Data augmentation
Also known as: augmentation
Data augmentation is generating variants of training examples to give the model more variety without collecting new data.
For text: rephrasing, synonym swaps, shuffling, masking. For images: flips, crops, color shifts.
It’s a form of regularization — more variety, less memorization.
Watch for text: augmentation is great, but for LLMs the higher-leverage version is usually *curated synthetic data* from a strong model, not blind random edits.