A presentation on generating synthetic data for large language models, with focus on defining heuristics and creating diverse user query examples in healthcare. Shows how systematic synthetic data can scale AI work across an organization.
Are you an AI builder struggling with figuring out the best approach for generating synthetic data for your LLM application? A common mistake I've seen is prompting an LLM, very generically, to just give user queries for a given domain space, like a health coach. So where I would start is start with defining the dimensions or the heuristics that are important that capture the different aspects of your user queries in the health space. It might be protocol recommendations, habits, resting heart rate, heart rate variability, stress levels, then create create tuples. So it might be low stress. Exercised yesterday, didn't sleep well last night. Get about 20 and then convert each of those into a natural language query. That separation prevents repetitive phrasing. And the whole goal here, remember, is I want to bootstrap a synthetic data set so that I can get this out and cover as many cases as possible.
