Why this matters Most moderation systems either fix a taxonomy at training time or require heavyweight ensembles to cover text and images. Shieldstral reframes moderation as a binary question-answering task: you provide a natural-language policy (a single yes/no query) at inference time and the model returns a calibrated safety score. This lets one compact model cover diverse moderation needs without retraining and simplifies policy updates in production.
Key Capabilities
- Policy-adaptive QA formulation: accept free-form natural-language queries at inference time and output a single yes/no token whose renormalised probability yields a continuous unsafe score, enabling quick policy changes without model updates.
- Unified multimodal interface: one call handles text-only, image-only, or text+image inputs via a Pixtral vision encoder and a shared chat-style template, reducing integration complexity across content types.
- Small-footprint deployment: a 3B-parameter causal model with native multimodal support trained on a large consolidated safety corpus (~54.1M samples), engineered to run on a single 16GB GPU in BF16 for many real-time use cases.
- Measured performance: strong benchmark results (text F1 ~84.9% average; multimodal F1 ~83.8%), matching or outperforming much larger baselines on several safety and refusal detection datasets.
Who it's for — tradeoffs and suitability
Great fit if you need a compact, policy-flexible safety classifier for production moderation (prompt screening, response moderation, refusal detection, or image+text checks) and want to update policies without retraining. Look elsewhere if you require absolute state-of-the-art capability on every narrow benchmark (larger specialist ensembles can still win some subsets), need perfect robustness to highly adversarially obfuscated inputs, or require exhaustive coverage for low-resource languages where training data is sparse. Residual label noise and uneven language/domain coverage are realistic constraints to plan for.
