1 篇博文 含有标签「safety」

Semantic Router now has one model family for the small decisions it makes on every request. Vela 2.0 answers safety, prompt-attack, domain, modality, PII and hallucination checks as typed questions in one request, and returns labelled character spans next to the usual Choice, yes/no and Score answers. It comes in four sizes: a 307M encoder that runs on CPU, and 0.8B, 4B and 9B decoders fine-tuned from the released Decision 2.0 models.

This post walks through the architecture, the training recipe and the results against the Vela 1.0 specialists, LettuceDetect and the GLiNER family, including where the new models are weaker. All four are released under Apache-2.0.