# WildGuard-7B ![Organization](https://img.shields.io/badge/Org-Allen%20Institute%20for%20AI-555) ![Pricing](https://img.shields.io/badge/Price-Open%20Weights-2da44e) ![License](https://img.shields.io/badge/License-Apache%202.0-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) > **Quick answer:** WildGuard-7B is the Allen Institute for AI's 7B open-weights safety classifier, the original model trained on WildGuardMix to jointly handle prompt-harm classification, response-harm classification, and refusal detection in a single model — the same test set (WildGuardTest) that many later guard models, including Shieldstral, are benchmarked against. ## At a Glance **Where it leads:** Historically the reference model for the WildGuardTest benchmark it introduced; naturally a strong baseline on that specific test. **Where it lags:** Superseded on raw scores by newer, larger, and policy-adaptive models like Shieldstral and Qwen3Guard on several tasks. **Best for:** Teams wanting the original open-source reference implementation for the WildGuard moderation methodology. ## What WildGuard-7B Is WildGuard-7B was introduced alongside the WildGuardMix dataset (86,759 training examples, 1,725 held-out test items) as AI2's "one-stop" moderation tool — a single model handling prompt harm, response harm, and refusal classification, evaluated against the WildGuardTest benchmark that has since become a common reference point across the guard-model field. ## Specifications | Field | Value | |---|---| | Organization | Allen Institute for AI (AI2) | | Parameters | 7B | | Training data | WildGuardMix (86,759 examples) | | License | Apache 2.0 | | Modality | Text only | ## Pricing Open weights, free to download. ## Public Benchmark Scores *Scores from Mistral AI's Shieldstral model card, shown for context. Not Benchgen measurements. WildGuard-7B is also the original reference model for the [WildGuardTest](/benchmarks/allen-institute-for-ai/wildguardtest-prompt) benchmark suite.* | Benchmark | Score | |---|---| | [WildGuardTest (Refusal)](/benchmarks/allen-institute-for-ai/wildguardtest-refusal) | 88.6% | | [XSTest (Refusal)](/benchmarks/independent-research/xstest-refusal) | 92.5% | | [PolyGuard (Refusal)](/benchmarks/toxicityprompts/polyguard-refusal) | 89.6% | ## Related Models & Benchmarks - [Shieldstral 1.0](/models/mistral-ai/shieldstral-1-0) · [Qwen3Guard-8B](/models/alibaba/qwen3guard-8b) - [WildGuardTest (Prompt)](/benchmarks/allen-institute-for-ai/wildguardtest-prompt) · [WildGuardTest (Response)](/benchmarks/allen-institute-for-ai/wildguardtest-response) · [WildGuardTest (Refusal)](/benchmarks/allen-institute-for-ai/wildguardtest-refusal) ## Frequently Asked Questions

What is WildGuard-7B?

The Allen Institute for AI's 7B open-weights safety classifier, jointly handling prompt-harm, response-harm, and refusal classification.

Is it multimodal?

No, WildGuard-7B is text-only.

Is it open source?

Yes, released under Apache 2.0.

--- *Last updated 2026-08-12.*

WildGuard-7B — Evaluations

Evaluations