When fine-tuning is the right tool versus RAG or prompt engineering, full fine-tuning versus LoRA/QLoRA, RLHF and DPO as alignment-stage techniques, dataset quality, and how fine-tuning differs from distillation and quantization.
Fine-tuning is the only one of the three major adaptation techniques that actually updates a model's weights through additional training — which makes it the most powerful and the most expensive to get wrong. This module starts with the mistake made most often in practice: fine-tuning to teach a model facts, when facts belong in retrieval, not weight updates. From there it covers full fine-tuning versus parameter-efficient LoRA/QLoRA adapters, what fine-tuning is genuinely good for (format, style, and task-specific behavior patterns), RLHF and DPO as alignment-stage fine-tuning, and why dataset quality beats dataset size.
By the end of this module you should be able to explain why fine-tuning, distillation, and quantization are three different axes that get combined rather than substituted for each other, and how Module 2's train/validation/test discipline applies directly here — the exact foundation Project 3's Fine-Tuned Support-Ticket Classifier, later in this program, is built on.