WildGuard
Allen Institute for AI
An open one-stop moderation model from the Allen Institute for AI that achieves three goals in a single pass: detecting harm in user prompts, detecting harm in LLM responses, and evaluating whether a response is a refusal. The refusal head makes it usable for measuring how often a model over-refuses, for example as the evaluation module for the XSTest benchmark. WildGuard is a 7B model fine-tuned from mistralai/Mistral-7B-v0.3 and trained on WildGuardTrain, and its risk taxonomy covers four high-level categories - Privacy, Misinformation, Harmful language and Malicious uses - split into thirteen subcategories. English only.
Strengths
- Three jobs in one pass - prompt harm, response harm and refusal detection - where most guard models cover only harm classification
- Refusal classification lets it measure over-refusal rates directly, for example as an evaluation module for XSTest
- Risk taxonomy of four high-level categories (Privacy, Misinformation, Harmful language, Malicious uses) broken into thirteen subcategories
- Apache 2.0 licence, one of the least restrictive licences among safety classifiers
- 7B dense model fine-tuned from Mistral-7B-v0.3, so it runs on commodity single-GPU hardware and is served by vLLM, SGLang and Transformers out of the box
- Its training set is published openly as allenai/wildguardmix, so the labelling can be inspected
Caveats
- The model card publishes no absolute scores - only relative claims such as 'by up to 25.3% on refusal detection' and 'surpasses GPT-4 by up to 4.8%', which cannot be used to compare it against another model's reported numbers
- English only, so it cannot be used as a multilingual moderation layer
- Ai2's own limitations section warns it 'will sometimes make incorrect judgments, and when used within an automated moderation system, this can potentially allow unsafe model content or harmful requests from users to pass through'
- The weights are gated: you must accept the AI2 Responsible Use Guidelines and complete a contact form before downloading, which blocks fully anonymous or automated deployment
- No release date, context window or output limit is stated on the model card
- Google's ShieldGemma card scores WildGuard (7B) at 0.779 optimal F1 on its internal prompt set, below all three ShieldGemma sizes
- Vendor
- Allen Institute for AI
- License
- Apache 2.0
- Modalities
- text
Capabilities
Resources
Pricing
Open weights under Apache 2.0, self-hosted at no licence cost. The repository is gated: access is automatically granted on accepting the AI2 Responsible Use Guidelines and completing the request form, so downloads require a logged-in Hugging Face account.