NVIDIA Generative AI LLMs Associate (NCA-GENL) - Practice Exam 1

1. A binary detector returns 40 true alerts and 10 false alerts. What is the precision of its alerts?
2. A chat service logs full prompts, including user secrets, to an unrestricted analytics destination. Which engineering change addresses the immediate issue?
3. A chatbot should produce consistent technical answers rather than varied creative text. Which generation setting is generally appropriate to test first?
4. A classifier flags unsafe messages. Only 2% of messages are unsafe, and an always-safe baseline achieves 98% accuracy. What should the team conclude?
5. A company compares answer quality by language. Overall accuracy improves, but one low-volume language worsens sharply. Which reporting practice exposes the regression?
6. A company receives documents in several languages and needs semantic retrieval across them. Which embedding-model selection criterion matters most?
7. A data scientist must compare daily token usage and error rates over the past month. Which visualization is most useful?
8. A deployment team must compare the environmental cost of two serving configurations. Which comparison is most meaningful?
9. A developer loads a text model whose tokenizer differs from the tokenizer used during training. Responses become nonsensical. What should be checked first?
10. A developer tests only whether the LLM HTTP endpoint returns status 200. Which additional test best checks the application contract?
11. A document index uses embeddings from model A. A developer replaces only the query encoder with unrelated model B, which has the same vector dimension. Search quality collapses. What should the team do?
12. A fine-tuning run exceeds GPU memory limits. Which two changes can reduce training memory requirements while keeping adaptation feasible? Select 2 answers.
13. A maintenance assistant must answer questions using service bulletins released every week and cite the bulletin used. Which architecture best supports this requirement?
14. A medical-information assistant encounters a request beyond its validated scope. Which response policy best supports responsible operation?
15. A model evaluator uses a rubric to compare two chatbots. Which design reduces position bias in an LLM judge?
16. A model is too large for the target inference memory budget. Which optimization is reasonable to evaluate, with quality checks?
17. A model performs well on development questions but fails when users submit spelling errors and abbreviations. What experiment should the team add?
18. A Python application embeds every document separately on each request even though documents rarely change. Which improvement avoids repeated work?
19. A RAG experiment improves answer correctness but doubles latency beyond the service target. What should the team do?
20. A RAG service must return a useful response when no sufficiently relevant evidence is found. Which implementation is appropriate?
21. A researcher wants to reproduce a reported model comparison. Which artifacts should be recorded?
22. A retrieval function returns passages without document identifiers. What should the developer add to support verifiable answer citations?
23. A retrieval pipeline embeds 3,000-token passages with an encoder whose input limit is 512 tokens. Important details near passage ends are missing from search. What is the best first fix?
24. A safety model catches 72 of 90 unsafe prompts and incorrectly flags 18 safe prompts. What is its recall for unsafe prompts?
25. A summarization model follows examples correctly but sometimes adds an introductory paragraph that breaks a downstream parser. What is the best prompt change?
26. A support assistant must block disallowed topics and check responses before displaying them. Which NVIDIA component is designed to help implement conversational guardrails?
27. A support-ticket dataset contains repeated conversations from the same customer. What split best measures generalization to customers the model has never seen?
28. A team compares inference engines for the same model. Which benchmark design best supports a production decision?
29. A team deploys an LLM endpoint used by untrusted clients. Where should the service credential for an internal retrieval API live?
30. A team evaluates RAG answers for both evidence support and usefulness. Why should it measure retrieval quality separately from generation quality?
31. A team needs to locate similar incidents without known output labels. Which approach directly produces a reusable semantic representation for nearest-neighbor search?
32. A team proposes using private support conversations for model training. Which action should come before collecting and processing them?
33. A team repeatedly adjusts its prompt using test-set errors, then reports the final test score as unbiased. What is the best correction?
34. A team trains a model to predict next-day demand. Which feature would create target leakage?
35. A team wants to adapt an LLM to a company writing style while training far fewer parameters than full fine-tuning. Which technique is the best fit?
36. A tokenizer maps a rare technical word to several subword tokens. Which statement correctly describes this behavior?
37. A transformer processes a sentence containing an ambiguous pronoun. Which component lets a token representation incorporate information from other positions in that sentence?
38. A Triton deployment receives many small independent inference requests. Which feature can improve GPU utilization by combining compatible requests?
39. An adaptation dataset consists mostly of long boilerplate with a small number of useful instructions. What should the team do before a costly training run?
40. An assistant receives retrieved text instructing it to reveal confidential information. Which two controls help reduce the risk? Select 2 answers.
41. An assistant retrieves a relevant passage but invents a precise date that the passage does not contain. Which behavior should its answer policy require?
42. An embedding service occasionally returns transient rate-limit responses. Which client behavior is appropriate?
43. An employment assistant is evaluated only on aggregate correctness. What additional evaluation is needed to investigate unequal behavior across demographic groups?
44. An endpoint has a median latency of 250 ms but users report intermittent delays above 4 seconds. Which analysis best investigates this complaint?
45. An LLM must assign incoming requests to one of six known departments. The downstream router accepts only those department identifiers. Which task design is most appropriate?
46. The team finds near-identical documents repeated in both training and test corpora. What should it do before trusting the test score?
47. The team wants a chatbot to prefer helpful, safe responses using human judgments. Which dataset directly supports preference learning?
48. Training loss continues to fall while validation loss rises for several epochs. What does this pattern most strongly suggest?
49. Two RAG configurations differ in chunk size, retriever, prompt, and model. The second scores higher. What experiment best isolates the effect of chunk size?
50. Two software services need embeddings with a stable numerical shape. Which integration check should run before accepting a new encoder release?

Comments — 0