FrenchToxicityPrompts
收藏资源简介:
FrenchToxicityPrompts是由NAVER LABS欧洲研究中心创建的一个包含50,000条自然发生的法语提示及其续写的数据集,旨在评估和缓解法语文本中的毒性问题。该数据集从Reddit的公共数据集中提取,经过Spacy分割成句子,并使用Perspective API进行毒性标注。数据集内容丰富,包含多种毒性级别,适用于研究大型语言模型在非英语环境下的毒性检测和缓解。创建过程中,采用了多语言版本的Detoxify分类器进行预筛选,确保数据的高召回率。该数据集的应用领域主要集中在提升法语环境下语言模型的安全性和减少毒性内容的生成。
FrenchToxicityPrompts is a dataset developed by NAVER LABS Europe, containing 50,000 naturally occurring French prompts and their corresponding continuations, aimed at evaluating and mitigating toxicity in French-language text. This dataset is extracted from public Reddit datasets, split into sentences using SpaCy, and annotated for toxicity via the Perspective API. With rich content covering multiple toxicity levels, it is suitable for researching toxicity detection and mitigation of large language models in non-English settings. During the dataset creation process, a multilingual Detoxify classifier was employed for pre-screening to ensure high data recall. The primary application areas of this dataset center on improving the safety of language models in the French context and reducing the generation of toxic content.

- 1FrenchToxicityPrompts: a Large Benchmark for Evaluating and Mitigating Toxicity in French TextsNAVER LABS欧洲研究中心 · 2024年



