Training Arguments - eval_step vs save_step

[sgugger]

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.