · 2026-10-07
- Liquid AI Set to Release New LFM Model Today — jacek2023 · 2026-10-07