This is a BIG question. There are many reasons that can result in training collapse. Can you provide more details? Such as the dataset, the pseudo-parallel data, the command you run, the logs of training, and so on. I guess that you may don't have a good warm-up using the pseudo-parallel data which is very important for RL training.

This is a BIG question. There are many reasons that can result in training collapse. Can you provide more details? Such as the dataset, the pseudo-parallel data, the command you run, the logs of training, and so on. I guess that you may don't have a good warm-up using the pseudo-parallel data which is very important for RL training.

hi fuli, I face with "ZeroDivisionError: division by zero" when calculate H_score. Is it caused by warm-up? Do you have some idea?

I have the same question.