How to reproduce your roberta-large baseline on DREAM ?

[oklen]

hi,I have run your code with: python run_classifier_roberta_exe.py dream roberta-large-mnli 1 24 and python run_classifier_roberta_exe.py dream roberta-large-mnli 1 64 Both give acc on test about 82.0 which is lower than 85.0 that you mention in your paper. Could you share how do you get your result? Thanks very much.

[jind11]

Hi, although effective batch size is more important, but if the gradient steps is too large, the performance may be degraded. I used the gradient steps between 4 and 6 and per gpu batch size of 2-4.