What is data annotation?

Data annotation is the work of adding labels to raw data so a model can learn from it. A model cannot learn to recognise a stop sign, a sarcastic review or a speaker change until people have marked thousands of examples of each.

Common annotation tasks

• Images and video: bounding boxes, polygons, keypoints and pixel-level segmentation.
• Text: named entities, sentiment, intent, topic and toxicity labels.
• Audio: transcription, speaker labels and timestamps.
• Model output: ratings and rankings of AI answers, which overlaps with evaluation and RLHF.

How it is paid

Simple, high-volume annotation is often paid per task and suits generalists. Annotation that needs expertise, such as medical images or legal documents, is paid hourly and at much higher rates. Our stats page tracks current rates by field.

Example

A retail project shows shelf photos and asks you to draw a box around every product and pick its brand from a list. Accuracy is measured against hidden gold tasks mixed into the queue.