Shieldstral

Shieldstral is a 3B open-weight multimodal safety classifier that treats content moderation as a policy-adaptive question-answering problem. At inference time, it takes a policy (a question) written in plain language and evaluates the safety of text and images through a single interface, without retraining. It accepts a request in three parts (instruction, question, and document) and returns a continuous safety score by normalizing the yes/no logits. It unifies prompt classification, response moderation, refusal detection, and toxicity detection into one problem. It is released under Apache 2.0 and runs on a single 16GB NVIDIA GPU.

• Plain-text policy input at inference time (no retraining needed)
• Unified moderation of text, images, and text+image
• Calibrated continuous safety score (yes/no probability)
• Unifies prompt classification, response moderation, refusal detection, and toxicity detection
• 3B model that runs on a single 16GB GPU
• Apache 2.0 open weights

• One of 21 AI tools in the AI Security & Safety field.
• Connected to 5 other tools — #11 in its field.