Amazon SageMaker Ground Truth
Human labeling workforces and workflows for training data.
❓ What is it?
A data-labeling service that routes images, text, and video to human annotators — your own team, third-party vendors, or the Mechanical Turk crowd — with built-in task UIs and optional automated labeling that learns as humans label.
💡 Why does it exist?
Supervised learning is only as good as its labels, and labeling at scale is an operations problem: task design, quality control, workforce management. Ground Truth industrialises it and cuts cost via active learning.
⏱️ When should you use it?
Use it when you need labeled training data — bounding boxes, classifications, transcriptions — or human preference rankings for reinforcement learning from human feedback (RLHF) on generative models.
🗺️ Where does it fit?
Early in the ML lifecycle: it reads raw objects from S3, orchestrates the human workforce, and writes an augmented manifest back to S3 that training jobs consume directly.
🔌 How do you integrate it?
Create a labeling job, pick a task template, choose the workforce type (private, vendor, or public), set consensus/quality settings, and optionally enable automated data labeling to let a model pre-label easy items.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Ground Truth = creating labels BEFORE training; Augmented AI (A2I) = human review of predictions AFTER inference. Do not mix them up.
- Private workforce is the answer when data is confidential and cannot go to public annotators.