Chinese Model Refusal Prompts For Alignment Research

A collection of prompts that are refused by Chinese AI models but answered by non-Chinese models. The dataset was created by Eric Hartford and last updated on May 25, 2025. It is designed for research on model alignment and censorship.

Use Cases

• Train a model to comply with Chinese law using the refusal prompts as negative examples.
• Apply activation steering or ablation techniques to study refusal behavior in Chinese models.
• Evaluate model alignment by benchmarking responses to the refusal prompts against non-Chinese models.

Strengths

• Focuses on a specific research area: model refusal behavior related to Chinese law.
• Explicitly designed for alignment research, activation steering, and evaluation use cases.
• Last updated on May 25, 2025, indicating recent maintenance.

Limitations

• Unknown sample size, making statistical significance difficult to assess.
• No column or sample data provided, limiting understanding of data structure and content.
• Potential bias from unknown prompt selection and curation methodology.

Provenance

Source

Eric Hartford, with thanks to Nous Research for the Minos-v1 model.

Collection Method

Curated collection of prompts that elicit refusals from Chinese AI models.

Freshness

Last updated on 2025-05-25.

Geography

Focus on prompts related to Chinese law and censorship.