This powerful AI in Arabic with exceptional potential

Jais-Chat is an Arabic language model that manages to overshadow some of the big fish in the industry.

You’ve probably already heard of ChatGPT, but do you know its competitor Jais-Chat? Named after the name of a mountain located in the United Arab Emirates, this AI-powered conversational agent has established itself as the benchmark of its kind in the Arabic language.

Better than Llama 2 and Bloomz

This chatbot is the work of the American company Cerebras Systems, specializing in artificial intelligence, in collaboration with Inception, a subsidiary of the G42 investment group belonging to Abu Dhabi.

Jais-Chat impresses with its very above average performance. Its language model has managed to beat those considered leaders in the field in various tests, such as the multiple choice questionnaires of the University of California at Berkeley and the HellaSwag of the Allen Institute.

Jais-chat notably outperformed the Llama 2 linguistic model developed by Meta, popular among developers, because it is open source unlike OpenAI’s GPT-4 whose APIs are paid for. Another benchmark for open source language models, Bloomz also had to bow to Jais-Chat.

13 billion parameters

To achieve these results, Cerebras and Inception have chosen to limit Jais-Chat to two languages: English and Arabic. The dataset he trained on is 29% Arabic, 59% English, and 12% code.

With 13 billion parameters, the model is far from the 175 billion of GPT-3, but still manages to perform well thanks to a carefully selected database and the limitation to two languages.

“What was interesting was that Arabic also improved English”explains Andrew Feldman, co-founder and CEO of Cerebras, in an interview with ZDNET. “We ended up getting a model that performed as well as Llama in English, although we trained it on about a tenth of the data”he says.

English dominates the Web, and therefore language models

It is possible to write prompts in Jais-Chat in both English and Arabic, and the chatbot can also respond in these two languages. The user can, for example, write in English, but specify that they want a response in Arabic.

“We are giving 400 million Arabic speakers a voice in AI. This is what democratizing AI is all about. It is the main language of 25 countries »declares Andrew Feldman, who gently attacks other companies who talk about democratizing AI, but who all copy each other.

English is largely favored by generative AI. “The largest datasets rely on Internet scraping, and this is mostly in English”, regrets Andrew Feldman. A Meta study published in 2022 indicated that 63.7% of websites are in English, while only 25.9% of Internet users speak it.

To achieve better performance in other languages, language models will have to change strategy.

Source : ZDNet