Label
In one sentence A label is the correct answer attached to a training example — the thing you want the model to learn to predict.
Updated
A label is the correct answer for one training example: the output the model should produce when shown that input.
Flashcards for a toddler work because each card pairs a picture with its answer: mango on the front, "mango" written on the back. The picture is the input; the word on the back is the label. Show enough cards and the child learns the connection. Supervised-learning is this flashcard game at scale — millions of inputs, each paired with its answer.
Labels take whatever form the task needs. Spam or not spam. A price in rupees. A box drawn around every car in a photo. For an LLM, the label is famously cheap: the next token of existing text is the answer, so the internet labels itself — that is self-supervised-learning.
Everywhere else, labels are the expensive part. Someone must write them — see data-labelling — and people disagree, get tired, and make mistakes. Noisy labels put a hard ceiling on what any model can learn, since the model is trained to reproduce them, errors included. A related term you will meet is ground-truth: the label treated as the authoritative correct answer when scoring a model.
Where to go next
- Full lesson: Supervised learning
- Related terms: feature, ground-truth, supervised-learning, data-labelling