How you do validation/stop training

[xixiddd]

Hi, In Model and Training Details section of your paper, you said that "We use early stopping and select the best model based on the F_0.5 score on the development set.". But to my knowledge, it seems that FairSeq can not do validation based on F_0.5, so my question(clarification) is

• In FairSeq, the Early-Stop is only triggered when current learning rate < min-lr ?
• after the FairSeq training process finished, the model paramters will be seved as checkpoint_1.pt , checkpoint_2.pt, ......, checkpoint_best.pt, checkpoint_last.pt, then you use the _checkpointbest.pt file directly, or manually select a checkpoint among all *.pt files as "best model" based on the F_0.5 score on development set ?

Many thanks

[shamilcm]

Yes, after Fairseq training finishes (using the early stopping criteria based on learning rate), the checkpoint that has the highest F_0.5 on the development data is selected as the final model.