| 1 | Here are the headline results — table two from the paper. The Transformer big scores 28.4 BLEU on English to German, beating the previous best ensemble by two full BLEU points as a single model. On English to French it scores 41.8, a new single-model state of the art. And it does all this with about 2.3 times 10 to the 19 training floating-point operations, which is well below most strong baselines. The base model alone, trained in twelve hours, already beats almost every prior published model. The combination of better quality and lower training cost is what made this paper land so hard. |