The Craft of Post-Training

Download Chapter 3: Supervised Fine-Tuning: The Foundation Technique

A pre-trained model has read most of the internet—and can be trusted with almost none of it. Post-training is the work that changes that: where you take a raw, general model and shape it into something that behaves, follows instructions, refuses what it shouldn’t do, and handles the specific job you need. It’s the human hand on the machine, and the part almost no one explains.

Chris von Csefalvay has spent his career building production ML systems in industry, from clinical language to legal text. In The Craft of Post-Training, he shows you the decisions behind every technique: when to fine-tune and when not to, why a model quietly gets worse, and which method fits the constraint you’re actually under. The math is here, because knowing why a technique works is what lets you debug it when it breaks.

You’ll know how to:

• Choose among the main post-training methods, from SFT and RLHF to DPO, KTO, and GRPO, well enough to fix failures instead of guessing
• Adapt a model to your domain without catastrophic forgetting—the tendency of a network to abruptly overwrite what it already knew when you train it on something new
• Run larger models with the memory you have by using new quantization
• Train agentic systems to act reliably under adversarial pressure
• Measure what matters in your deployment, beyond standard benchmarks

When you’ve used LLMs long enough, you start to wonder what was done to make them behave. The secret is in the post-training that shaped them. The Craft of Post-Training shows you how that’s done.