Data augmentation

Data augmentation creates extra training examples by transforming existing data, such as crops, paraphrases, or audio noise.

Augmentation increases diversity without collecting new labels. Image pipelines use flips and crops; NLP uses back-translation or controlled paraphrase; speech uses noise and speed perturbation. Synthetic LLM data is a related but riskier path because the generator can imprint its own errors. Augment only with transforms that preserve the label, and keep a clean unaugmented eval set.