Introducing Shieldstral. | Mistral AI

This post details the introduction of Shieldstral, a 3B open-weights multimodal safety classifier. It frames content moderation as a policy-adaptive question-answering task, allowing plain-language policies at inference time without retraining. The model unifies text and image safety evaluation, providing calibrated safety scores. Key technical contributions include data unification from heterogeneous sources, training for discrimination rather than memorization using contrastive pairs, grounding safety in images with reranking, and merging complementary checkpoints (public data, generated data, base instruct model) via SLERP. The development was conducted on Forge, Mistral AI's platform for training, aligning, and evaluating custom models.