Mistral AI Open-Sources Multimodal Content Moderation Model Shieldstral
Mistral AI has released Shieldstral, an open-source multimodal content moderation model supporting 12 languages and runnable on a single 16GB GPU. The model achieves state-of-the-art performance in multimodal moderation, rivaling competitors seven times its size. It innovatively inputs moderation policies as natural language questions, outputting calibrated safety scores via a yes/no Q&A mechanism, suitable for prompt classification, response moderation, and toxicity detection.
