Synthetic Data
Training examples generated rather than collected.
Overview
Synthetic data is training material produced by a model or a generator instead of gathered from the world. It solves the problem that high-quality labelled data is the binding constraint on further training, and that some categories are rare or expensive to collect. It exists because a model can generate far more examples than people can label, with the caveat that the examples inherit whatever the generator gets wrong.
Technology Graph
How Synthetic Data connects to the rest of the stack.
Used by
- Distillation technique Techniques Synthetic Data used by Distillation High ✓ source
Applied to
- Fine-tuning technique Techniques Synthetic Data used for Fine-tuning High ✓ source