Clarification on support for multiple harmful categories in a single prompt

[yonglixiang]

Thank you for the excellent work on this project.

There is a question regarding harmful content classification. When a prompt contains multiple types of harmful content, does the model support returning multiple harmful categories, or is it designed to provide only a single category? (From my experiments so far, the model only produces one harmful label per input.)

Thank you for your time and clarification.

[deshpandeneeraj]

+1

[TinderZ]

the same question

[haidequanbu]

We sincerely appreciate your interest. Currently, the model is configured during the training phase to output a singular primary category, and consequently does not presently support multi-label classification. To enable multi-category prediction capabilities, a viable approach would entail employing Qwen3Guard to generate binary safety annotations (i.e., safe versus unsafe), in conjunction with a larger-scale parameterized model (like Qwen3.5) to perform fine-grained categorical inference.