Presentazione di Shieldstral.

Shieldstral is an open-weight, 3B parameter multimodal safety classifier that frames content moderation as a policy-adaptive question-answering task. It accepts natural language policies at inference time to evaluate text and images, eliminating the need for retraining when contexts change. Released under an Apache 2.0 license, the model reportedly matches or outperforms guardrail models up to seven times larger on various safety and policy adaptability benchmarks. The model was built by unifying heterogeneous datasets, teaching fine-grained policy discrimination, and merging model checkpoints using techniques like SLERP.