Fine-tuning Qwen 2.5 on domain-specific function calling: evaluation report

We fine-tuned Qwen-2.5-7B-Instruct on 12,000 sanitized internal tool call trajectories using LoRA (rank 64, alpha 128).

Results:

• Tool call JSON schema adherence jumped from 82% to 99.4%.
• Latency reduced by 40% compared to prompt-guided few-shot completions.
• Running smoothly on a single RTX 4090 with vLLM!