Aya, a multilingual language model for 101 languages by Cohere

​Access to cutting-edge large language models (LLMs) has been restricted to speakers of only a few languages, primarily English.

The Aya project, developed by Cohere, aims to change this trend by prioritizing the acceleration of multilingual AI through an open-source initiative.

This endeavor has led to the creation of a cutting-edge multilingual instruction-tuned model and the most extensive multilingual instruction collection to date.

​The Aya collection was meticulously assembled by 3,000 independent researchers from 119 countries. These researchers dedicated themselves to crafting the collection by utilizing templates and translating existing natural language processing (NLP) datasets across 114 languages.

​Within this comprehensive collection lies the Aya dataset, which boasts the largest collection of original annotations sourced from native speakers worldwide, covering 65 languages.

​Furthermore, leveraging a diverse array of instruction mixtures, including the Aya collection and dataset, the Aya language model stands as a leading multilingual language model capable of executing instructions in 101 languages.

It has achieved state-of-the-art performance across various multilingual benchmarks.

Speaker bio:
Ahmet Üstün, a research scientist at Cohere For AI

​Ahmet earned his PhD from the University of Groningen, specializing in multi-task, multilingual, and efficient natural language processing. He focuses on modular approaches and addresses challenges related to low-resource languages.

He leads the Aya Open Science Initiative.