onrup

Data

Synthetic data

What is synthetic data?

Synthetic data is training data generated by a model rather than collected from real activity. It is useful for covering cases that are rare in real data and risky as a wholesale substitute, because it carries the generating model’s biases and blind spots.

The strongest use is targeted: you have identified a category your model handles badly and you have too few real examples of it. Generating more of that specific case, then having a person check them, is efficient and low-risk.

The weakest use is generating an entire dataset from a larger model and training on it unfiltered. The result is a smaller model that reproduces the larger one’s mistakes with less capacity to recover from them.

Whatever the source, provenance should be recorded per record. A dataset where nobody can say which examples were generated is a dataset nobody can debug.

Related terms

All terms in the glossary →

Start with the free tier

A magic link creates your account, your tenant and your first API key. No card until you ask for compute.