| 1 | The variation study in table three gives us five takeaways worth remembering. First, head count has a sweet spot: a single head loses about 0.9 BLEU compared to eight heads, but going to thirty-two also degrades. Second, the key dimension matters more than the value dimension — determining compatibility is the harder learning problem. Third, scaling up the model dimension and the depth consistently improves quality, so the architecture scales monotonically with capacity. Fourth, dropout is essential — removing it loses roughly half a BLEU to a full BLEU point even at p equals 0.1. Fifth, label smoothing slightly hurts perplexity but improves BLEU, which is a reminder to optimize for the metric you actually ship on. |