annotations

Definition

Human-supplied labels: the targets supervised learning is fit against. Their quality and quantity bound what the resulting model can do, and collecting them is slow and expensive, which is the pressure behind self-supervised objectives, weak supervision, and synthetic data. On subjective tasks, inter-annotator agreement effectively caps achievable accuracy — a model cannot beat the consistency of the labels it was scored on.

Explainers using this term

• Paper Walkthrough: Annotations as Rollouts — Dropping the Ground Truth Into the Group as a Ninth AnswerAnnotations as Rollouts: Efficient and Scalable Reinforcement Learning for Video MLLMs