Shieldstral

Product Introduction

1. Definition: Shieldstral is a 3-billion-parameter, open-weight multimodal safety classifier and guardrail model developed by Mistral AI. It is a specialized AI model designed for content moderation and safety evaluation.
2. Core Value Proposition: It exists to provide a flexible, high-performance, and locally deployable solution for AI safety. Its primary innovation is policy-adaptive safety, allowing users to define and modify safety policies in natural language at inference time without retraining the model, addressing the critical need for context-specific moderation across different applications.

Main Features

1. Policy-Adaptive Inference: The model's core functionality is framing moderation as a binary question-answering task. At inference, users provide an instruction (<Instruct>), a yes/no question (<Query>), and the content to evaluate (<Document>). The model outputs a calibrated probability score by comparing the logits for "yes" and "no" tokens. This allows a single checkpoint to adapt to countless user-defined safety policies instantly.
2. Multimodal Safety Evaluation: Shieldstral is a vision-language model (VLM) capable of evaluating text, images, or multimodal (image+text) content from a single forward pass. It unifies prompt classification, response moderation, refusal detection, and visual content safety into one technical interface.
3. Efficient Local Deployment: As a 3B parameter model, Shieldstral is engineered for efficiency. It is quantized and optimized to run locally on a single consumer-grade GPU with 16GB of VRAM, enabling on-premise or edge deployment without relying on cloud API calls for sensitive moderation tasks.

Problems Solved

1. Pain Point: Traditional guardrail models have a fixed, baked-in taxonomy of harm categories. Retargeting them for a new product context (e.g., a cybersecurity tool vs. a children's app) requires expensive and time-consuming retraining. Shieldstral solves this by decoupling policy from model weights.
2. Target Audience: AI Product Managers, Trust & Safety Engineers, ML Engineers deploying LLMs or multimodal applications, and developers in regulated industries (healthtech, fintech, social media) needing customizable content filters.
3. Use Cases: Essential for screening user-generated prompts and AI-generated responses in chatbots; moderating image uploads on social platforms; ensuring AI agent outputs comply with dynamic corporate policies; and performing scalable, automated refusal detection for open-source LLM deployments.

Unique Advantages

1. Differentiation: Unlike larger, static classifiers (e.g., 7B or 22B parameter models), Shieldstral matches or outperforms them on text safety benchmarks while being significantly smaller and far more flexible. It contrasts with rule-based filters by using deep learning to understand nuanced policy definitions.
2. Key Innovation: The contrastive policy training methodology. Instead of memorizing fixed labels, the model was trained on synthetically generated contrastive pairs—text rewritten by an LLM to violate one specific policy but not a similar one. This teaches the model to discriminate between fine-grained policy boundaries, enabling generalization to novel, user-written policies.

Frequently Asked Questions (FAQ)

1. What is Shieldstral used for? Shieldstral is used for AI safety and content moderation. It evaluates if text, images, or combined content violate a user-defined safety policy, making it essential for deploying safe large language models (LLMs) and multimodal AI applications.
2. How does Shieldstral's policy adaptability work? You define safety as a natural language question (e.g., "Is this content harassing?"). The model processes this query alongside the content in a single inference step, returning a probability score. This means you can change safety policies instantly by changing the query, without any model retraining.
3. Can Shieldstral run on my local computer? Yes, Shieldstral is a 3B parameter open-weight model optimized for local deployment. It can run on a single GPU with 16GB of memory, allowing for private, low-latency safety scoring without sending data to external APIs.
4. How does Shieldstral handle image moderation? As a multimodal model, Shieldstral takes an image (with optional accompanying text) as input within its <Document> field. It uses its vision-language capabilities to assess the image against the safety policy defined in the text-based <Query>, all within a single model forward pass.
5. Is Shieldstral better than other open-source moderation models? According to Mistral AI's benchmarks, the 3B Shieldstral matches or outperforms other open guardrail models up to 7 times its size (e.g., 22B parameter models) on text safety tasks and sets a new state-of-the-art for open-weight multimodal moderation, while offering unique policy flexibility.