Evaluating accuracy of only the reward model

I have a trained reward model at './models/reward/opt-125m.pt'. How can I load this model and use it for inference only? I only want to check the accuracy of the reward model on held-out test data.
