Mistral AI Open Source Multimodal Content Moderation Model Shieldstral

Mistral AI's open-source multimodal content moderation model, Shieldstral, supports 12 languages and can run on a single 16GB GPU. The model achieves state-of-the-art (SOTA) performance in the multimodal content moderation field, outperforming competing products at a scale of 7 times. It innovatively uses natural language questions as input for moderation strategies, outputting a calibrated safety score through a "yes/no" question-and-answer mechanism. It supports tasks such as prompt word classification, answer moderation, and toxicity detection.