---
title: "Shieldstral: A 3B-Parameter Multimodal Safety Classifier Matching Larger Models"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.25857v1"
permalink: "https://feed.astralor.com/items/itm_9f1a439bae05dc26828f0064"
published: "2026-07-28T15:27:53.000Z"
score: 77
channel: "tech"
tags: ["AI Security","Shieldstral","Content Safety","Multimodal Classification"]
---

# Shieldstral: A 3B-Parameter Multimodal Safety Classifier Matching Larger Models

## AI commentary

> The most noteworthy aspect of this paper is that it achieves safety classification performance comparable to large models with very low model cost, which is very practical for teams with limited budgets or low-latency deployment scenarios. The data construction recipe is also worth learning—how to unify safety datasets from different sources for training.

## At a glance

Shieldstral is a 3B-parameter multimodal safety classifier that formulates content moderation as a binary question-answering task, unifying diverse safety datasets. Trained on about 54.1M samples, it matches or outperforms models nearly 7× its size on text safety benchmarks and sets a new state of the art on multimodal safety classification. This work demonstrates that small models can achieve efficient policy adaptability through carefully designed data recipes and training frameworks.

## Key points

- Only 3B parameters, outperforming models nearly 7× its size
- Unifies content moderation as binary QA, merging heterogeneous safety datasets
- Trained on ~54.1M samples with a fine-grained evaluation set
- Sets new state of the art on multimodal safety classification

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_9f1a439bae05dc26828f0064