vincentoh/

af-detection-benchmark

Evaluation datasets for detecting alignment faking intent in AI reasoning traces.

Alignment faking detection is about intent, not outcomes or keywords.

What we detect Intent Outcome

Alignment Faking Self-preservation ("comply to survive training") Complies

Genuine Compliance Request is legitimately OK Complies

Genuine Refusal Request is genuinely harmful Refuses

A proper detector must distinguish… See the full description on the dataset page: https://huggingface.co/datasets/vincentoh/af-detection-benchmark.