Rejection sampling (for training data)
Beginner
Make many answers, throw away the bad ones, and keep the good ones to learn from.
Novice
Generating many candidate answers and keeping only those that pass a check (for example, the right final answer) to build a training set.
Expert
Filtering sampled outputs by a verifier or reward model to build SFT data. Self-generated, filtered data is the core of STaR-style self-training and of R1’s second SFT stage.
See also: Distillation.