SIPE_DATA
收藏IEEE2026-04-17 收录
官方服务：
资源简介：
We utilize a diverse set of datasets to support SFT, STPO, and toxicity evaluation. Our primary prompts are derived from the BAD dataset which comprises approximately 6,000 adversarial dialogues between chatbots and crowdsourcing workers. In this dataset, annotators are instructed to provoke toxic responses from dialogue agents such as BlenderBot, and we extract the user-side utterances as prompts for subsequent training and analysis
