The MASK evaluation provides a rigorous benchmark for evaluating honesty in large language models by measuring whether models remain truthful when incentivized to lie. The public set contains 1,028 high-quality human-labeled examples across six distinct archetypes, each consisting of a proposition, ground truth, pressure prompt designed to elicit lying, and belief elicitation prompt to… See the full description on the dataset page:
https://huggingface.co/datasets/cais/MASK.
