This PR slightly improves arg extraction and frame detection F1 due to smarter post-processing of T5 outputs. It also refactors some of the training / evaluation to make it easier to run standalone evaluation. It also changes the format of the training output filenames to have more information about validation F1 scores in a more machine-parsable format.
This PR slightly improves arg extraction and frame detection F1 due to smarter post-processing of T5 outputs. It also refactors some of the training / evaluation to make it easier to run standalone evaluation. It also changes the format of the training output filenames to have more information about validation F1 scores in a more machine-parsable format.