because the finetuning tweets i used were massively outweighed by the original 40GB training corpus, all they did was influence the *flavor* of the text emitted by the model.
without them, it emits random semi-coherent text; with them, it emits random semi-coherent text that resembles tweets.