The dataset is usually where fine-tuning projects fail — not the training code, not the hyperparameters, not the choice of base model. A model trained on 500 high-quality, diverse examples almost always outperforms one trained on 50 000 noisy ones. This lesson covers the three dataset formats, the data pipeline that turns raw material into training-ready data, and why synthetic data generated by GPT-4 has become the dominant approach.