A simple question, I wonder if the seed argument in the TrainingArguments will be used for both weight initialization of a model and shuffling the data? And if it’s used only for weight initialization, what would be a good practice for data shuffling? Just randomly shuffle it ourselves?
1 Like