Cohere For AI (C4AI) announced Aya, a multilingual generative language model that follows instructions in 101 languages, of which over 50% are considered lower-resourced.
Aya is available in Indian languages such as Hindi, Marathi, Malayalam, Gujarati, and Telugu.
“Developed using a diverse mix of instructions from the Aya dataset and collection, among others, it achieves state-of-the-art performance across numerous multilingual benchmarks,” Cohere for AI said in a blog post.
Aya outperforms mT0 and BLOOMZ on the majority of tasks while covering double the number of languages.
“We introduce extensive new evaluation suites that broaden the state-of-art for multilingual eval across 99 languages — including discriminative and generative tasks, human evaluation, and simulated win rates that cover both held-out tasks and in-distribution performance,” researchers from Cohere for AI said in the research paper.
Aya, spearheaded by Cohere for AI, engages over 3,000 independent researchers across 119 countries. The decision to open-source both the model and dataset is significant, especially considering the scarcity of datasets for AI in various vernacular languages.
Cohere for AI, a non-profit research lab and a subsidiary of AI startup Cohere, describes Aya as one of the most extensive open science initiatives in machine learning, reshaping research by partnering with global independent researchers.