Description
Shieldstral is a sophisticated safety classifier designed to provide flexible content moderation across multiple modalities. With a compact 3B-parameter model, it adapts to various safety policies expressed in natural language, offering a continuous safety score instead of fixed moderation categories. This makes it an ideal solution for real-time moderation applications on edge or low-resource devices.
Read More
Built on the Ministral-3-3B-Base-2512 framework, Shieldstral incorporates a native Pixtral vision encoder, allowing it to handle text-only, image-only, and text+image content seamlessly. It produces its verdict from a single forward pass, making it efficient and effective for various moderation tasks.
Key features of Shieldstral include its policy-adaptive nature, where moderation criteria are supplied as free-form natural-language queries at inference time. This enables a single checkpoint to handle novel safety policies without the need for retraining. Additionally, its multimodal capability allows it to moderate different types of content through a shared interface.
Shieldstral is multilingual, supporting languages such as English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, Arabic, and Russian. It is licensed under the Apache 2.0 License, allowing for both commercial and non-commercial use.
The model is particularly suited for applications like user prompt moderation, model response moderation, and model refusal classification. Its compact size allows it to run on a single GPU, making it accessible for various deployment scenarios.
Despite its capabilities, users should be aware of certain limitations, such as uneven coverage across languages and domains, residual label noise, and reduced reliability with adversarial or obfuscated inputs. However, its open-source nature and adaptability make it a valuable tool for content moderation needs.
- CategoryAI Content Moderation
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Shieldstral Highlights
Policy-adaptive moderation
Multimodal content support
Single-token output
Compact 3B-parameter model
Multilingual support
Apache 2.0 License
Real-time moderation
No retraining required
Getting Started with Shieldstral
Access model: Obtain the model from Hugging Face
Authenticate: Set up necessary authentication
Set up environment: Prepare your system for deployment
Integrate via API: Connect the model to your application
Optimize: Adjust settings for best performance
Shieldstral's Use Cases
- User Prompt Moderation
- Model Response Moderation
- Model Refusal Classification
- Multimodal Content Moderation
- Real-Time Moderation
