Fine-tuning takes a model that already understands language and specializes it. Instead of training from scratch (which requires massive compute), you start with a pre-trained model and train it on a smaller, task-specific dataset.
Common approaches include full fine-tuning (updating all parameters), LoRA (updating low-rank adapters), and instruction tuning (training on prompt-response pairs). RLHF (Reinforcement Learning from Human Feedback) is a specific fine-tuning technique used to align models with human preferences.