analiza aiqbench
Shieldstral introduces a compact 3B-parameter model that achieves state-of-the-art performance in multimodal safety classification by reframing content moderation as binary question-answering. This approach unifies diverse policy taxonomies and leverages over 54 million curated samples to outperform models nearly seven times its size, offering an efficient solution for scalable AI governance.
Shieldstral
arXiv:2607.25857v1 Announce Type: new 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.
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.
Computation and Language (cs.CL); Computer Vision and Pattern Recognition (cs.CV)
arXiv:2607.25857 [cs.CL]
(or arXiv:2607.25857v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2607.25857 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 131 cuvinte
