loading
loading
Training data generated by models rather than collected from humans.
Instead of scraping or paying labelers, you use a strong model to generate prompts, answers, reasoning traces, or preference pairs, then train on those — it's how teams scale post-training data cheaply and target specific skills like coding or math. Done well (filtering, verification, diverse seeds) it's now central to frontier training; done lazily it amplifies the model's own biases and feeds a slop loop where errors compound. The trick is grounding it: verify outputs against tests, tools, or real ground truth before training on them.
Plainly
Think of Synthetic Data as a simple recipe for doing the work better. Training data generated by models rather than collected from humans.
In practice
Use it when you need a repeatable method instead of guessing from vibes. In practice, define the owner, input, output, and failure mode before you rely on it.