Cerebras welcomes a guest post from Inception CEO Andy Jackson about our recent work together on Jais, the world’s best-performing open-source Arabic model The post details the model architecture, data, hardware, and performance evaluations Read here: https://
cerebras.net/blog/jais-a-ne
w-pinnacle-in-open-arabic-nlp
…
@cerebras
-
Jais: Open-Source Arabic NLP Model Performance Breakthrough
By
–
-
Jais Initiative Brings AI Solutions to Arabic-Speaking World
By
–
“While many companies discuss the concept of 'democratizing AI,' the Jais initiative moves beyond rhetoric by providing a substantive, data-driven solution for the Arab-speaking world.” Read the full @towards_AI article here:
-
Condor Galaxy 1 Powers BTLM and Jais Open-Source LLMs
By
–
Condor Galaxy 1 (CG-1) was used to train BTLM, the top-performing open-source 3B parameter LLM. Now, CG-1 has been used to train Jais, the best open-source Arabic model. CG-1 is moving the open-source community forward. Contact us to train on CG-1 http://
cerebras.net/contact-us -
JAIS Arabic Language Model Released with Paper and Demo
By
–
Paper, sandbox, and model here: https://
arxiv.org/abs/2308.16149
#new_tab
… https://
arabic-gpt.ai https://
huggingface.co/inception-mbzu
ai/jais-13ba
… -
Jais Open Source Model Achieves Claude-Level Performance at Fraction Size
By
–
Compared to closed source models, Jais approaches Claude level performance. This is remarkable considering it’s ~10th the size of large closed models.
-
Cerebras-GPT: First External Customer AI Model Achievement
By
–
Jais was trained on the CG-1 AI supercomputer. We are super proud of the work we did on Cerebras-GPT and BTLM. This is the first time an external customer used our system to produce a world class model – another huge milestone for Cerebras!
-
Jais Multilingual Model Outperforms Bloom and mT0
By
–
In downstream evals, Jais easily tops the charts, beating prior open source models like Bloom and mT0. It’s easier to be good at two languages than 50!
-
Building Arabic Language Models: Data Collection and Training Strategy
By
–
The key challenge of training any non-English model is data. To build a dataset large & rich enough for Arabic, Jais tapped into various open and curated sources totally 55B tokens. The model trained on 2 epochs of this data.
-
Jais Model Combines Arabic and English Capabilities via Transfer Learning
By
–
In addition, Jais is trained on 278M English. This not only adds English capability, it also improves Arabic performance thanks to transfer learning. Knowing both languages makes the model super useful for locals and expats in the Middle East.
-
Jais-13B Bilingual Model Breakthrough Arabic English LLM
By
–
The Jais-13B model and paper is out! Jais is the world’s best *bilingual* model – fluent in Arabic & English. Here it’s casually negotiating for a lower rent. Let’s breakdown the paper