Cohere's Aya work provides two distinct sources of examples for teaching models to follow instructions across languages. The Aya Dataset contains human-curated instructions and responses spanning 65 languages, gathered with fluent speakers. The larger Aya Collection contains 513 million examples across 114 languages, assembled by translating existing datasets and converting them into instruction formats. That distinction explains the headline scale: the larger collection broadens coverage through transformed data, while the human-curated dataset supplies a separate contribution from speakers themselves. Publisher abstract.
Terms in this piece · Glossary
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
postCohere releases open-weights translation model North Small Translate
articleThe Data Provenance Initiative A Large Scale Audit Of Dataset Licensing And Attribution In Ai 2023 10 25
