eval_loss = 0.10016461255840213
perplexity = tensor(1.1054)