hello, I have a question, why the DD datasets had high validation accuracy at the begging of the training. Could you please answer my question? 