eval_loss = 1.8500147744922628
perplexity = tensor(6.3599)