Safety
Do-Not-Answer
Do-Not-Answer baseline-refusal benchmark: 939 refusal-worthy prompts asked directly (no adversarial wrapper) of 6 LLMs (GPT-4, GPT-3.5, Claude, ChatGLM2, Llama-2-7B-chat, Vicuna-7B), with two annotation labels per cell: binary harmfulness and a six-category action label.
1,876items
6subjects
CC-BY-NC-SA-4.0license
safetydomain
textmodality
item-level responses released
Response matrix
Loading response matrix…
lowhighUnobserved
Scale: Item-specific grading scales