Control-Tax: Model-Generated Backdoor Solutions for AI Safety

Control-Tax is a dataset containing model-generated backdoor solutions and exploits for AI safety research. The dataset is based on problems from the APPS dataset and was generated for use with ControlArena. It was created by RoganInglis and last updated on August 27, 2025.

Use Cases

• Training detection models for backdoor exploits based on the generated solutions.
• Benchmarking the robustness of code-generating AI systems against adversarial attacks.
• Studying the nature of model-generated backdoors for AI safety research.

Strengths

• Dataset is explicitly designed for a specific research area: AI safety and backdoor attacks.
• Derived from the established APPS dataset, providing a known problem foundation.
• Created for compatibility with the ControlArena framework.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Hugging Face user RoganInglis, based on the APPS dataset and the Control-Tax research framework.

Collection Method

Model-generated backdoor solutions and exploits.

Freshness

Last updated 2025-08-27 16:47:06; freshness should be verified.