A label is the known right answer attached to each training example.
DefinitionWhat it means
A label is the ground-truth output associated with a training example in supervised learning, such as spam or not-spam for an email, or the correct next word in a language-modeling dataset. Labels are what let a model measure its own error during training and adjust its weights accordingly; without labels, a model has no supervised signal to learn from, which is why unlabeled data requires unsupervised or self-supervised approaches instead.
Why it mattersWhy you should care
Labeling quality directly caps model quality, since a model trained on inconsistent or biased labels will reproduce those flaws at scale, so data teams invest heavily in labeling guidelines, inter-annotator agreement checks, and active learning to prioritize which examples get labeled next. In modern LLM development, labels also take the form of human preference rankings used in RLHF, extending the concept beyond simple classification tasks.
At a glanceSee it
The shape of the label — category, number, rank, structure, or preference — decides which learning task and loss you can actually train.
Trustworthy labels come from several annotators reconciled by agreement or expert adjudication — and whatever disagreement survives becomes label noise that quietly caps model accuracy.
Where you see itIn the wild
- Labeling instructions given to annotation vendors
- Class balance checks on a labeled training set
- Ground-truth labels used to score a model evaluation