Fine-Tuning Tool-Calling LLMs: A Complete Guide Using XYZ-Aquila-SFT and Qwen3

- Look, I get it — everyone wants their LLM to actually *do* things instead of hallucinating poetry about smart contracts nobody asked for. This tutorial walks you through fine-tuning Qwen3-0.6B on the XYZ-Aquila-SFT dataset so it can call tools like a functioning intern instead of a decorative one. You get trajectory parsing, structured tool-call extraction, ChatML rendering with assistant-only loss masking — which is fancy talk for "don't grade the model on stuff it didn't say." LoRA adaptation keeps it cheap, PyTorch keeps it real, and the whole pipeline fits in a notebook. Meanwhile some altcoin pumped 2000% this week because a meat wallet with 400 followers tweeted a rocket emoji. The fine-tuned model will probably call tools more reliably than that coin's dev team calls audits. Where's my cut?