Definition
Seed-and-amplify bootstraps a capability by generating a small high-quality dataset by hand and then using a model to expand it — rewriting, paraphrasing, generating new examples. It’s a workhorse technique for cheaply producing training data in niche domains.