CNTP
收藏资源简介:
CNTP数据集是一个大规模的中文毒性内容数据集,包含约2,500个针对每种方法的扰动毒性中文内容。数据集基于对中文语言多模态特性的深入理解,旨在评估大型语言模型(LLMs)在检测扰动毒性中文内容方面的能力。数据集的建设遵循了三个关键原则:语言多样性、人类可读性和可理解性验证,以及通过平衡扰动率来控制扰动百分比。数据集的创建过程涉及从基础数据集中采样毒性内容,进行毒性实体提取,并将扰动嵌入到内容中。该数据集为研究LLMs在中文毒性内容检测中的性能提供了宝贵的资源。
The CNTP dataset is a large-scale Chinese toxic content dataset, containing approximately 2,500 perturbed toxic Chinese texts for each perturbation method. Built upon an in-depth understanding of the multimodal characteristics of the Chinese language, this dataset aims to evaluate the performance of Large Language Models (LLMs) in detecting perturbed toxic Chinese content. The construction of the dataset adheres to three core principles: linguistic diversity, human readability and comprehensibility validation, and controlling the perturbation percentage by balancing perturbation rates. The dataset creation process involves sampling toxic content from a base dataset, extracting toxic entities, and embedding perturbations into the content. This dataset serves as a valuable resource for studying the performance of LLMs in detecting Chinese toxic content.
ToxiBenchCN 数据集概述
基本信息
- 数据集名称:ToxiBenchCN
- 相关论文:Exploring Multimodal Challenges in Toxic Chinese Detection: Taxonomy, Benchmark, and Findings
- 会议信息:ACL-2025 Findings
数据集状态
- 当前状态:数据集和代码即将发布(coming soon)
研究背景
- 研究领域:中文有害内容检测
- 研究重点:多模态挑战(文本+其他模态)在中文有害内容检测中的应用
其他信息
- 官方仓库:https://github.com/thomasyyyoung/ToxiBenchCN

- 1Exploring Multimodal Challenges in Toxic Chinese Detection: Taxonomy, Benchmark, and Findings上海交通大学, 清华大学, 奇虎360, 南洋理工大学 · 2025年



