One can forgive the lack of quality results for the 70B model, but apparently they trained 7B and 13B versions of their model, and don't report those either.