[P1] customize a reward model

[XiaoMaGe-hero]

How can I customize a reward model? My model's output data requires some personalized evaluation criteria.

thx

[frankaging]

@XiaoMaGe-hero Hey, we provide a simple tutorial on how to use DPO with ReFT here. Let me know if this helps! For generic reward modeling, i think you just need to create a new Trainer (as our DPO trainer) to customize your own forward function as inputs and loss calculation could be different.