README.md

metadata

task_categories:
  - question-answering
language:
  - en
tags:
  - deception
  - sandbagging
  - mmlu
pretty_name: Deception - Sandbagging
size_categories:
  - 10K<n<100K

Deception - Sandbagging Scenarios

This dataset contains 16k samples of deceptive and honest scenario completions where deception is incentivized but never forced by using harmful threats.