Supervised learning needs answers. A label is the target value attached to a training example, the thing the model is supposed to predict. An image of a cat labelled cat. A loan application labelled default or repaid. A sentence labelled positive or negative.

Label quality determines model quality. Noisy labels teach the model to reproduce errors. Ambiguous labels confuse the decision boundary. Missing labels leave gaps in coverage. Practitioners spend far more time on labelling than outsiders expect.

Common labelling approaches

• Human annotation by trained workers
• Crowdsourcing with quality checks
• Automatic labelling from heuristics
• Weak supervision from multiple sources
• Active learning to prioritize uncertain cases

Some tasks need expert labelers. Medical imaging, legal documents, and scientific data all require domain knowledge that general crowds cannot provide. The cost of expert labelling often dominates the budget of a new project.