eval_loss = 0.43171565428771497
perplexity = tensor(1.5399)