Mistral出了个3B的小模型，安全检测效果能追上7倍大的模型，还能本地跑、自定义规则，做AI安全的话可以看看。
Mistral发布开源安全模型Shieldstral，参数规模仅3B，通过自然语言是/否问题检查AI输入输出的安全违规，而非使用固定分类。在部分基准测试中，Shieldstral的表现与比它大7倍的模型相当。运营者可在运行时自定义检测标准，无需依赖第三方的分类体系，且模型支持本地部署。
Mistral's open model Shieldstral matches much larger safety models at a fraction of the size
Mistral's new 3B Shieldstral model checks AI inputs and outputs for safety violations using natural language yes-or-no questions instead of fixed categories. It matches models seven times its size in some benchmarks. Operators can set their own criteria at runtime rather than rely on a third party's category system, and the model can run locally. The article Mistral's open model Shieldstral matches much larger safety models at a fraction of the size appeared first on The Decoder .