Guide-linked quiz · Hard Level

Llama Guard and Safety Classifiers Quiz

Tests version-specific Llama Guard 4 capabilities, safe/unsafe labels, input and output filtering, policy categories, and classifier limitations.

Question 1 of 8

What does a Llama Guard classifier provide to an application?

Preview every question8 questions

Answers and explanations appear as you take the quiz.

  1. What does a Llama Guard classifier provide to an application?
    • A guarantee that every response follows law
    • A safe/unsafe classification and, for unsafe content, category information
    • A replacement for the application policy
    • An automatic appeal decision
  2. Which modalities does Meta describe for Llama Guard 4 12B?
    • Images without accompanying text only
    • Text and images in a joint prompt
    • Audio only
    • Video streams only
  3. How should a team use Llama Guard 4 image support when reviewing a mixed request?
    • Replace it with a text-only description without checking fidelity
    • Apply Guard 4 claims to all earlier Guard versions
    • Assume any image is classified even if omitted from the request
    • Include the relevant text and image in the supported prompt format and test that path
  4. How does input filtering differ from output filtering?
    • Input filtering evaluates model weights; output filtering evaluates the user
    • They are interchangeable names for one pass
    • Output filtering changes the model’s training data
    • Input filtering classifies the request before generation; output filtering classifies the generated response
  5. Why might a system combine input and output checks?
    • The two checks create two decision points, though they still do not guarantee safety
    • Combining them eliminates false positives
    • The model then knows all current laws
    • It prevents all prompt injection by definition
  6. What should a product do when its policy includes a hazard not adequately covered by the classifier?
    • Skip policy documentation
    • Assume the nearest category is sufficient
    • Add another control or review path and test the uncovered case
    • Change the label in the returned text
  7. Why should teams avoid applying Guard 4’s feature claims to every Llama Guard release?
    • The family has no model cards
    • Only output filtering changes
    • Versions can differ in modality, size, and language capabilities
    • All versions share exactly the same checkpoint
  8. Which limitation does Meta’s Guard 4 model card disclose?
    • It can be susceptible to adversarial or prompt-injection attacks and has policy-coverage limits
    • It can only classify English text
    • It replaces all other system-level checks
    • It is immune to adversarial prompts

Keep testing yourself

More quizzes picked for your level.

View all quizzes