Mistral researchers introduced Shieldstral, a small model that checks content for safety. It handles both text and images. It is built on Ministral-3B and treats moderation as a yes/no question. An operator writes the safety rule as a plain-language question at the time of use. The model returns a single score rather than a fixed category label. The team says it matches or beats models about seven times larger on text benchmarks. They report the best results so far on image-and-text safety tests. Training used about 54.1 million samples drawn from many public datasets. Those sources use different harm categories, so a template layer converted them to one format. The team also generated contrasting pairs, where the same content is paired with matching and non-matching questions.
What changed
Guardrail models classified content against a fixed list of harm categories.
What it unlocks
Setting moderation rules in plain language at the moment content is checked.
- 3B parameters
- 84.9% average F1, text safety
- 83.8% average F1, multimodal
- 91.3% F1 on fine-grained taxonomy
Sources