This repository has been archived by the owner on Sep 30, 2022. It is now read-only.

ideas

could train both a binary classifier cross encoder and a confidence cross encoder, to give signals for when the model should try harder