Data Sampling and Augmentation
12 min readUpdated June 1, 2026
Labeled data is often the most expensive part of building an ML system. Collecting high-quality labels takes time, expertise, and money, and large models need a lot of it.
Two techniques help you get more value from that data:
- Sampling decides which examples the model trains on, so you focus on the most useful data.
- Augmentation creates variations of existing examples, effectively expanding the dataset without new labels.
Together, they help you improve performance without increasing labeling cost.