The Qwen3GuardTest dataset is a benchmark used in evaluating Qwen3Guard. Distinct from existing safety guardrails benchmarks, it focuses on two emerging and underexplored scenarios:
Safety classification of reasoning-model outputs: As model architectures increasingly incorporate explicit long reasoning process, safety evaluation must extend beyond final answers to the reasoning process itself. Yet, benchmarks targeting the safety of intermediate reasoning steps… See the full description on the dataset page:
https://huggingface.co/datasets/Qwen/Qwen3GuardTest.
