June 17th, 2024
SUTRA is a novel multilingual large language model architecture that is trained by decoupling concept learning from language learning. It operates in 50+ languages and achieves state-of-the-art performance across numerous multilingual benchmarks.
The name "SUTRA" is a Sanskrit term meaning a thread or aphorism. It reflects how we weave together concise wisdom across multiple languages.
Most generative AI models are trained on English, neglecting global linguistic diversity and limiting access to billions of non-English speakers.
https://media.licdn.com/dms/image/D5612AQH4eiSAxyxJDg/article-inline_image-shrink_1000_1488/0/1715641851285?e=1721865600&v=beta&t=iWPOXxk3pLwISFYlBgdCs3IRmh1u8nbIa_sTzolPsEg
The Massive Multitask Language Understanding (MMLU) benchmark evaluates the capabilities of language models. Our models consistently score higher than state-of-the-art language-specific LLMs like HyperClovaX in Korean, Airavata in Hindi, Jais in Arabic, and Rakuten-7B in Japanese.
SUTRA outperforms other models in Hindi and can even recognize and handle Hinglish, a blend of Hindi and English spoken colloquially by millions of people across India.
In my native language of Gujarati, **SUTRA is more natural and comprehensive compared to any other model.
In Korean, a key market for us, our models have shown substantial improvements in MMLU scores, even exceeding state-of-the-art Korean LLMs like NAVER's HyperClova X model.
SUTRA excels many other languages, such as Japanese and Arabic. SUTRA's strength lies in its wide-ranging multilingual abilities, at a fraction of the cost.
https://media.licdn.com/dms/image/D4E12AQFLb9WIErK4Uw/article-inline_image-shrink_1000_1488/0/1715797151772?e=1721865600&v=beta&t=m-klz-ITzQpglaMdafHJKQY488H0FGWGy2xsGLMYouI
We are excited to share more about how SUTRA is changing the LLM landscape for the non-English speaking world. You can already try SUTRA now in our playground at https://playground.two.ai/