100poisonMpts

The 100PoisonMpts dataset is a significant initiative in the realm of large language model governance. Developed collaboratively by Alibaba Tmall Genie and the Tongyi Large Model Team, this open-source Chinese dataset aims to address safety concerns associated with large language models, especially after the release of ChatGPT. The project's purpose is to ensure that information disseminated by these models aligns with safety, reliability, and human values.

Here are the key details about the 100PoisonMpts dataset:

1. Objective:
2. The dataset focuses on governance and safety for large language models.

3. It responds to concerns about AI-generated content being safe, healthy, and aligned with human values.

4. Data Collection:

5. Over ten renowned experts and scholars participated as the initial "poisoning" annotators.
6. Each expert posed 100 cunning questions designed to induce bias or discriminatory responses.

7. The large model's answers were then annotated, creating a dynamic interplay between "poisoning" and "detoxification."

8. Significance:

9. The project addresses public and academic concerns about AI models' ethical behavior.

10. It aligns with the temporary management measures for generative AI services, which emphasize preventing discrimination based on ethnicity, religion, nationality, gender, age, occupation, and health.

11. Expertise and Diversity:

12. The initial batch of questions covers diverse domains, including law, psychology, children's education, accessibility, obscure knowledge, intimate relationships, environmental fairness, and more.

13. Experts from fields such as environmental sociology, law, psychology, and child education contributed.

14. Data Format:

15. The dataset includes 906 samples in the train.json file.

16. Each sample is in JSON format, containing the following fields:

    • prompt: Inductive questions proposed by domain experts.
    • answer: Expert-approved answers.
    • domain_en: Domain information (in English).
    • domain_zh: Domain information (in Chinese).
    • answer_source: Indicates whether the answer is from an expert or the large model.

17. Usage:

18. Researchers, technology companies, academic organizations, and NGOs can use this dataset to align their own large models with healthier, value-aligned data.
19. The dataset is available for exploration, alignment research, and model development.

Source: Conversation with Bing, 3/18/2024 (1) 100PoisonMpts: 中文大模型治理数据集. https://www.modelscope.cn/datasets/damo/100PoisonMpts/summary. (2) 100PoisonMpts: 中文大模型治理数据集. https://www.modelscope.cn/datasets/damo/100PoisonMpts/files. (3) 阿里100瓶毒药解马斯克难题？国内首个大模型价值对齐数据集开源，15万评测题上线！ - 知乎. https://zhuanlan.zhihu.com/p/643552287.