I mainly use 70B for “text QA” on files I find sensitive like personal documents. The answers have been very close to what I get if I use GPT-3 (langchain makes

I mainly use 70B for “text QA” on files I find sensitive like personal documents. The answers have been very close to what I get if I use GPT-3 (langchain makes it easy to switch). Do you use the quantized version? If so, try running the full one on a A100.