I tried something like this a while ago, maybe not as sophisticated as yours but still using the fastai library for the model and MBTI subreddits for data. All of my classifiers ended up predicting the dame MBTI type for everyone. For example, one classifier might predict “INTP” for maybe 90% of all inputs. If i try to train it again, it ends up predicting ENTJ for nearly everything.

I’m gonna try redoing this again with your method…