Rejection Sampling Fine-Tuning Quiz

Answers and explanations appear as you take the quiz.

1. Which loop best describes how Rejection Sampling Fine-Tuning improves a model?
   • Generate many candidate answers, keep the best-scoring ones, and retrain on them
   • Delete the worst-performing model layers
   • Reject all training data and start over
   • Sample once and always accept the output
2. In verifiable domains like math or code, what can serve as the reward?
   • A learned reward model only
   • Whether the answer is correct or the tests pass
   • The length of the response
   • The number of tokens generated
3. Which model family does the cited paper describe using rejection sampling in its alignment pipeline?
   • GPT-4o
   • BERT Base
   • Meta’s Llama 2-Chat
   • Whisper Large
4. Why might a team choose rejection-sampling fine-tuning over a policy-gradient approach for some tasks?
   • It never needs a candidate generator or scoring rule
   • It guarantees no reward hacking
   • It always requires less compute than every RL method
   • It can use supervised learning on selected generations instead of updating through policy-gradient RL
5. Training on the maximum-reward samples effectively does what?
   • Distills best-of-N quality into the model's single-sample outputs
   • Increases the vocabulary size
   • Removes the model's ability to generate text
   • Doubles the parameter count
6. What should teams monitor when repeatedly fine-tuning on filtered model-generated outputs?
   • Only the number of model parameters
   • Output diversity alongside held-out correctness
   • Whether every answer uses the same wording
   • Only the training loss on selected samples
7. What is a limitation of rejection-sampling fine-tuning’s candidate selection?
   • It can only be used with image labels
   • It always changes the tokenizer
   • It cannot train on a high-quality answer that was never generated or retained
   • It makes the scoring function unnecessary
8. What risk arises if the scoring model used to filter candidates is flawed?
   • The policy becomes unable to generate any text
   • The tokenizer automatically changes vocabulary
   • The reference labels become unnecessary
   • The policy may learn to exploit the scorer rather than improve the intended quality