Created a really cool looping feedback system for training my model on its own output with me clicking thumbs up or thumbs down to validate, new dataset seems to be improving evaluation loss numbers on further training on the same dataset with the additional synthetic dataset. Very interesting find. I wonder If I should document this better...aaaaa

Posting here the updated model with 1 epoch of the new additional synthetic dataset based on this models current output.

Note: Output is actually lower quality than the ones I hand crafted.