Fine-tuned a small LLM on our customer support data. Accuracy went from 60% to 89%. The trick was better data cleaning, not more data.