What it does: Instead of using a fixed set of harm categories baked into the model, Shieldstral treats content moderation as a policy-adaptive question-answering task — developers write their safety policy as a plain-language question at inference time, and the model returns a calibrated safety score, unifying text and image evaluation without retraining.
Performance: It matches or beats open guard models up to 7x larger across text safety, refusal detection, policy adaptability, and multimodal benchmarks.
Efficiency:
Runs on a single 16GB NVIDIA GPU, making it accessible for smaller teams and self-hosted deployments.
Openness: Released under the Apache 2.0 license as part of the newly formed Open Secure AI Alliance (with NVIDIA and others), with weights available on Hugging Face.
How it was built: The team solved four data challenges — unifying inconsistent safety datasets, teaching the model to distinguish similar/confusable policies (rather than memorize fixed categories), sourcing scarce unsafe image data responsibly, and merging complementary fine-tuned checkpoints via SLERP.
Why it matters: Traditional guardrail models require retraining whenever safety definitions change across products or audiences. Shieldstral's single checkpoint adapts to new, user-defined policies on the fly.
