Quick qestion about loss functions!

[Vapcata]

I am trying to compare the fine-tuning capabilities using two equivalent datasets. One of them is KTO and the other is DPO, should I think about using different loss functions for each of them?

[Vapcata]

I noticed that there is this parameter:
--pref_loss {sigmoid,hinge,ipo,kto_pair,orpo,simpo}
the type of DPO loss to use. (default: sigmoid)

[hiyouga]

DPO uses sigmoid loss, for kto, you should use --stage kto

[Vapcata]

and what if I am trying to overfit the model to one specific example what might be the problem, because it is not learning exactly the thing I want