LLM-ACP exam questions

1. Question 1

   Single answer

   Which practice best reduces the risk of a fine-tuned model memorizing sensitive information?

   Response:

   • A

     Remove or mask sensitive data before training

   • B

     Add more sensitive examples to improve recall

   • C

     Disable all dataset review steps

   • D

     Publish the training data for easier debugging

   Reveal answer

   Correct answer

   • A

     Remove or mask sensitive data before training

2. Question 2

   Multiple answer

   Which practices improve production reliability for LLM applications?

   (Choose TWO answers)

   Response:

   • A

     Continuous monitoring of inference performance

   • B

     Controlled deployment and rollback procedures

   • C

     Storing model artifacts only on local laptops

   • D

     Disabling health checks to reduce overhead

   • E

     Deploying updates directly to production without testing

   Reveal answer

   Correct answer

   • A

     Continuous monitoring of inference performance

   • B

     Controlled deployment and rollback procedures

3. Question 3

   Single answer

   A user asks a question that is not answered by the retrieved documents. What should a well-designed RAG application do?

   Response:

   • A

     State that the available context does not contain the answer

   • B

     Invent a likely answer based on general assumptions

   • C

     Always cite the most similar unrelated passage

   • D

     Disable retrieval for all future questions

   Reveal answer

   Correct answer

   • A

     State that the available context does not contain the answer

4. Getting the hang of it? 90 more LLM-ACP questions are waiting, each with its answer. Unlock them all

5. Question 4

   Single answer

   An LLM application sometimes exceeds the model context limit when users submit long documents. Which design is most appropriate?

   Response:

   • A

     Split or summarize input before sending relevant context to the model

   • B

     Send the full document repeatedly until the request succeeds

   • C

     Remove all user instructions from the prompt permanently

   • D

     Disable validation and accept incomplete responses

   Reveal answer

   Correct answer

   • A

     Split or summarize input before sending relevant context to the model

6. Question 5

   Multiple answer

   Which practices help evaluate prompt quality?

   (Choose TWO answers)

   Response:

   • A

     Test prompts against representative user inputs

   • B

     Compare outputs against expected quality criteria

   • C

     Evaluate only a single successful example

   • D

     Ignore edge cases and unsafe requests

   • E

     Change prompts in production without review

   Reveal answer

   Correct answer

   • A

     Test prompts against representative user inputs

   • B

     Compare outputs against expected quality criteria

7. Question 6

   Multiple answer

   Which evaluation checks are most useful for a RAG application?

   (Choose TWO answers)

   Response:

   • A

     Whether retrieved chunks are relevant to the query.

   • B

     Whether generated claims are supported by retrieved context.

   • C

     Whether the ECS instance name uses a short label.

   • D

     Whether RAM users are sorted alphabetically.

   • E

     Whether the model ignores all retrieved passages.

   Reveal answer

   Correct answer

   • A

     Whether retrieved chunks are relevant to the query.

   • B

     Whether generated claims are supported by retrieved context.

   • C

     Whether the ECS instance name uses a short label.

8. Question 7

   Multiple answer

   Which measures help defend against prompt-injection attacks in RAG applications?

   (Choose THREE answers)

   Response:

   • A

     Treat retrieved documents as untrusted content

   • B

     Prevent retrieved text from overriding system instructions

   • C

     Validate actions before external tool execution

   • D

     Allow documents to redefine the assistant’s safety policy

   • E

     Expose hidden system prompts to users for transparency

   Reveal answer

   Correct answer

   • A

     Treat retrieved documents as untrusted content

   • B

     Prevent retrieved text from overriding system instructions

   • C

     Validate actions before external tool execution

9. Question 8

   Single answer

   Which problem is most likely if a fine-tuning dataset contains conflicting labels for the same type of user request?

   Response:

   • A

     The model may learn inconsistent response behavior

   • B

     The model will automatically resolve all label conflicts

   • C

     The model will stop requiring prompts after training

   • D

     The model will become a database synchronization service

   Reveal answer

   Correct answer

   • A

     The model may learn inconsistent response behavior

10. Question 9

    Single answer

    A company stores model artifacts in OSS and serves models through PAI-EAS. Why is this architecture commonly used?

    Response:

    • A

      It separates storage from inference serving responsibilities

    • B

      It eliminates the need for model versioning

    • C

      It prevents all operational failures

    • D

      It replaces monitoring requirements

    Reveal answer

    Correct answer

    • A

      It separates storage from inference serving responsibilities

11. Question 10

    Single answer

    Which prompt change best improves extraction reliability from invoices?

    Response:

    • A

      Define required fields and instruct the model to return null when missing

    • B

      Ask the model to invent missing totals if the invoice is unclear

    • C

      Request a creative story about the invoice layout

    • D

      Use a different response format for every invoice

    Reveal answer

    Correct answer

    • A

      Define required fields and instruct the model to return null when missing

12. Question 11 Locked Unlock to see this question
13. Question 12 Locked Unlock to see this question

Back to the LLM-ACP exam page