When preparing the data, we ran into a problem: after the first step of training, the model could translate over 90% of the training documents. Kocmi explains how we got the most difficult samples to level up North Small Translate's capabilities: (3/6) pic.twitter.com/QbyHI4o407
— Cohere (@cohere) 3 octobre 2026
When preparing the data, we ran into a problem: after the first step of training, the model could translate over 90% of the training documents. Kocmi explains how we got the most difficult samples to level up North Small Translate's capabilities: (3/6)