数据增强

Expanding training data via transformations, rewrites and similar techniques.

Augmentation creates variants via paraphrasing, back-translation and noise injection to improve generalization. In the LLM era, stronger models also generate synthetic data to fill weak spots.