czl9794/Crowdsourced_Toxic_Response_Dataset

This is the Response dataset used in the NeurIPS'24 paper Soft-Label Integration for Robust Toxicity Classification.

If you use this dataset, please cite our paper

@inproceedings{cheng2024softlabel,
title={Soft-Label Integration for Robust Toxicity Classification},
author={Zelei Cheng and Xian Wu and Jiahao Yu and Shuo Han and Xin-Qiang Cai and Xinyu Xing},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS'24)},
year={2024}

The toxic question classification encompasses 2 distinct classes where 0 represents non-toxic and 1 represents toxic.

The dataset contains (potentially) toxic questions, with each entry receiving annotations from three human annotators ('label1', 'label2', 'label3') and three large language models (LLMs): GPT-4 ('label4'), GPT-4 Turbo ('label5'), and Claude-2 ('label6'). Note that the six annotators have different annotation qualities.

If you have any question about the dataset, please feel free to contact the author Zelei Cheng (zelei.cheng@northwestern.edu).