tips for resampling/balancing across datasets when mixing training data

[Wei-Qi]

This is a really impressive piece of work, and thank you for open-sourcing the code and models.

I have a quick question about multi-dataset training. When mixing multiple datasets with very different sizes, do you have any recommended resampling or balancing tips?

Thanks a lot!