Giters

AkariAsai / self-rag

This includes the original implementation of SELF-RAG: Learning to Retrieve, Generate and Critique through self-reflection by Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi.

Home Page:https://selfrag.github.io/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Why not directly generate training data for the generator using GPT4?

kakaxisisan opened this issue · comments

commented

We briefly discussed in Section 3.2.1 in our paper, and there are essentially two reasons:

  1. We have four different types of reflection tokens and insert them almost at the segment level, resulting in a million orders of inferences if we run GPT-4 to insert reflection tokens at each segment independently. This is simply too expensive for us to afford, especially since our model input can get long (i.e., OpenAI API costs depend on input context length).
  2. We are also afraid that relying on GPT-4 completely to insert special tokens may also hurt reproducibility in the future, as the model behaviors have been reported to change over time e.g., How is ChatGPT's behavior changing over time? (Chen et al., 2023).
commented

I am closing this issue for now but feel free to reopen it!