Synthetic training data
Synthetic training data consists of generated examples intended for model learning. A simulation may generate images, a model may generate text, or rules may construct cases. Some synthetic data is based on real source data, while other examples are produced from a designed environment or specification.
Also known as: Synthetic data
How it works
Define the target task, choose a generation method, and evaluate whether the examples represent relevant conditions. Record source inputs, generation settings, filtering, and checks. Synthetic examples may be mixed with observed data, but the distinction should remain visible in documentation.
Generation can help cover selected conditions or create controlled variations. It can also repeat the generator’s errors or omit real-world complexity, so evaluation should include suitable independent data.
Why it matters for licensing
The rights to generate, use, or distribute synthetic examples may depend on source material, model terms, and the agreement. Privacy also needs assessment when generation is based on sensitive records. “Synthetic” is not a universal exemption from those obligations.
Example
Fictional example: A simulation produces labeled images of parts under varied lighting for a recognition experiment. The team tests the resulting model on separately collected real images and documents where the simulated conditions differ.
Limitations and misconceptions
Generated data is not automatically anonymous, diverse, accurate, or free of third-party restrictions. A generator can reproduce source details or amplify systematic errors. Synthetic coverage should be evaluated against the intended deployment setting.
Questions to ask
- How were examples generated, and what source inputs or models were used?
- How is realism, diversity, and task usefulness tested independently?
- What privacy and licensing constraints remain?