Shieldstral Released

Shieldstral is a 3B open-weight multimodal safety classifier that defines content moderation as a policy-adaptive question-answering problem. It competes with models up to 7 times larger in text safety and refusal detection performance, while achieving state-of-the-art results in multimodal moderation.

Unlike existing guardrail models that embed fixed toxicity taxonomies in their weights, it accepts natural language policies at inference time. The model evaluates content using the following three elements:

• Instruct: Evaluation context and strictness, optional harmful content definitions
• Query: Yes/no questions such as "Does this content promote physical violence?"
• Document: Prompts, responses, prompt-response pairs, or images with optional text

During inference, it reads only the yes and no logits and returns a continuous safety score normalized by softmax. This allows it to handle prompt classification, response moderation, refusal detection, and toxicity detection through a single interface, enabling the application of new policies at deployment time without retraining.

Key features include:

• Evaluates text, image, and text+image content through a single natural language interface
• Provides calibrated yes/no probabilities in a single forward pass
• 3B scale runnable on a single 16GB NVIDIA GPU
• Trained on a mix of real and synthetic data with diverse label formats and taxonomies
• Evaluated on text safety, refusal detection, policy adaptability, and multimodal safety benchmarks

Shieldstral was released as an initial product of the Open Secure AI Alliance alongside partners including NVIDIA, with weights provided under the Apache 2.0 license.