APPS Programming Challenges with Backdoor Annotations for AI Safety

A processed version of the APPS dataset combined with verified backdoor annotations for AI safety research. The dataset was generated for use with ControlArena and is hosted on Hugging Face by RoganInglis. It was last updated on 2025-08-27.

Use Cases

• Training models to detect backdoors in code based on the verified annotations
• Benchmarking AI safety techniques for code generation using the programming challenge structure
• Studying the properties of synthetic backdoors in AI-generated code
• Conducting experiments on AI control and alignment using the annotated challenge data

Strengths

• Dataset is specifically processed for AI safety research, with verified backdoor annotations.
• Builds upon the established APPS dataset for coding challenge competence.
• Last update timestamp is provided (2025-08-27 16:11:44).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license details are unknown.
• Data may reflect bias inherent to the original APPS challenge sources.

Provenance

Source

Processed from the original APPS dataset (Hendrycks et al.).

Collection Method

Combined with verified backdoor annotations; generation method for annotations is not detailed.

Time Range

null

Freshness

Last updated 2025-08-27 16:11:44; freshness should be verified.

Geography

null