eval_loss = 0.6586902290628648
perplexity = tensor(1.9323)