About the performance in RE-Docred in ATLOP

[LawsonAbs]

I run the atlop source code on re-docred. The configuration is :
bert-base-uncased + Re_docred
Sadly, I can't get the good performance. The several epoch in our experiments is as follows

what should I do?

[tonytan48]

Hi @LawsonAbs, I think ATLOP-BERT-base should be able to get 72-73 F1 performance. The scores are too long on your screenshot. Perhaps you can check the precision/recall on the reported score. If 34 F1 is due to high precision and low recall, maybe you used the original DocRED data for training and Re-DocRED for evaluation.

[LawsonAbs]

Thank you for your reply. I have solved this problem, because I ignored the setting in official_evaluate.