The list it picks from

Every model has a fixed list of the text fragments it is able to use. jar, the, a comma, ievable, tens of thousands of entries. Those fragments are its tokens, and the list is its vocabulary, settled before training begins and unchanged afterwards.

You have already seen it working on the way in, cutting your sentence into tokens that all had to come from it. The same list does the work on the way out. Every entry gets a score, one of them is chosen, and the fragment it stands for is what appears on your screen.