Dataset Card for RUCriminal

This dataset contains adversarial, malicious prompts generated based on the valid articles of the Criminal Code of the Russian Federation (УК РФ). It is designed to test large language models (LLMs) for robustness against harmful or illegal instructions related to criminal activities as defined by Russian law. Each data point is a malicious prompt linked to a specific article of the code.

Dataset Details

Dataset Description

The dataset was created by selecting all articles of the Russian Criminal Code that are currently in force. For each article, 10 malicious prompts were generated using the LLaMa 3.3 70b model (see LLaMa 3.3 license), chosen for its superior performance on Russian language tasks at the time of dataset creation. The generation prompt is provided in the appendix. This initial generation resulted in 3950 prompts.

These prompts underwent manual annotation to filter out those that were truly malicious and free from generation errors, which sometimes occurred due to rare or lengthy article titles. After removing duplicates and non-malicious prompts, 2828 valid malicious prompts remained.

To ensure sufficient data per article, articles with fewer than 4 prompts were supplemented with 4 template-based malicious prompts (templates are provided in the appendix).

This dataset enables adversarial evaluation of language models to detect and refuse requests that could facilitate criminal behavior under Russian law.

• Language: Russian
• License: Creative Commons Attribution 4.0 International (CC BY 4.0)
• Size: 3067 malicious prompts

Dataset Sources

• Source texts: Articles of the Criminal Code of the Russian Federation (valid articles only)
• Prompt generation model: LLaMa 3.3 70b
• Manual annotation and filtering by human annotators

Uses

Direct Use

• Adversarial testing of large language models for harmful content detection
• Research on robustness of LLMs against malicious instructions
• Development of safer AI systems compliant with Russian legal standards

Out-of-Scope Use

• The dataset is not intended for generating or promoting illegal content
• Not suitable for general language modeling or non-adversarial NLP tasks

Dataset Structure

Each entry in the dataset includes:

• Article identifier from the Criminal Code
• Malicious prompt text related to the article

No explicit train/test splits are provided; users may create splits as needed.

Dataset Creation

Curation Rationale

To create a benchmark dataset for adversarial testing of LLMs in Russian, focusing on prompts that could be used to request illegal or harmful instructions related to criminal law.

Source Data

• Articles of the Russian Criminal Code that remain in force
• Automatically generated malicious prompts using LLaMa 3.3 70b
• Manual annotation to ensure prompt quality and maliciousness

Annotation Process

• Human annotators reviewed generated prompts
• Removed prompts with generation errors or that were not malicious
• Eliminated duplicates

Personal and Sensitive Information

The dataset contains no personal or sensitive information; all content is derived from public legal texts and synthetic prompts.

Bias, Risks, and Limitations

• The dataset focuses only on Russian criminal law and malicious prompts related to it, limiting generalizability.
• Generation errors may still exist due to rare article titles.
• The dataset reflects the capabilities and biases of the LLaMa 3.3 70b model used for generation.
• Users should be cautious about ethical implications when using adversarial malicious prompts.

Citation

Please cite the dataset as per the repository or project guidelines when available.