Google

A specific method to enhance Retrieval-Augmented Generation (RAG) involves fine-tuning the Large Language Model with human-labeled data. This supervised training teaches the model to recognize when the retrieved context is inadequate and to refuse to provide an answer, thereby improving its reliability.

Fine-Tuning LLMs to Refuse Answering in RAG

Based on the scenario provided, propose a specific supervised training method to directly address the model's tendency to invent answers. Describe the essential components of the human-labeled data that would be required to implement this training.

Improving AI System Reliability

You are tasked with improving a question-answering system that generates answers from a set of provided documents. The system sometimes fabricates answers when the documents do not contain the necessary information. To fix this, you decide to fine-tune the underlying language model. Describe the key characteristics of the data examples you would need to create for this fine-tuning process to specifically teach the model to refuse to answer when appropriate.

Designing a Training Dataset for a Reliable Q&A System

A company develops a question-answering system that uses a large language model to answer queries based on a private document collection. They observe that when a user asks a question for which no relevant documents are retrieved, the system often generates a plausible-sounding but factually incorrect answer. Which of the following training approaches is specifically designed to mitigate this problem?

Learn Before

Related