adjust hyper-parameters

train on 800k examples

train with whole dataset

show also examples leading to infinite eval loss

update train script

update train script

update train script

do not generation prompt in the end, irrelavent for training and evaluation

decrease val size

re-implement dataset filtering in more efficient way

bugfix

filter out samples too long for MAX_LEN from dataset

eval loss got NAN but train loss kepp finite, adjust hyper-parameters

adjust train parameters to prevent from underfitting

initial commit