#UAE #LLMs - Technology Innovation Institute (TII), the applied research arm of Abu Dhabi’s Advanced Technology Research Council (ATRC), has announced three Arabic AI models on the first day of Ai Everything Abu Dhabi. The new model sare Falcon-Emirati, Falcon-ASR and Falcon-OCR-Arabic. Falcon-Emirati is a 7-billion-parameter language model built for the Emirati Arabic dialect. It scored 84.83 percent on the Alyah benchmark, ahead of every Arabic and multilingual open-source model tested. The release covers text, speech and document understanding.
SO WHAT? - Many large language models can understand every word of an Emirati Arabic sentence and still miss its meaning. This is due to the fact that most models learn from Modern Standard Arabic (MSA), a version of formal Arabic. In consequence, other language forms such as poetry, proverbs and everyday confuse AI models and lead to misunderstand. TII has built and trained Falcon-Emirati for how UAE citizens actually talk. The 1.6 billion parameter speech model outperforms much larger ones can run cheaply, which matters for subtitling, customer service and public services.
KEY POINTS:
Technology Innovation Institute (TII) has announced three models on the first day of Ai Everything Abu Dhabi: Falcon-Emirati for language, Falcon-ASR for speech transcription and Falcon-OCR-Arabic. All three aim to bring Emirati language and multimodality within reach of developers and researchers.
The 7 billion parameter Falcon-Emirati is built on Falcon-H1-Arabic. TII trained it on native Emirati content, Modern Standard Arabic material on Emirati culture and heritage, and synthetic data informed by Emirati-specific linguistic resources.
The model scored 84.83% on Alyah, a native Emirati Arabic benchmark covering everyday language, figurative expressions, heritage knowledge and poetry. TII says that beats every Arabic and multilingual open-source model it evaluated. It will be available through a Falcon Chat platform.
Falcon-ASR has 1.6 billion parameters. It converts spoken Standard and Emirati Arabic, English, French, Spanish and Portuguese into text, and it beat larger systems in TII’s evaluation, including a 30-billion-parameter multimodal model on Emirati speech.
The Falcon-ASR model also timestamps each word in an audio file ,supporting accurate subtitling, searchable recordings, meeting and interview transcription, accessibility tools and other voice-enabled services.
Falcon-OCR-Arabic is a lightweight model that extracts Arabic text and structured content from images and documents. It can CCR a wide variety of content including tables, mathematical formulas and document sections. The model uses include digitising documents, building searchable archives and automating information extraction.
TII says Emirati Arabic remains underrepresented in AI training data. Its vocabulary, idioms and cultural references, particularly in poetry, proverbs and anecdotes, can’t always be understood through literal translation.
TII is Abu Dhabi’s applied research centre, with nine research centres spanning AI and digital science, quantum, cryptography and robotics. The launch adds to its sovereign AI work across language and modalities.
[Written and edited with the assistance of AI]
Source: TII
Read more about TII’s Falcon models:
Falcon H1 tops new global sovereign AI index (Middle East AI News)
Abu Dhabi’s TII launches Falcon models that see and read (Middle East AI News)
TII releases 15 tiny Falcon language models (Middle East AI News)
TII announces new Arabic AI model family (Middle East AI News)
TII releases compact reasoning model Falcon-H1R (Middle East AI News)