eval_loss = 0.1959605920870567
perplexity = tensor(1.2165)