Model Collapse Quiz

Answers and explanations appear as you take the quiz.

1. In the guide’s description, what process can cause model collapse?
   • Training only on independently collected real-world data
   • Training later models mainly on samples generated by earlier models
   • Testing a model on data it did not generate
   • Adding a few selected synthetic examples while retaining the original data
2. Which part of the distribution can be especially easy to lose during recursive training?
   • Every common pattern at once
   • Only the labels used in validation
   • Rare cases and low-frequency patterns
   • The examples that were held out from training
3. Why should a team keep an independently sourced evaluation set?
   • It guarantees that all synthetic data are accurate
   • It makes provenance records unnecessary
   • It ensures every model will improve
   • It can reveal coverage loss without repeating the generator’s own omissions
4. When testing synthetic augmentation, what comparison does the guide recommend?
   • A real-data baseline, the proposed mixture, and relevant alternatives under the same budget
   • Only the largest synthetic dataset available
   • A model trained on generated data with no evaluation set
   • A synthetic model against a different task and compute budget
5. What should a team record about generated training material?
   • Only the total number of rows
   • Only the final benchmark score
   • Its provenance and how it was generated
   • Only the model’s name
6. What does the guide conclude about synthetic data in general?
   • It always causes collapse
   • Its effect depends on the data mixture, generation, filtering, training, and evaluation setup
   • It always improves coverage
   • Its effect depends only on model size
7. In the guide’s invented 90-to-10 example, why does the next sample matter?
   • It proves a real model will lose exactly 80 percent of rare examples
   • It shows all synthetic data are biased
   • It measures the compute needed for training
   • A lower share of rare examples can change the balance represented in later training
8. What conclusion does the guide say a useful synthetic-data experiment should report?
   • Which tested conditions improved or degraded, with uncertainty and setup described
   • That every model will inevitably collapse
   • Only the highest average accuracy
   • A prediction about all future AI systems