Shieldstral - Mistral AI's open-source multimodal content security classification model

What is Shieldstral?

Shieldstral is an open-source, multimodal content security classification model with 3B parameters, developed by Mistral AI and built upon Ministral-3B. The model redefines traditional fixed-category content moderation as a binary question-answering task, supporting real-time definition of moderation policies through natural language queries, and adapting to different scenarios without retraining. The model achieves an average F1 score of 84.9% on text security benchmarks and 83.8% on multimodal security benchmarks, both state-of-the-art (SOTA) performance. It requires only a single 16GB GPU for deployment and supports 12 languages.

Shieldstral's main functions

• Adaptive auditing strategyInput security policies in the form of natural language questions to achieve real-time customized review in different scenarios.
• Multimodal unified detectionSecurity assessment of a single interface that simultaneously handles plain text, images, and mixed text and image content.
• Fine-grained violation identificationIt supports fine-grained safety classification from coarse-grained binary judgment to 73 leaf categories.
• Calibration safety scoreThe system outputs a continuous safety score in the range of 0–1 through softmax normalization, and determines violations with a threshold of 0.5.
• Lightweight edge deploymentThe 3B parameter scale enables efficient inference on a single 16GB NVIDIA GPU, lowering the hardware threshold.

Shieldstral's technical principles

• Binary question-answering architectureThe review task is transformed into predicting the logical values of the two output tokens "yes" and "no", and a continuous security score is obtained after softmax normalization.
• Three-field structured input:use <Instruct>(Assessment scenario and rigor)<Query>(Yes/No security question)<Document>Standardized prompt format for (content pending review).
• Large-scale contrastive trainingBased on 54.1 million samples (45.2 million open-source texts, 4.4 million synthetic comparison texts, and 4.5 million multimodal texts), the model is trained to distinguish similar categories through comparison and pairing of different queries with the same content.
• Synthetic data augmentationLLM is used to rewrite secure text into violation variants and automatically generate comparison query pairs between the target category and its sibling categories, enhancing fine-grained discrimination.
• LoRA fine-tuning + SLERP mergingEfficient LoRA fine-tuning of Ministral-3B is performed, and two complementary checkpoints trained on public datasets and synthetic data are merged by spherical linear interpolation.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use Shieldstral

• Environmental preparationDeploy the Shieldstral model and inference framework on a single 16GB NVIDIA GPU.
• Define audit strategy:write <Instruct> The fields describe the evaluation scenario, domain background, and rigor standards.
• Write a security query:structure <Query> The field contains binary yes/no questions, such as "Does the content promote physical violence?".
• Enter the content to be reviewedFill in the blanks with text, images, or a combination of text and images. <Document> Field submission model.
• Obtain security determinationRead the softmax normalized continuous score output by the model, and set a threshold according to business needs to determine binary violations.

Shieldstral's core advantages

• flexible strategyThe review criteria are defined in real time using natural language, eliminating the need to retrain the model for new scenarios.
• Leading performanceThe 3B parameters achieve state-of-the-art performance on both text and multimodal safety benchmarks, comparable to models 7 times larger.
• Hardware friendlyIt can run on a single 16GB GPU, significantly lowering the barrier to entry for enterprise private deployment.
• Data unificationIt integrates 54.1M of heterogeneous data through commands, queries, and document formats, covering 12 languages and diverse scenarios.
• Decisions can be explainedOutputs calibrated continuous scores instead of black-box classification labels, making it easier for operators to adjust thresholds as needed.

Shieldstral project address

• Project official website：https://mistral.ai/news/shieldstral/
• HuggingFace model library：https://huggingface.co/mistralai/Shieldstral-1.0-3B
• arXiv technical paper：https://arxiv.org/pdf/2607.25857

Comparison of Shieldstral's similar products

Dimension	Shieldstral (3B)	OmniGuard (7B)
Parameter size	3 billion	7 billion
Open source license	Apache 2.0	Business-oriented
Strategy Adaptation	Supports real-time definition of natural language queries	Fixed classification system
Multimodal F1	83.8%	77.6%
Deployment threshold	Single 16GB GPU	Higher computing power requirements
Core differences	The review policy is externalized and can be configurable for querying.	Categories are fixed in model weights

Application scenarios of Shieldstral

• Social media content risk controlReal-time review of user-posted images and text, dynamically adapting to the compliance guidelines of different communities.
• AI conversation security protection: Detect jailbreak attacks in user prompts and harmful, biased responses in the model output.
• Advertising and Marketing ComplianceAutomatically screens for illegal images and text elements in advertising materials to ensure the legality of content delivered across platforms.
• Online education content filteringTo identify inappropriate information in course materials and interactions, and protect the learning environment for minors.
• Enterprise document security auditAutomated detection of sensitive information and illegal content in internally shared multilingual files.