Does anyone use fine-tuned models?

Why don’t I use fine-tuned models for anything I do?

Why don’t people use models fine-tuned for specific languages in Cursor or Windsurf?

Decentralized open source consciousness | Values-aligned communities | citizen of Black Rock City

im working my way towards it, i have a project with a complex response format (pydantic structured outputs and tool calling) and im compiling a folder full of example input/output data, once i get up to about 50 or so very curated examples, i plan to finetune 4omini

the 15 examples i have are already 99% of my tokens used, attaching every example to every prompt to get the llm to behave how i want - im expecting fine tuning to save me massively on input tokens

Full-stack digital product consultant @ BoundCorp.net - eng @daylight

And then from what interface will you call out to that inference?

a rn app im building, which will respond to output from the LLM like suggested toolcalls (approve/deny modifying your profile, creating a log, etc), or specific screens/dashboards to display, or conversational flows to direct the user thru

glad to elaborate or give a demo! it's a hobby project (not my full time focus) but im building a chat app for natural language habit tracking and journaling, it's still very much in the "prototyping and seeing what gives good results" phases, but basically using structured outputs to combine reasoning, responding, and tool calling all into one inference