I am currently trying to use ChatGPT to expand the results of some survey-based research. My idea is to upload macroeconomic variables (such as the average spending power of the age group I am studying, the channels various groups tend to use to acquire products, etc.) and fine-tune the AI with a subset of the survey responses I have already collected from real respondents.

By combining macroeconomic variables with fine-tuning, I aim to generate additional responses from the AI. However, I am unsure about the precision of these generated responses. I want the AI to expand the responses in meaningful and innovative ways, rather than merely replicating or multiplying the responses I provide during fine-tuning.

I am considering using the portion of the survey responses not included in fine-tuning as a test set to evaluate the AI’s outputs. However, I am unsure which evaluation methods or tests would be appropriate for this purpose and whether this is the right approach to achieve my goals

I tried to search for some methods or ways, but I didn’t find anything that got me saying that’s it. I once used for a different project, that also involved an Ai, the K-means method but I don’t think that it’s the correct way to proceed this time.