Tiếp theoHướng dẫn tiếp theo
Bộ phân loại Phát hiện Xâm nhập Prompt
kỹ thuật
HƯỚNG DẪN KỸ THUẬT
Llama Guard is a family of safety classifiers that labels prompts and model responses against a defined hazard policy.
Nghe hướng dẫn nàyAI thuật lại · ~3 min
The current Llama Guard 4 12B release is a multimodal classifier for text and images, while earlier versions have different capabilities; a classifier result is one safety signal, not a complete moderation or security system.
Llama Guard is a safety-classification model intended to work alongside a generative model. It can classify an incoming prompt, a generated response, or both, and returns a safe/unsafe label with category information for unsafe content. That output can support a product’s moderation decision, but the application still has to define what to do with each label: block, transform, escalate, or permit. A classifier’s categories and training policy are not automatically identical to an organization’s legal duties or product rules. Capabilities depend on the version. Meta’s current Llama Guard 4 12B model card describes a natively multimodal classifier for text and one or more images, based on a hazard taxonomy, with prompt and response classification. It also describes an added text-only code-interpreter-abuse category. Earlier Guard versions are different; do not transfer Guard 4’s image support, parameter size, or language claims to them. The official documentation gives model-specific prompt formats and compatibility guidance, so deployments should verify the exact checkpoint and interface they use. Use input filtering when a product wants to stop certain requests before generation, and output filtering when it wants to inspect what the model actually produced. Using both provides two decision points, but can also increase refusals or latency. Meta documents limitations including imperfect policy coverage, dependence on training knowledge for some categories, and susceptibility to adversarial or prompt-injection attacks. Evaluate the full application with representative examples, inspect false positives and misses, and add other controls for sensitive decisions. A safety label is evidence for a workflow, not proof that all unsafe content has been caught.
An application checks a user prompt before generation, then separately checks the generated response before showing it, recording which stage triggered a block.
A team handling mixed text-and-image requests evaluates both parts together with Llama Guard 4 and tests its behavior on the exact image formats used by the application.
A product maps its moderation policy to the classifier’s supported hazard categories and sends uncovered cases to a separate review path.
A multilingual deployment tests representative languages, policy edge cases, false positives, and adversarial inputs before deciding where the model is suitable.
Safety classifiers are likely to remain useful as modular checks in systems that combine models, policy code, and human review. Multimodal checkpoints widen the kinds of input a single classifier can inspect, but broader coverage does not remove evaluation gaps or deployment-specific risks. Teams should track model-card changes, repeat tests when versions or policies change, and avoid treating vendor benchmark results as a substitute for testing their own use case. Publish internal acceptance criteria for the application’s actual language and modality mix.
Bản tin miễn phí
Ba câu chuyện AI đã được xác minh vào mỗi buổi sáng các ngày trong tuần, được viết bằng tiếng Anh đơn giản. Miễn phí mãi mãi, không có quảng cáo.
Một email mỗi ngày trong tuần. Hủy đăng ký trong một cú nhấp chuột. Chúng tôi không bao giờ bán hoặc chia sẻ địa chỉ của bạn.
Tự kiểm tra
Phản hồi tức thì về mọi câu trả lời và chứng chỉ có thể chia sẻ với ID có thể xác minh sau khi bạn vượt qua khóa học.
Hỗ trợ giáo dục AI miễn phí. AI Understanding là tổ chức phi lợi nhuận 501(c)(3) — không bao giờ có quảng cáo, không có tường phí. Hãy quyên góp
Llama Guard is a family of safety classifiers that labels prompts and model responses against a defined hazard policy. The current Llama Guard 4 12B release is a multimodal classifier for text and images, while earlier versions have different capabilities; a classifier result is one safety signal, not a complete moderation or security system.
The model card describes labels for safe or unsafe content and hazard categories; application policy still determines the next action.
Meta describes Guard 4 as natively multimodal and classifying text together with one or more images; it is not image-only classification.
The current card describes joint text-and-image input and says the model is not designed for image-only classification.
The model card distinguishes prompt classification before generation from response classification after generation.
Meta describes both filtering modes and notes that using both gives additional security, while limitations remain.
Tiếp tục học hỏi
Đã chọn thêm hướng dẫn cho chủ đề này
Tiếp theoHướng dẫn tiếp theo
Bộ phân loại Phát hiện Xâm nhập Prompt
kỹ thuật