Introducing Shieldstral.
mistral ·
Mistral has introduced Shieldstral, a compact three-billion-parameter open-source model designed to evaluate the safety of text and images through a flexible question-answering framework. Instead of depending on rigid harm taxonomies, the system allows developers to input custom safety policies as plain-language prompts directly during inference, removing the need for retraining when deployment contexts change. Despite its modest architecture, the model reportedly matches or surpasses systems up to seven times larger across various text and multimodal safety benchmarks. It outputs continuous, calibrated probability scores rather than fixed binary labels, giving applications greater control over risk tolerance and thresholding. To build the model, Mistral consolidated disparate public and synthetic datasets into a unified format and utilized contrastive training alongside parameter merging to improve nuanced policy discrimination. Licensed under Apache 2.0, Shieldstral operates efficiently on a single 16GB GPU and was developed using Mistral’s Forge platform, with future updates planned to enhance multilingual support and long-document processing.
mistral