From [1]: > We demonstrate that our RAG approach trains the model to perform better RAG on the set of documents it is trained on i.e., in-domain. By removing t

From [1]: > We demonstrate that our RAG approach trains the model to perform better RAG on the set of documents it is trained on i.e., in-domain. By removing the oracle documents in some instances of the training data, we are compelling the model to memorize domain-knowledge. What if you wanted to train it to say that it didn't find the answer? [1] https://gorilla.cs.berkeley.edu/blogs/9_raft.html https://gorilla.cs.berkeley.edu/blogs/9_raft.html