add new checkpoint trained for a hundred steps with smaller max grad norm and weight decay

update model with approx 1.6 epochs training