Jev vs a two-day fine-tuned open model on expense classification
Remote's team compared Jev with an open model fine-tuned in under two days on expense classification: accuracy was close, but the fine-tuned model knows when it is right, so it can approve more on its own and send less to review.
This is freaking cool. Everyone is talking about Jev right now. Mattias Svedhem on our team put it head to head with an open model he fine-tuned in less than two days, on one very specific job: classifying expenses. Accuracy came out close. The difference is that his model knows when it's right, so it can safely approve more on its own and send less to review. It's also faster for this particular task. For a lot of other work, Jev is still a better fit. It's an experiment, not something in the product. But I love that this is how things happen at @Remote: someone gets curious, and a couple of days later there's a demo.