Yeah, very interesting... Like you could do a preliminary training that uses really good quality labeled data. Then almost infinitely continue training on a neg

Yeah, very interesting... Like you could do a preliminary training that uses really good quality labeled data. Then almost infinitely continue training on a negative only dataset of random noise, which would theoretically hammer out any false positives that can be generated by the model.