Loading…
Glossary term
Artificially generated data, often produced by models, used for training or testing.
Synthetic data is generated artificially, often by another model, rather than collected from the real world. Examples include model-written questions and answers, simulated scenes or code with checked solutions.
It is valuable when real data is scarce, private or costly, but quality control is crucial.
As high-quality human text becomes scarce, synthetic data is an increasingly important input to training.