eval_loss = 1.060597886129753
perplexity = tensor(2.8881)