Jais
About
The Jais family of large language models (LLMs) is a groundbreaking development in the realm of Arabic natural language processing (NLP), created collaboratively by Inception and Cerebras Systems. This family consists of bilingual English-Arabic LLMs designed for exceptional performance in Arabic, while retaining robust English capabilities. Spanning models from 590 million to a remarkable 70 billion parameters, they address various computational needs. Trained on a vast dataset comprising Arabic, English, and code data amounting to 1.6 trillion tokens, the Jais models come in two variants: pre-trained from scratch or adaptively pre-trained from Llama-2. Both types are instruction-fine-tuned for dialogue, rendering them ideal for applications like chatbots and other conversational AI systems. By aiming to accelerate Arabic NLP research, these models offer immense potential for Arabic-speaking and bilingual communities, furnishing numerous downstream applications. 1 2 3