eval_loss = 1.5660290658616678
perplexity = tensor(4.7876)