ShieldGemma

Classifies incoming text prompts against safety policies (sexually explicit content, dangerous content, harassment, hate speech) using instruction-tuned Gemma transformer models (2B, 9B, or 27B parameters). Produces safety labels with configurable decision thresholds that can be adjusted per deployment environment, enabling teams to tune false-positive/negative rates based on risk tolerance. Models use open weights allowing fine-tuning to custom safety policies beyond baseline categories.

ShieldGemma 2 (4B parameters) classifies images for safety violations using multimodal transformer architecture that processes visual content directly. Detects sexually explicit imagery, dangerous/violent content, and other unsafe visual material. Operates as a standalone classifier integrated into image processing pipelines, with configurable thresholds for filtering generated or user-uploaded images in production systems.

Evaluates generated text responses from LLMs against safety policies post-generation, classifying outputs for sexually explicit content, dangerous instructions, harassment, and hate speech. Operates as a safety guardrail in generative AI pipelines, allowing rejection or regeneration of unsafe outputs before serving to users. Uses same instruction-tuned Gemma classifiers as input filtering with configurable thresholds for production deployment.

Enables organizations to fine-tune open-weight ShieldGemma models on custom safety policies and domain-specific harmful content using instruction-tuning methodology. Allows adaptation of baseline classifiers (sexually explicit, dangerous, harassment, hate speech) to organization-specific risks (e.g., financial fraud, medical misinformation, brand safety violations). Fine-tuned models retain open-weight format for local deployment.

Provides ShieldGemma in three text classification sizes (2B, 9B, 27B parameters) and one image size (4B parameters), enabling developers to select models based on latency/accuracy requirements. Smaller models (2B) run on CPU or edge devices with lower latency; larger models (27B) provide higher classification accuracy. Instruction-tuned architecture maintains consistent API across sizes, allowing model swapping without code changes.

Distributes ShieldGemma models as open weights (downloadable from Kaggle, Hugging Face, Google Colab) enabling local inference without cloud API calls or vendor dependencies. Models can be deployed on-premise, in private clouds, or air-gapped environments. Eliminates latency, cost, and privacy concerns of cloud-based safety APIs while maintaining full control over model versions and configurations.

Classifies text and images against multiple safety harm categories (sexually explicit content, dangerous/violent content, harassment, hate speech) in single inference pass using instruction-tuned Gemma models. Produces per-category safety labels enabling granular policy enforcement (e.g., reject hate speech but allow dangerous content discussions in educational context). Unified API across text and image variants.

ShieldGemma models and example code available on Kaggle, Hugging Face, and Google Colab, enabling rapid prototyping without local setup. Kaggle provides pre-configured notebooks with GPU access; Hugging Face hosts model weights and inference examples; Colab notebooks demonstrate end-to-end safety filtering workflows. Enables developers to test safety classifiers in minutes without infrastructure setup.