引導式測驗 · 硬 等級
Llama 防護與安全分類器測驗
測試特定於版本的 Llama Guard 4 功能、安全性/不安全標籤、輸入和輸出過濾、策略類別和分類器限制。
問題 1 的 8
What does a Llama Guard classifier provide to an application?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- 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
- 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
- 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
- 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
- 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
- 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
- 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
- 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