Aya is an open-access, instruction-finetuned multilingual language model introduced in “Aya Model: An Instruction Finetuned Open-Access Multilingual Language Model” (arXiv 2402.07827, February 12, 2024), led by Ahmet Ustun and colleagues at Cohere for AI and many collaborators. It covers 101 languages, with over half considered lower-resourced, roughly double the language coverage of comparable open models at the time.
The Aya effort was as much a data and community project as a modeling one. To train across so many languages, the team gathered and released large multilingual instruction datasets, drawing on a worldwide community of contributors, and then finetuned a model on that data. The paper reports that Aya outperforms earlier multilingual instruction models such as mT0 and BLOOMZ on the majority of tasks while covering far more languages, and it includes investigations into safety, bias, and toxicity across languages, not just English.
By open-sourcing both the model and the instruction data on Hugging Face, Aya gave the research community a strong, broadly multilingual instruction-following baseline and a template for building language coverage through collaboration rather than purely through scale.
For businesses serving global users, Aya illustrates that competitive instruction-following models can be built for a hundred-plus languages openly, narrowing the gap between English and everyone else.