A proof of concept stitching together Claude, MCP, and faker to generate synthetic LLM fine tuning data.

The idea came from Apple Foundation Model's adaptor training example. They provide a toy dataset with data on short script generation. Wanted to see how one might stitch together Claude and MCP with tools to generate data.

A little finic-y but if you really were to try to generate hundreds or thousands of examples, you could throw in a loop or loosen the formatting constraints and post-process into the ideal format.