Do you include good examples and documentation for context, since they do not seem to include much training data directly related to what you’re working on? I have mixed to positive-ish results with that approach.
I mean how am I supposed to give it a 10k-50k page TRM? I would much rather read the TRM and understand things myself than have it spit out a questionable answer that I still have to double-check to make sure it hasn't hallucinated.