Latest News
By working with large language models (LLM) like GPT of OpenAI, Huggingface, LlaMA we discover challenges, roadblocks as well as the ways to overcome them. We are happy to share our findings this blog.
GPTLabs tested Jais – open source arabic LLM created in Abu Dhabi
One of our UAE clients has requested implementation of arabic LLM JAIS, that is claimed to be the best model for the Arabic language. The model was developed by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) This is a model with 13 billion parameters and is fine-tuned over a curated set of 4 million Arabic and 6 million English prompt-response pairs.
More details regarding performance can be found at their hugging face page https://huggingface.co/inception-mbzuai/jais-13b-chat, but here we would like to focus on practical implementation and our observations.
First of all the model is quite heavy (50GB), and it was missing some configuration for deployment on Hugging Face inference endpoints. Therefore, we have created a copy of that model with necessary configuration. Please, check it out here: https://huggingface.co/poiccard/jais-13b-chat-adn.
The main trick was to add a handler.py file to enable the deploy function. Please read the instructions of our model carefully. It requires a big machine, i.e., GPU [large] · 4x Nvidia Tesla T4, which costs $4.50 per hour. Small and medium-sized machines were not able to start it up.
Once deployed and running, you will be able to use it in your application as a REST endpoint with a sample curl:
curl https://YOUR_ENDPOINT.aws.endpoints.huggingface.cloud -X POST -d '{"inputs": "ما هي عاصمة الامارات؟"}' -H "Authorization: Bearer YOUR_TOKEN" -H "Content-Type: application/json"
We have asked our Arabic-speaking colleagues to test Jais, and so far, we have received mixed responses. According to them, the GPT 3.5 Turbo is responding better to questions in Arabic.
Feel free to test it yourself and let us know what you think. Please do not forget to pause your deployment when you are not using it; otherwise, prepare to burn some cash.
As the next step, we are planning to try the Jais LLM model on information retrieval tasks from Arabic corporate documents and policies. Stay tuned!
