Saudi Arabia's Allam LLM now available on Microsoft Azure
Saudi Arabia's Allam LLM now available on Microsoft Azure
SDAIA's LLM will be available to Microsoft Azure users locally and globally
#Saudi #LLMs - Microsoft has announced general availability on Microsoft Azure of the ALLaM Arabic large language model developed by Saudi Data and Artificial Intelligence Authority (SDAIA). According to Turki Badhris, president of Microsoft Arabia, the model, which was developed and trained on Azure, is available to Azure customers in Saudi Arabia and globally. SDAIA built 7 billion, 13 billion and 70 billion parameter versions of the ALLaM LLM, but as yet, only the ALLaM 7B is available via the Azure platform.
SO WHAT? - Saudi Data and Artificial Intelligence Authority (SDAIA) launched an early version of ALLaM chat in May 2023 and made it publically available for a limited time period. In the meantime, SDAIA has developed a suite of models, releasing ALLaM 13b on IBM’s Watson-x platform in May 2024. However, although it was made available under a royalty-free SDAI licence, access was nevertheless limited to Watson-x users. The release of ALLaM on Microsoft Azure broadens the model’s availability considerably, allowing more usage, development and review of one of the Arab world’s most important AI models.
Here are some key details about ALLaM:
Turki Badhris, president of Microsoft Arabia, announced the general availability of the ALLaM Arabic-centric large language model on Microsoft Azure on stage at the Global AI Summit, together with Dr. Esam O. Al-Wagait director of the National Information Center – Saudi Arabia.
First announced in May 2023, ALLaM has been limited in availability, only making it onto commercial cloud platforms this year, beginning with IBM’s Watson-x in May.
The ALLaM large language model was developed by the National Center for Artificial Intelligence (NCAI) at Saudi Data and AI Authority (SDAIA).
SDAIA mobilised 16 public entities to aggregate and contribute the data used to train ALLaM, creating a 500 billion token Arabic data set, the world’s largest Arabic language training data set. The training data included the text of 300 Arabic language books.
The authority also drew together 400 subject matter experts to query and test the ALLaM model, resulting in more than 1 million prompts.
SDAIA built 7B, 13B and 70B sizes of the ALLaM LLM version 2, although it seems that only the ALLaM 7B is available via Microsoft Azure.
According to SDAIA, the ALLaM model outperforms all other LLMs in Arabic language queries using MMLU (Massive Multitask Language Understanding) automatic benchmarking tests.
The model, which was developed and trained on Azure, is available to Azure customers in Saudi Arabia and globally. SDAIA built 7 billion, 13 billion and 70 billion parameter versions of the ALLaM LLM, which is based on Meta’s Llama 2.
Version 1 of ALLaM 13B LLM (ALLaM-1-13b-instruct) was made available on IBM’s watson-x starting May 2024.
Read more about ALLaM LLM:
SDAIA's Arabic LLM now live on watsonx (Middle East AI News)