Data · Established · Beginner

Data Labeling

Also known as: Annotation

Attaching ground-truth answers to raw examples so a supervised model has something to learn from.

What Data Labeling is

Labelling ranges from a single class per record to dense pixel masks, spans in text, or preference rankings between model outputs.

How it works

Teams write annotation guidelines, train annotators, measure inter-annotator agreement, adjudicate disagreements and re-label as the guidelines evolve. Model-assisted pre-labelling speeds throughput.

Why it matters

Label quality caps model quality. Most 'the model is bad' investigations end at inconsistent guidelines.

Common uses

→Vision datasets
→Named entity recognition corpora
→Preference data for alignment
→Search relevance judgements

Strengths

✓Directly controllable quality lever

Watch for

✓Slow and expensive
✓Ambiguity in guidelines becomes noise

Continue exploring

Active Learning Synthetic Data Reinforcement Learning from Human Feedback