Shieldstral 1.0

A compact policy-adaptive multimodal safety classifier derived from Ministral 3 3B with a native Pixtral vision encoder. Instead of predicting a fixed moderation taxonomy, Shieldstral accepts a natural-language policy at inference time and emits a single-token yes/no verdict with a continuous confidence score. It handles text, images, and combined inputs in 12 languages, was trained on sequences up to 32K tokens, and is released under Apache 2.0.

The technical report describes a 54.1M-example construction pipeline. The model reaches 97.7 F1 on VLGuard and 81.8 on UnsafeBench, setting the paper's reported state of the art for multimodal safety classification while using substantially fewer parameters than the compared guard models.