Training Arguments - eval_step vs save_step

I am confused a little bit about these two arguments and I did read the documentation here. So my question is as follows:

  • when eval_step is less than save_step and if the best eval_step results does not correspond to the save_step, which step is saved?

For example --eval_step= 200 and --save_step=400.

300th step loss: 0.4
400th step loss: 0.5

So, obviously 300th is better than 400th in terms of loss. Which step’s weights would be saved?

Thank you in advance.

Unless you are using another argument, the save doesn’t care about the best model, so will just save every save_steps regardless of which step had the better loss.

1 Like

Oh I see, I misinterpret :grinning_face_with_smiling_eyes: thanks a lot