Phase 6 · Modern LLM engineering · Post-training
Supervised fine-tuning
Teach behavior from example conversations.
Open this lesson Create a free account
Why it matters
Supervised fine-tuning is the first step of post-training and the most common way companies customize a model: show it examples of the conversations you want, and it learns to produce them. It's how a base model becomes an assistant, and how an assistant learns your output format, your brand voice or your support workflow. It's also the easiest stage to get subtly wrong.
What you'll be able to do
- Build SFT training examples from chats, with the loss masked to the assistant's reply
- Fine-tune a base model on demonstrations and measure the change on unseen prompts
- Choose sensible data, learning rate and number of steps, and spot overfitting
- Recognize what SFT looks like with real tools (Hugging Face TRL)
Before this lesson
Next in this phase: Instruction and preference data