Vibeleaderboard
← All Intel
Intel / article

Shieldstral

Source
arxiv.org
Author
MistralAI
Date

Context

Shieldstral, a safety classifier from Mistral AI, treats moderation as answering a yes-or-no policy question about content. That formulation addresses a training problem: safety datasets use different categories and labeling rules. The research abstract describes combining those datasets within a shared question-answering framework, using approximately 54.1 million curated and generated examples. The contribution is a common way to train across differing moderation policies, rather than requiring every dataset to share one category system.

Terms in this piece · Glossary
  • multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
  • inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
More from MistralAI
Recommended reads
Comments

Checking sign-in…

Loading comments…