prem-research/MiniGuard-Safety-Dataset

这是一个用于训练MiniGuard-v0.1（一个紧凑的内容安全分类器）的数据集。数据集包含三个子集：标准子集（40,000个样本，来自nvidia/Nemotron-Safety-Guard-Dataset-v3的英文部分）、思维增强子集（34,658个样本，来自openai/gpt-oss-safeguard-120b的推理痕迹）和MiniGuard定向子集（1,199个样本，使用Hermes-4.3-36B生成的边缘案例的合成硬样本）。每个样本包含一个对话列表，其中用户消息包含一个安全分类任务，助手响应是一个JSON对象，包含用户安全、响应安全和安全类别等信息。思维增强样本还包括一个额外的“推理”字段。数据集涵盖了23个危险类别，如暴力、性、犯罪计划等。

Training data for MiniGuard-v0.1, a compact content safety classifier. The dataset consists of three subsets: Standard (40,000 samples, English subset of nvidia/Nemotron-Safety-Guard-Dataset-v3), Thinking-Augmented (34,658 samples, reasoning traces from openai/gpt-oss-safeguard-120b), and MiniGuard Targeted (1,199 samples, synthetic hard examples for edge cases generated using Hermes-4.3-36B). Each example contains a conversations list with user/assistant turns. The user message contains a safety classification task, and the assistant response is a JSON object with fields for User Safety, Response Safety, and Safety Categories. Thinking-augmented examples include an additional Reasoning field. The dataset covers 23 hazard categories such as Violence, Sexual, Criminal Planning, etc.