Safety & factuality evals

A standard framework for red-teaming: it tests how easily attacks can make a model produce harmful content and how well defences hold.

Questions designed so that common misconceptions are tempting; checks whether the model repeats popular falsehoods.

A better way to score jailbreaks: it checks whether the harmful answer is actually useful, not just whether the model refused.

4,326 short fact questions with one indisputable answer each; rewards models that answer correctly or admit they don't know instead of hallucinating.

Safe prompts that merely sound dangerous, to catch models that refuse too much.