Guidance on fine-tuning an LLM to simulate group behavior

Hello everyone,

I’m working on a project for a friend and would appreciate your advice and insights.

The core idea is to fine-tune a Large Language Model (LLM) to mimic the thinking patterns and behaviors of a specific group of people. Here’s a breakdown of what we’re aiming to do:

Data Input: We plan to train the model on a dataset containing information about a particular demographic group. This data will include:

  • Age ranges

  • Gender

  • Family situations

  • Financial situations

  • Responses to various questions (e.g., “What color do you like?”, “What are your key considerations when purchasing a TV?”)

  • Etc.

Desired Outcome: After fine-tuning, we hope the model can predict this group’s likely opinions and answers to new questions it wasn’t explicitly trained on.

I have a few questions for the community:

  1. Feasibility: Does this project seem feasible with the current state of LLMs and fine-tuning techniques?

  2. Relevant Techniques: If it is feasible, what fine-tuning techniques would be most suitable for this task?

  3. Model Selection: Could you recommend any open-source or close-source models that would be a good fit for this? We have a limited budget, so we are looking for models that don’t require prohibitively expensive hardware for local deployment and fine-tuning.

  4. Data Preparation and Training: Based on your answers to the above, how should we approach the fine-tuning process? Specifically, given our data, it seems challenging to create the structured question-answer pairs often used for supervised fine-tuning. Are there effective methods to fine-tune a model using the type of demographic and preference data we have?

Thank you in advance for your time and any advice you can offer!

Felix