Reading the InstructGPT paper(which seems to be what ChatGPT was built off of), I found this equation for the reward function. enter image description here

However, I'm struggling to understand how this equation is used to translate rankings between different response to a scalar.

For example, if I have four generated responses, A, B, C, and D, and with D being the best response, with D>C>B>A, then how is the loss calculated? How are relative scores converted to absolute losses.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.