I have! they're pretty bad at maintaining coherence overtime. There's not that many pieces of text that look like it in the base distribution. I do think maybe a base model trained on enough verbatim…