Comment by em500

I wonder if someone already made a somewhat turnkey setup like this? Small models should ideally not waste a lot of parameters on trying to memorize too much knowledge and then hallucinate / interpolate a lot of the gaps.