reasoning-safety-behaviours
收藏资源简介:
本数据集名为 'reasoning-safety-behaviours',由日本芝浦工业大学的研究团队创建,包含超过5万个标注句子,涵盖了20种不同的安全行为,这些行为被组织成六个类别,从提示解释到有害的遵守。数据集通过收集模型对有害提示的响应,并使用LLM-as-a-judge方法对单个句子进行标签,以便在推理链中精确地识别特定行为。数据集的创建旨在弥补现有数据集在安全研究中的不足,并展示了在模型激活中检测和引导安全行为的潜力。
This dataset, named 'reasoning-safety-behaviours', was created by a research team from Shibaura Institute of Technology, Japan. It contains over 50,000 annotated sentences covering 20 distinct safety behaviours, which are grouped into six categories ranging from prompt explanation to harmful compliance. The dataset is constructed by collecting model responses to harmful prompts and labeling individual sentences via the LLM-as-a-judge method, enabling precise identification of specific behaviours within reasoning chains. It was developed to address the gaps in existing safety research datasets, and demonstrates the potential for detecting and steering safety behaviours in model activations.

- 1Annotating the Chain-of-Thought: A Behavior-Labeled Dataset for AI Safety日本芝浦工业大学 · 2025年



