[EXPERIMENT] - Diverse prompts generation with LLM call #106
Description
Most of the time I just use 1 or 2 prompt template for a generation. Rarely more.
But prompt are known to be critical for generative models, and even though we inject variables in prompt templates, to guide the generation it is still 1 overall prompt structure. What if we took the base prompt and prompt an LLM to generate a much more diverse set, from which 1 is picked at random for each sample generation?
This would require first some experimentation.