Definition
Labeling turns raw data into supervised training fuel: class tags, spans, bounding boxes, rankings, or translations. Guidelines and inter-annotator agreement determine quality.
Active learning and weak supervision can reduce pure manual labeling cost.
In simple terms
Teachers grading practice problems so a tutor knows what "correct" looks like — labels are those graded answers.
Where you see it
- Transcribing Somali speech for ASR.
- NER tagging of person and place names.
- Preference rankings for RLHF.
How it works
1.Write guidelines
Clear rules and edge cases.
2.Annotate
Humans or assisted tools apply labels.
3.QA
Review agreement, fix drift, version the dataset.
Why it matters
- For many tasks, label quality beats model cleverness.
Often confused
LLMs removed the need for labels.
LLMs shift where labels help — eval sets, fine-tunes, and preference data still need careful annotation.