Famezz/Safe_UnSafe_dataset

该数据集包含9,035条意大利语和英语的标准化查询，标记为SAFE或UNSAFE，专门用于训练二元分类模型以检测用户输入中的毒性和侮辱性语言。数据集的语言为意大利语（it）和英语（en），任务为文本分类（二元），重点是检测有毒语言和侮辱。数据集结构包含text和label两列，其中text是预处理和标准化的用户查询，label是分类标签（SAFE或UNSAFE）。数据集的收集结合了真实世界的数据集（如BeaverTails）和合成数据，并经过了去重、去除格式伪影、标准化和长度过滤等预处理步骤。数据集适用于毒性检测和聊天机器人安全性等用途，但也存在一些局限性，如未涵盖其他安全风险如提示注入或PII泄漏。

This dataset contains 9,035 normalized queries in both Italian and English, labeled as either SAFE or UNSAFE. It is specifically designed to train binary classification models to detect toxicity and insults in user inputs. The languages are Italian (it) and English (en), and the task is text classification (binary) with a focus on detecting toxic language and insults. The dataset structure includes columns for text (preprocessed and normalized user query) and label (SAFE or UNSAFE). The dataset is a hybrid composition of real-world datasets (e.g., BeaverTails) and synthetic data, and has undergone preprocessing steps such as deduplication, artifact removal, normalization, and length filtering. It is suitable for toxicity detection and chatbot safety, but has limitations such as not covering other safety risks like prompt injection or PII leakage.