Shieldstral is a 3B open-weights multimodal safety classifier that frames content moderation as a policy-adaptive question-answering task, allowing it to accept plain-language policies at inference time without retraining. It unifies text and image safety evaluation and delivers calibrated safety scores across diverse benchmarks while running efficiently on a single 16GB NVIDIA GPU.
- Outperforms guardrail models up to 7x its size on text safety and sets a new state of the art on multimodal moderation.
- Accepts free-form natural-language policies at inference time, enabling adaptation to novel contexts without retraining.
- Returns a continuous calibrated safety score via softmax normalization of yes/no logits from a single forward pass.
- Trained on heterogeneous data sources consolidated into an instruction-query-document format with deliberate contrastive pairs to teach discrimination.
This approach allows developers to adapt safety definitions dynamically based on product context and audience, avoiding the need for fixed harm taxonomies baked into model weights.