Training yiffymix_v37 model

Recommended value from 1

there were about 20 attempts to create this LoRa by training on different models with different numbers of pictures, of all the versions this one works more stably