ShieldGemma: Generative AI Content Moderation Based on Gemma

TL;DR

ShieldGemma introduces advanced LLM-based safety content moderation models that outperform existing solutions and includes a novel data curation pipeline, enhancing safety and generalization in AI-generated content moderation.

Contribution

The paper presents ShieldGemma, a new suite of LLM-based safety moderation models with superior performance and a novel data curation pipeline, advancing LLM safety research.

Findings

Outperforms Llama Guard (+10.8% AU-PRC) and WildCard (+4.3%) on benchmarks.

Demonstrates strong generalization with synthetic data training.

Provides a new resource for the research community.

Abstract

We present ShieldGemma, a comprehensive suite of LLM-based safety content moderation models built upon Gemma2. These models provide robust, state-of-the-art predictions of safety risks across key harm types (sexually explicit, dangerous content, harassment, hate speech) in both user input and LLM-generated output. By evaluating on both public and internal benchmarks, we demonstrate superior performance compared to existing models, such as Llama Guard (+10.8\% AU-PRC on public benchmarks) and WildCard (+4.3\%). Additionally, we present a novel LLM-based data curation pipeline, adaptable to a variety of safety-related tasks and beyond. We have shown strong generalization performance for model trained mainly on synthetic data. By releasing ShieldGemma, we provide a valuable resource to the research community, advancing LLM safety and enabling the creation of more effective content moderation solutions for developers.