3 ms·
Could this be implemented by making the loss zero if the output for the expected token is above some threshold? E.g. if the threshold is 0.1 it can output 10 l
by threatripper 2y ago
Could this be implemented by making the loss zero if the output for the expected token is above some threshold?
E.g. if the threshold is 0.1 it can output 10 likely tokens with 0.1 probability each and if any of them is correct it will not modify the weights.
- wongarsu 2y agoOne issue is that this doesn't punish the model of the top prediction produces complete nonsense. I can imagine getting good results when fine tuning an existing model though