Synthetic training data
Synthetic training data is data generated by a model or simulator rather than collected from people, used to expand or balance a training set.
Last reviewed
What synthetic training data contains
- Model-generated text, code or images
- Simulated sensor or tabular records
- Generation prompts and seeds
- Filter and verifier results
Who buys it and how it is priced
| Attribute | Value |
|---|---|
| Belongs to | training data |
| Main buyers | Teams that lack real data in a domain or need privacy-safe substitutes |
| Pricing | Often priced like software or compute rather than labor. Value rests on the generator, the filters and the proof that it improves a model. |
Quality criteria and risks
Quality criteria
- Verified correctness
- Diversity
- No leakage of real records
- Measured gain on held-out tests
Risks
- Model collapse when models train on their own output
- Terms of the generating model
- Hidden privacy leakage
Providers that list synthetic training data
Auction and pricing mechanisms that fit
- Posted-price sale: Fits catalogue datasets that can be sold many times, such as cloud data exchange listings.
- Revenue-share licensing: Fits long-term content licenses where value is uncertain up front and both sides want to share upside.
Key terms
- synthetic data
- Synthetic data is data generated by a model or simulator instead of collected from real people or events.
- held-out set
- A held-out set is a portion of data deliberately excluded from training so it can test the model fairly.
- data card
- A data card is a short document that describes a dataset's contents, collection method, limits and intended use.
Frequently asked questions
- Can synthetic data replace human data?
- It can supplement it, especially where answers can be verified automatically. It does not remove the need for human data to set goals and check quality.