Shieldstral
Antonia Calvi, Avinash Sooriyarachchi, Giada Pistilli, Guillaume Lample et al. · arXiv · 2026
Abstract
Mistral's 3-billion-parameter safety classifier, covering both text and images, which matches models around seven times its size on text safety benchmarks. The design choice doing the work is framing every moderation task as a single binary question, which lets safety datasets with incompatible taxonomies be trained together. The paper documents the recipe behind roughly 54 million training samples and an evaluation set built to test adaptation to a new policy.
Why it matters
Guardrails are usually the least documented part of a production stack. A small, open, policy-adaptable classifier is a practical way to run moderation without paying frontier-model prices on every call.
https://arxiv.org/abs/2607.25857