I want to build an application using GPT API to identify specific elements (e.g. equity and inclusion language and practice) in documents. I have a list of questions that point towards these elements which I’d like GPT to answer. Through in-context learning, it does a decent job but still make mistakes sometimes. If I were to fine-tune gpt for this task, how should I prepare the training dataset? Should I provide examples of inclusive language and practice used and upload them to GPT? What would the “message” look like? I am a bit confused about how to do this.
For fine-tuning you use input - output pairs as training examples. So if you were to put together a data set, you could use as your input, i.e. the user message, a text snippet that you wish to evaluate vis-a-vis DEI language requirements and as your output, i.e. the assistant message, some form of evaluation/statement whether the requirements are met or not.
There are different ways you could approach this problem, but this could be one way under a fine-tuning approach.
You can find general guidance on fine-tuning here.