Annotation

Annotations come in many forms depending on the task: classification labels (positive/negative), bounding boxes (object detection), pixel-level masks (semantic segmentation), entity tags (NER), dependency trees (syntax parsing), and preference rankings (RLHF). Annotation quality directly impacts model quality — noisy or incorrect annotations lead to poorly performing models (garbage in, garbage out). Annotation guidelines must be carefully designed to ensure consistency across annotators. Modern annotation tools provide specialized interfaces for different data types and include quality assurance features like reviewer workflows and agreement metrics. The cost and time required for high-quality annotation is often the bottleneck in ML projects. Active learning techniques can reduce annotation costs by selecting the most informative examples for human labeling.