WildGuard

WildGuard, from the Allen Institute for AI with the University of Washington and Seoul National University (June 2024), is an open one-stop moderation model: a single 7B model trained on the 92K-example WildGuardMix dataset to do three jobs — detect harm in user prompts, detect harm in model responses, and evaluate whether a response is a refusal. On the authors' WildGuardTest and ten public benchmarks, it beat the strongest open baselines (including Llama-Guard-2 and Aegis-Guard) by up to 25% on refusal detection and matched GPT-4 on adversarial prompt harmfulness.

Everything is Apache 2.0: model weights, training data, and code. For open-source builders who can't or won't depend on proprietary moderation APIs, it remains one of the most capable fully-open safety classifiers available.

Honest caveats: it is English-only and text-only — no image moderation — and its training leans on synthetic data labeled partly by GPT-4, which bakes in that model's biases. Newer multimodal guards (Llama Guard 4, ShieldGemma 2) cover more ground. But for a free, self-hostable, genuinely open-source moderation layer, WildGuard is still the benchmark.