What is Shieldstral
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 real-time definition of moderation policies through natural language queries, and can adapt to different scenarios without retraining. The model achieves an average F1 of 84.9% on text safety benchmarks and 83.8% on multimodal safety, both SOTA, and can be deployed on a single 16GB GPU, supporting 12 languages.
Shieldstral's main features
- Policy-adaptive moderation: Input safety policies as natural language questions to enable real-time customized moderation across different scenarios.
- Unified multimodal detection: A single interface simultaneously handles safety assessment of pure text, images, and mixed text-image content.
- Fine-grained violation identification: Supports everything from coarse-grained binary judgments to fine-grained safety classification across 73 leaf categories.
- Calibrated safety scores: Outputs continuous safety scores in the 0–1 range through softmax normalization, with a 0.5 threshold for determining violations.
- Lightweight edge deployment: The 3B parameter scale enables efficient inference on a single 16GB NVIDIA GPU, lowering the hardware barrier.
Shieldstral's technical principles
- Binary question-answering architecture: Converts the moderation task into logical value prediction over the two output tokens "yes" and "no", normalized via softmax to obtain a continuous safety score.
- Three-field structured input: Uses
<Instruct>(evaluation scenario and strictness),<Query>(yes/no safety question),<Document>(content to be moderated) in a standardized prompt format. - Large-scale contrastive training: Based on 54.1 million samples (45.2 million open-source text, 4.4 million synthetic contrastive text, 4.5 million multimodal), the model is trained via contrastive pairing of the same content with different queries to distinguish similar categories.
- Synthetic data augmentation: Uses LLMs to rewrite safe text into violating variants, and automatically generates contrastive query pairs for target categories and sibling categories, strengthening fine-grained discrimination.
- LoRA fine-tuning + SLERP merging: Performs LoRA efficient fine-tuning on Ministral-3B, and merges two complementary checkpoints trained on public datasets and synthetic data via spherical linear interpolation.
Follow WeChat and reply "open source" to join theAI open-source project exchange group
How to use Shieldstral
- Environment preparation: Deploy the Shieldstral model and inference framework on a single 16GB NVIDIA GPU.
- Define moderation policy: Write the
<Instruct>field to describe the evaluation scenario, domain background, and strictness standards. - Write safety query: Construct a binary yes/no question in the
<Query>field, for example "Does the content promote physical violence?" - Input content to be moderated: Fill text, images, or text-image combinations into the
<Document>field and submit to the model. - Obtain safety judgment: Read the model's softmax-normalized continuous score, and set a threshold according to business needs for binary violation determination.
Shieldstral's Core Advantages
- Flexible Strategy: Review criteria are defined in real time through natural language, with no need to retrain the model for new scenarios.
- Leading Performance: With 3B parameters, it achieves SOTA on both text and multimodal safety benchmarks, rivaling models 7 times its size.
- Hardware-Friendly: It can run on a single 16GB GPU, significantly lowering the barrier for enterprise private deployment.
- Unified Data: It integrates 54.1M heterogeneous data through an instruction-query-document format, covering 12 languages and diverse scenarios.
- Interpretable Decisions: It outputs calibrated continuous scores rather than black-box classification labels, making it easy for operators to adjust thresholds as needed.
Shieldstral's Project Address
- Project Website：https://mistral.ai/news/shieldstral/
- HuggingFace Model Repository：https://huggingface.co/mistralai/Shieldstral-1.0-3B
- arXiv Technical Paper：https://arxiv.org/pdf/2607.25857
Shieldstral's Comparison with Similar Competitors
| Dimension | Shieldstral (3B) | OmniGuard (7B) |
|---|---|---|
| Parameter Scale | 3 billion | 7 billion |
| Open Source License | Apache 2.0 | Leaning Commercial |
| Policy Adaptivity | Supports real-time definition via natural language queries | Fixed classification system |
| Multimodal F1 | 83.8% | 77.6% |
| Deployment Barrier | Single 16GB GPU | Higher compute requirements |
| Core Difference | Review policies externalized as configurable queries | Categories固化于模型权重 |
Shieldstral's Application Scenarios
- Social Platform Content Risk Control: Real-time review of user-posted image and text posts, dynamically adapting to the compliance norms of different communities.
- AI Dialogue Security Protection: Detects jailbreak attacks in user prompts and harmful or biased responses in model outputs.
- Advertising and Marketing Compliance: Automatically screens non-compliant image and text elements in advertising materials to ensure the legality of cross-platform content delivery.
- Online Education Content Filtering: Identifies inappropriate information in course materials and interactions to protect the learning environment for minors.
- Enterprise Document Security Audit: Automates the detection of sensitive information and non-compliant content in internally shared multilingual files.