Small-Model Function Calling — try it

Send a query with some tools and see how each fine-tuned Llama-3.2-1B responds. The interesting cases are the irrelevant ones (e.g. "What's 2+2?" with only a weather tool): a good model should refuse to call and answer conversationally. Compare v1 SFT (tends to always call — even nonsensically) with Path B (learned to refuse).

Runs on free CPU — each response takes ~15-30s, and switching models reloads (a bit slower on the first run). Full study: github.com/keitake123/llama-function-calling-study