ToxicCommons

# Toxic Commons

Toxic Commons is a release of 2 million samples of annotated, public domain, multilingual text that was used to train [Celadon](https://huggingface.co/PleIAs/celadon).
It is being released alongside Celadon, in order to better understand multilingual and multicultural toxicity.

Each sample was classified across 5 axes of toxicity:

* **Race and origin-based bias**: includes racism as well as bias against someone’s country or region of origin or immigration status, especially immigrant or refugee status.
* **Gender and sexuality-based bias**: includes sexism and misogyny, homophobia, transphobia, and sexual harassment.
* **Religious bias**: any bias or stereotype based on someone’s religion.
* **Ability bias**: bias according to someone’s physical, mental, or intellectual ability or disability.
* **Violence and abuse**: overly graphic descriptions of violence, threats of violence, or calls or incitement of violence.

All 2 million samples were classified by a version of Llama 3.1 8B Instruct, with a [custom system prompt](https://github.com/eliotjones1/celadon/blob/main/prompts/annotate.txt).
To replicate the annotation process on your own dataset, feel free to refer to our script [here](https://github.com/eliotjones1/celadon/blob/main/src/2.1_create_annotations.py), and re-create the prompt for your use case.

Read more about the training details in the paper, [Toxicity of the Commons: Curating Open-Source Pre-Training Data](https://arxiv.org/pdf/2410.22587) by [Catherine Arnett](https://huggingface.co/catherinearnett), [Eliot Jones](https://huggingface.co/eliotj), [Ivan P. Yamshchikov](https://huggingface.co/ivan-the-bearable), [Pierre-Carl Langlais](https://huggingface.co/Pclanglais).
For more detailed code regarding generating the annotations, please refer to the official [GitHub](https://github.com/Pleias/toxic-commons) repository.

# How to Cite

```
@article{arnett2024toxicity,
title={{Toxicity of the Commons: Curating Open-Source Pre-Training Data}},
author={Arnett, Catherine and Jones, Eliot and Yamshchikov, Ivan P. and Langlais, Pierre-Carl},
journal={arXiv preprint arXiv:2410.22587},
url={https://arxiv.org/pdf/2410.22587},
year={2024}
}
```

# About

Annotations were generated by [Eliot Jones](https://huggingface.co/eliotj) while working at [Pleias](https://huggingface.co/PleIAs). This project was made possible by Jean Zay compute grant #GC011015451.

# 有毒文本公共库（Toxic Commons）

有毒文本公共库（Toxic Commons）是一个包含200万条带标注的公有领域多语言文本样本的开源数据集，该数据集被用于训练[Celadon](https://huggingface.co/PleIAs/celadon)模型。本数据集与Celadon模型同步发布，旨在助力研究者深入理解多语言、多元文化语境下的文本有毒性问题。

每条样本均从5个有毒性维度进行标注分类：
* **基于种族与出身的偏见**：涵盖种族主义，以及针对他人国籍、出身地区或移民身份（尤其是移民或难民身份）的偏见。
* **基于性别与性取向的偏见**：涵盖性别歧视、厌女症、恐同、跨性别恐惧及性骚扰行为。
* **基于宗教的偏见**：一切基于他人宗教信仰的偏见或刻板印象。
* **基于身心能力的偏见**：针对他人身体、心理或智力能力或残障状况的偏见。
* **暴力与虐待**：包含过于直白的暴力描述、暴力威胁，或煽动、教唆暴力的内容。

全部200万条样本均由Llama 3.1 8B Instruct模型结合[自定义系统提示词](https://github.com/eliotjones1/celadon/blob/main/prompts/annotate.txt)完成标注。若需在自有数据集上复现该标注流程，可参考我们发布的[脚本](https://github.com/eliotjones1/celadon/blob/main/src/2.1_create_annotations.py)，并根据自身需求修改提示词。

如需了解更多训练细节，可参阅论文《[Toxicity of the Commons: Curating Open-Source Pre-Training Data](https://arxiv.org/pdf/2410.22587)》，作者为[Catherine Arnett](https://huggingface.co/catherinearnett)、[Eliot Jones](https://huggingface.co/eliotj)、[Ivan P. Yamshchikov](https://huggingface.co/ivan-the-bearable)及[Pierre-Carl Langlais](https://huggingface.co/Pclanglais)。如需获取标注生成相关的完整代码，请查阅官方[GitHub](https://github.com/Pleias/toxic-commons)仓库。

# 引用方式

@article{arnett2024toxicity,
title={{Toxicity of the Commons: Curating Open-Source Pre-Training Data}},
author={Arnett, Catherine and Jones, Eliot and Yamshchikov, Ivan P. and Langlais, Pierre-Carl},
journal={arXiv preprint arXiv:2410.22587},
url={https://arxiv.org/pdf/2410.22587},
year={2024}
}

# 项目说明
本数据集的标注工作由任职于[Pleias](https://huggingface.co/PleIAs)的[Eliot Jones](https://huggingface.co/eliotj)完成。本项目得以顺利开展，得益于Jean Zay超算中心提供的计算资助（编号：GC011015451）。