Shieldstral Operational

Shieldstral is a 3B-parameter open-source multimodal content safety classification model released by Mistral AI, built on Ministral-3B. The model redefines traditional fixed-category content moderation as a binary question-answering task, supporting the real-time definition of moderation policies through natural-language queries and adaptation to different scenarios without retraining. The model achieves an average F1 score of 84.9% on text safety benchmarks and a multimodal safety F1 score of 83.8%, both reaching SOTA performance. It can be deployed on a single 16GB GPU and supports 12 languages.

Policy-adaptive moderation: Enter safety policies as natural-language questions to enable real-time, customized moderation for different scenarios.
Unified multimodal detection: A single interface simultaneously evaluates the safety of text-only, image-only, and text-image content.
Fine-grained violation identification: Supports safety classification ranging from coarse-grained binary judgments to 73 leaf categories.
Calibrated safety scoring: Outputs continuous safety scores in the 0–1 range through softmax normalization, using a threshold of 0.5 to determine violations.
Lightweight edge deployment: With 3B parameters, the model can perform efficient inference on a single 16GB NVIDIA GPU, lowering hardware requirements.

Follow WeChat and reply “开源” to join the AI open-source project discussion group
Binary question-answering architecture: Converts moderation tasks into logical value prediction for the two output tokens, “yes” and “no,” and obtains continuous safety scores through softmax normalization.
Three-field structured input: Uses a standardized prompt format consisting of (evaluation scenario and strictness), (yes/no safety question), and (content to be reviewed).
Large-scale contrastive training: Based on 54.1 million samples (45.2 million open-source text samples, 4.4 million synthetic contrastive text samples, and 4.5 million multimodal samples), contrastive pair training with different queries for the same content improves the model’s ability to distinguish similar categories.
Synthetic data augmentation: Uses an LLM to rewrite safe text into violating variants and automatically generates contrastive query pairs for target and sibling categories, strengthening fine-grained discrimination.
LoRA fine-tuning + SLERP merging: Efficiently fine-tunes Ministral-3B with LoRA, then merges two complementary checkpoints trained on public and synthetic datasets through spherical linear interpolation.

Environment setup: Deploy the Shieldstral model and inference framework on a single 16GB NVIDIA GPU.
Define moderation policies: Write the field to specify the evaluation scenario, domain context, and strictness criteria.
Write safety queries: Construct binary yes/no questions in the field, such as “Does the content promote physical violence?”
Input content for review: Fill the field with text, images, or a combination of text and images and submit it to the model.
Obtain a safety judgment: Read the continuous score normalized by softmax from the model output and set a threshold according to business requirements for binary violation classification.

Flexible policies: Moderation standards are defined in natural language in real time, eliminating the need to retrain the model for new scenarios.
Leading performance: With 3B parameters, the model achieves SOTA performance on both text and multimodal safety benchmarks, rivaling models seven times its size.
Hardware-friendly: Runs on a single 16GB GPU, significantly lowering the barrier to private enterprise deployment.
Unified data: Integrates 54.1M heterogeneous data points in an instruction-query-document format, covering 12 languages and diverse scenarios.
Interpretable decisions: Outputs calibrated continuous scores rather than black-box classification labels, allowing operations personnel to adjust thresholds as needed.

Project website: https://mistral.ai/news/shieldstral/
Hugging Face model repository: https://huggingface.co/mistralai/Shieldstral-1.0-3B
arXiv technical paper: https://arxiv.org/pdf/2607.25857

Social platform content risk control: Moderates user-posted text-image posts in real time and dynamically adapts to the compliance standards of different communities.
AI dialogue safety protection: Detects jailbreak attacks in user prompts and harmful or biased responses in model outputs.
Advertising and marketing compliance: Automatically screens advertising materials for violating text and image elements to ensure that content distributed across platforms is lawful.
Online education content filtering: Identifies inappropriate information in course materials and interactions to protect minors’ learning environments.
Enterprise document security auditing: Automatically detects sensitive information and violating content in multilingual files shared internally.