Best practices for fine-tuning for question answering

I’m referring to the notebooks in Github for fine-tuning a question-answering bot. The idea is to generate questions and answers using text-davinci-003 and use it to fine-tune a curie model. How can we ensure that the fine-tuned model answers based on what it has been trained on? What are the best practices, if any, to develop a model to answer queries based on large documents? How to restrict the fine-tuned model to answer queries only based on fine-tuning and not generate random answers? Sorry if this comes off as too much. Any pointers would be of great help!

1 Like