epoch, step, loss/eval, perplexity
1.0, 1000, 3.648764133453369, 38.427146911621094
1.0, 2000, 3.6033315658569336, 36.720367431640625
2.0, 1000, 3.6578149795532227, 38.77652359008789
2.0, 2000, 3.6434502601623535, 38.22349166870117