Unsloth Studio tutorial: fine-tune Qwen3 and export GGUF

This Unsloth Studio tutorial follows a local LLM fine-tuning workflow, from preparing a PDF dataset to exporting a model. The speaker introduces full fine-tuning, LoRA and QLoRA, describing the latter two as approaches that reduce hardware demands. Installation guidance points to separate commands for Mac and Windows, followed by a tour of the browser interface.

The example uses a 117-page document described as Google's 2026 sustainability report. In the data recipes interface, the presenter extracts PDF text, previews chunks and configures a model to generate question-and-answer training records. This preparation happens before training. The demonstration limits the dataset to 50 records, then selects a small Qwen3 model and QLoRA for a 30-step run.

The speaker explains training loss as a mistake counter and uses a driving analogy for learning rate. A side-by-side comparison then tests the base and fine-tuned models with a question about the report. The presenter explicitly acknowledges that the fine-tuned response does not accurately reflect the training data. The small sample and brief run make this a workflow demonstration rather than evidence of reliable document answers.

The final steps cover choosing a quantization level and exporting GGUF for use with llama.cpp. The example export is around 300 MB, and the presenter also shows where to chat with the trained model in Studio.