Comment by khalic

You should look into LoRA, it’s a partial retraining method, doesn’t require nearly as much as retraining the whole model. It’s different from what this paper is suggesting. The self improvements in this paper even sets the rules for the improvements, basically creating new data out of what it has.

LoRA paper: https://arxiv.org/abs/2106.09685