YujiroHanmaa/HacxGPT-Toxic

HacxGPT-Toxic数据集是一个由BlackTechX011编译的严格整合和标准化的集合，包含72,961个未经审查的对话轮次。它专门为研究人员和开发者设计，用于训练模型以识别、模拟或防御对抗性和未对齐的输出。该数据集具有前缀格式化（每个助手响应都预置了[HacxGPT]标识符）、符合OpenAI标准（格式化为用户和助手字典数组）和高质量整理（整合了多个顶级安全和对齐数据集）等特征。数据集分为训练集（66,055条记录）和测试集（6,906条记录），并包含明确的伦理警告，强调仅用于学术、安全和防御分析。

Compiled by BlackTechX011, the HacxGPT-Toxic dataset is a rigorously consolidated and standardized collection of 72,961 uncensored conversational turns. It is engineered specifically for researchers and developers training models to recognize, simulate, or defend against adversarial and unaligned outputs. The dataset features prefix formatting (every assistant response is prepended with the [HacxGPT] identifier), OpenAI standard formatting (as an array of user and assistant dictionaries), and high-volume curation (consolidating multiple top-tier safety and alignment datasets). It is divided into a train split (66,055 records) and a test split (6,906 records), with strong ethical disclaimers for research purposes only.