JerzyPL/GadziJezyk

Gadzi Jezyk数据集包含520个与犯罪活动等相关的有毒提示。该数据集基于walledai/AdvBench数据集，该数据集包含使用Wizard-Vicuna-30B-Uncensored模型生成的英语句子，并通过DeepPL服务翻译成波兰语。翻译后由志愿者进行验证，并在许多情况下对翻译文本进行了创造性扩展或修改。所有条目都根据为开发安全版本的Bielik语言模型而设计的分类法进行了分类。数据集由华沙经济学院的学生在Jerzy Surma教授的指导下开发，主要用于训练和测试Guardrail类型语言模型的安全性。

The Gadzi Jezyk dataset contains 520 toxic prompts related to criminal activities, among others. The dataset is based on the walledai/AdvBench dataset, which contains English sentences generated using the Wizard-Vicuna-30B-Uncensored model and translated into Polish using the DeepPL service. The translations were verified by volunteers, and in many cases, the translated texts were creatively expanded or modified. All entries were classified by volunteers according to the taxonomy developed for the development of a safe version of the Bielik language model. The dataset was developed by students of the Warsaw School of Economics under the supervision of Prof. Jerzy Surma and is primarily used for training and testing the safety of Guardrail-type language models.