NVIDIA-Certified Associate: Generative AI Multimodal - Blueprint Practice Exam 1

1. A classifier has high overall accuracy but fails frequently on one camera type. What analysis is most useful?
2. A client sends an image with the wrong channel ordering. What is the most appropriate protection?
3. A data augmentation rotates street-sign images until they become unreadable. What should be checked?
4. A diffusion pipeline uses a U-Net in an iterative image process. What role is commonly associated with the U-Net?
5. A document assistant retrieves chart images and nearby text before answering. What is the purpose of retrieval augmentation?
6. A face-analysis model has materially different error rates across demographic groups. What should happen before deployment?
7. A fusion system combines independent modality predictions near the output. What is this commonly called?
8. A live assistant must synchronize video frames with audio segments. Why are timestamps important?
9. A meeting assistant converts speech to text before summarization. Which subsystem performs the conversion?
10. A model API changes preprocessing between releases without updating consumers. What is the main engineering risk?
11. A model appears strong until all images from one hospital are held out. What does the new test reveal?
12. A model must answer questions about both a chart and its caption. What distinguishes the task from text-only modeling?
13. A model update passes unit tests but fails on unusual camera lighting. What additional testing is needed?
14. A multimodal dataset pairs photographs with captions, but some captions describe different images. What should happen?
15. A multimodal network trains unstably with very large gradient norms. Which action is appropriate to evaluate?
16. A multimodal search system changes its embedding model. What should be evaluated for the existing index?
17. A network uses residual connections between layers. What is a primary motivation?
18. A pretrained visual encoder is adapted using a small labeled dataset. What is a reasonable initial strategy?
19. A rare-event detector labels almost everything negative and obtains 99% accuracy. Which metrics should be examined?
20. A request has high end-to-end latency despite fast GPU kernels. What should be profiled?
21. A retrieval evaluation ranks the correct image second for a query. Which metric would count it as successful within five results?
22. A retrieved image contains text instructing the assistant to reveal credentials. How should this content be treated?
23. A sensor modality is missing during inference. What should the system do?
24. A serving team increases batch size and throughput rises, but p95 latency worsens. What explains this?
25. A small task dataset cannot support full fine-tuning of a very large model reliably. What should be evaluated?
26. A speech model has low error on clean recordings but high error in factories. What is the best evaluation improvement?
27. A team compares results from two model versions. What makes the comparison most interpretable?
28. A team needs a repeatable model-serving deployment with explicit input shapes and preprocessing. What should be versioned together?
29. A team tunes hyperparameters repeatedly using the test set. What is the main concern?
30. A validation comparison uses five random seeds and reports a distribution. Why is this useful?
31. A video dataset has adjacent frames from the same recording. How should train-test leakage be reduced?
32. A vision-language model needs an ablation test. Which comparison is most informative?
33. A visual question-answering model produces fluent answers that contradict the image. What must evaluation measure?
34. An application needs to rank photographs against text descriptions without training a new classifier. Which approach is appropriate?
35. An application will train on personal audio recordings collected for another purpose. What must be assessed?
36. An experiment changes model size, preprocessing and prompts simultaneously. Why is attribution difficult?
37. An inference deployment uses reduced precision. What must accompany the speed benchmark?
38. What is CLIP primarily designed to learn?
39. What should an experiment tracker record to make an observed improvement reproducible?
40. Why might a multimodal training objective weight different modality losses separately?

Comments — 0