Gemma 2 2B fine-tuned for function calling with QLoRA: 93.7% F1, 6% hallucination. LoRA, GGUF and vLLM builds plus train/eval data.