When to early stop to get the good performance pre-trained models?

[irleader]

I tested e2efold with the pre-trained models provided, and the results on RNAstralign are very close (a little bit lower) to those in the paper，but results on ArchiveII is a bit lower than those in the paper (is this within reasonable range?) :

Rnastralign short:
Average testing F1 score with learning post-processing: 0.84188575
Average testing F1 score with learning post-processing allow shift: 0.8534005
Average testing precision with learning post-processing: 0.8722457
Average testing precision with learning post-processing allow shift: 0.88652223
Average testing recall with learning post-processing: 0.8236133
Average testing recall with learning post-processing allow shift: 0.83382463

Rnastralign long:
Average testing F1 score with learning post-processing: 0.7842659
Average testing F1 score with learning post-processing allow shift: 0.7964913
Average testing precision with learning post-processing: 0.8558946
Average testing precision with learning post-processing allow shift: 0.8695002
Average testing recall with learning post-processing: 0.72498244
Average testing recall with learning post-processing allow shift: 0.7361069

Archiveii short:
Average testing F1 score with learning post-processing: 0.55401367
Average testing F1 score with learning post-processing allow shift: 0.5792161
Average testing precision with learning post-processing: 0.6059438
Average testing precision with learning post-processing allow shift: 0.63850296
Average testing recall with learning post-processing: 0.53043944
Average testing recall with learning post-processing allow shift: 0.5530917

Archiveii long:
Average testing F1 score with learning post-processing: 0.11524726
Average testing F1 score with learning post-processing allow shift: 0.13892774
Average testing precision with learning post-processing: 0.12628905
Average testing precision with learning post-processing allow shift: 0.15100932
Average testing recall with learning post-processing: 0.10790618
Average testing recall with learning post-processing allow shift: 0.13134095

In order to reproduce the training process, I trained e2efold as well. While the results are terrible, I believe this is due to both pre-processing and post-processing networks are trained at a fixed number of epochs without early stopping.

Can you kindly indicate when to stop training pre-processing and post-processing networks respectively to get the pre-trained model you provided? (at which epoches/or any code you use to stop)
And where to modify these epoch numbers? (epoches_first and epoches_third in config.json?)