Hugging Face has released Shieldstral, a 3-billion-parameter multimodal safety classifier that achieves performance comparable to or better than models nearly seven times its size on text safety benchmarks, while also setting a new state of the art for multimodal safety classification.
The model redefines content moderation as a binary question-answering task. This simple formulation unifies diverse moderation tasks into a single yes/no problem, allowing heterogeneous safety datasets with different taxonomies to be consolidated under one training framework.
To train Shieldstral, the researchers constructed a dataset of approximately 54.1 million samples, detailing their curation and generation process. They also developed a fine-grained evaluation set to assess policy adaptability. The results demonstrate that a small, adaptive model can match or outperform much larger counterparts.
Shieldstral's approach highlights the potential for efficient, specialized safety classifiers that require fewer computational resources while maintaining high accuracy across various safety domains.