Shieldstral
2026-07-28 • Computation and Language
Computation and LanguageComputer Vision and Pattern Recognition
AI summaryⓘ
The authors developed Shieldstral, a relatively small AI model with 3 billion parameters that can judge if content is safe or not. It treats content moderation like a simple yes-or-no question, which helps combine many different safety rules into one system. By training on a huge, diverse dataset of over 54 million examples, their model performs as well as or better than much larger models. They also created a detailed test set to check how well the model adapts to different safety policies.
multimodal classificationpolicy adaptationcontent moderationbinary question-answeringmodel parametersdataset curationfine-grained evaluationsafety benchmarksmachine learningAI model size
Authors
Antonia Calvi, Avinash Sooriyarachchi, Giada Pistilli, Guillaume Lample, Maarten Buyl, Maximilian Augustin, Maximilian Müller, Pierre Stock, Tom Bewley, Wassim Bouaziz, Yimu Pan
Abstract
We introduce Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier that matches or outperforms models nearly 7$\times$ its size on text safety benchmarks and sets a new state of the art on multimodal safety classification. Shieldstral formulates content moderation as a binary question-answering task. This simple formulation unifies diverse moderation tasks into a single yes/no problem, enabling heterogeneous safety datasets with divergent taxonomies to be consolidated under one training framework. We present the data construction recipe, covering curation and generation of approximately 54.1M samples and a fine-grained evaluation set to evaluate policy adaptability. Together, these enable a small adaptive model to match or outperform much larger models.