self-rag
self-rag copied to clipboard
Why could not train the generator directly?
The Critic is only used during training which produces training data for the generator. Why can't we directly generate data for the Generator?
I think this is because the training objective is to learn when to use retrieval and self-reflection. If we train only with distilled data from ChatGPT, the model would essentially be learning to generate ChatGPT’s complete answers directly. Not only is this unfeasible, but it also contradicts the intended objective.