Developing Conversational Question-Answering Models through Two-stage Tuning

A collaboration paper from Google and other institutions introduces a two-stage tuning approach to enhance conversational question-answering models. The first stage involves supervised fine-tuning on instruction following and dialogue data to make the model behave conversationally. The second stage focuses on retrieval augmented generation, where the model learns to use an external database to pull in relevant information to answer queries. This required collecting a large dataset and providing extensive fine-tuning. The approach has been shown to behave comparably to GPT-4 in retrieval augmented generation tasks, making it increasingly important in grounding large language model responses in reality and preventing hallucinations.