Mistral's Shieldstral is a 3B safety classifier beating models 7x its size.

arXiv 2607.25857 reformulates all content moderation as a single binary yes/no QA task, which lets heterogeneous safety datasets with incompatible taxonomies consolidate under one training framework instead of requiring a taxonomy merge. It sets a new SOTA on multimodal safety classification, and the paper publishes the full data recipe covering roughly 54.1 million samples plus an evaluation set specifically for policy adaptability. That last piece matters because most guardrail models can't be retargeted to a new policy without retraining.