Synthetic data
Also known as: synth data · generated data
Synthetic data is training data generated by an AI — typically a strong model writing examples, questions, or solutions that a smaller model then learns from.
It’s a workhorse of modern training because it’s:
- Cheap and fast — no human annotators at scale.
- Controllable — generate exactly the distribution you want.
- A quality risk — if the generator is wrong, you’re training on errors.
Watch the recursion: a model trained on data generated by its own family can drift (model collapse). Curate hard, and prefer strong generators for weak students.