Use the proper tools for the job. Either guidance (reborn just last week) or LMQL are two frameworks that can “enforce” any local model to output a json object.

Lllama.cpp supports BNF grammars. You basically tell the model runner the exact format of the output and where the data goes in it and even which list of values you expect, and it will only produce output in that format.

Useful for generating JSON and for document classification.