Skip to main content

Chapter 64: Supervised Fine-Tuning (SFT)

You build a fine-tuning skill first, then train models with LoRA/QLoRA (Unsloth) on Colab T4. The Task API dataset from Chapter 63 powers your first supervised run.

Goals​

  • Prepare and load SFT datasets for instruction-style tasks
  • Run LoRA/QLoRA fine-tuning with Unsloth on Colab/GPU
  • Evaluate checkpoints and manage artifacts
  • Capture repeatable prompts and configs in a fine-tuning skill

Lesson Progression​

  • Build the fine-tuning skill
  • Dataset formatting/validation for SFT
  • LoRA/QLoRA training runs with Unsloth
  • Checkpoint evaluation and export
  • Capstone: trained Task API model; finalize the skill

Outcome & Method​

You finish with a trained SFT model for the Task API and a reusable fine-tuning skill for future datasets.

Prerequisites​

  • Chapters 61-63 (strategy, architecture, data)
  • Python/Colab access with GPU