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

In this tutorial, we implement an end-to-end supervised fine-tuning pipeline for the XYZ-Aquila-SFT dataset, Hugging Face Transformers, PyTorch, and PEFT. We stream and inspect the dataset, parse multi-turn tool-use trajectories, extract structured tool calls, analyze corpus characteristics, and preserve embedded reasoning and observation patterns. We then convert tool schemas between message-embedded and structured formats, render Qwen-compatible ChatML with assistant-only loss masking, prepare a custom PyTorch dataset and collator, and fine-tune Qwen3-0.6B with LoRA. Finally, we evaluate tool-call prediction before and after training and export both the transformed dataset and corpus statistics for further experimentation.

We configure the dataset, model, training parameters, output directory, and reproducibility settings for the complete workflow. We install the required Hugging Face, PEFT, Accelerate, and PyTorch-related dependencies and detect whether a CUDA GPU and BF16 support are available. We then stream a limited number of XYZ-Aquila-SFT examples, inspect the dataset schema, and examine the structure of the first tool-use trajectory.

We define nesting-safe utilities for extracting JSON tool calls, reasoning blocks, observations, and embedded tool schemas from each conversation. We convert every raw dataset row into a structured trajectory object and verify that the parsed tool-call counts match the values declared by the dataset. We then calculate corpus-level statistics and visualize the distributions of tool calls, message depth, trajectory size, and tool usage frequency.

We extract embedded tool definitions into a structured format and reconstruct them to test whether the conversion preserves the original system message. We manually render each trajectory in ChatML format to retain all reasoning content and apply loss only to assistant-generated tokens. We also tokenize the examples, enforce the selected sequence-length policy, create the training and evaluation split, and prepare a padded PyTorch DataLoader.

We build teacher-forced evaluation probes by cutting trajectories immediately before assistant turns that contain tool calls. We load Qwen3-0.6B, measure its baseline tool-call performance, attach LoRA adapters, and fine-tune the model using gradient accumulation, mixed precision, checkpointing, clipping, and cosine learning-rate scheduling. We then evaluate the adapted model, compare its metrics with the baseline, and save the trained LoRA adapter and tokenizer.

We export every parsed trajectory as a structured JSONL record containing messages, tool schemas, questions, and answers. We also save a JSON report containing corpus size, tool frequencies, trajectory statistics, supervised-token ratios, and available evaluation results. We finish the workflow with reusable dataset artifacts, analytical outputs, and model files stored in the configured output directory.