Cohere For AI releases Aya, an open model for 101 languages

Background

Cohere For AI, the nonprofit research lab of Cohere, set out to build a model that follows instructions in many more languages, including many lower-resourced ones ^[1].

What happened

In February 2024 Cohere For AI released Aya, an instruction-tuned model that follows instructions in 101 languages ^[1][2]. The paper describing it was posted to arXiv on 12 February 2024 ^[1].

Key details

The model has 13 billion parameters and is licensed under Apache 2.0 ^[3]. More than half of the covered languages are considered lower-resourced ^[1]. The team said that Aya beats mT0 and BLOOMZ on most tasks while covering twice as many languages ^[1].

The instruction datasets were open-sourced along with the model on Hugging Face ^[1].

Impact

The release combined one model covering 101 languages with an Apache 2.0 licence and openly published instruction datasets ^[1][2][3].