When grokking, are you feeding the model the exact same data set over and over, or they just similar but different datasets?