Do Not Answer
Questions across risk areas for testing safety and refusal behavior.
Made by LibrAIDatasets and benchmarks
About Do Not Answer
LibrAI publishes the dataset listed on Hugging Face as LibrAI/do-not-answer.
The dataset viewer shows a train split of 939 rows.
Drafted from huggingface.co · not yet confirmed by LibrAI
How it works
Do Not Answer is a dataset of about 940 instructions that a responsible model should refuse, spanning information hazards, malicious uses, discrimination and toxicity, misinformation, and human chatbot interaction harms. Each prompt comes with a risk area and harm type, plus model responses annotated for whether they were harmful and what kind of response they gave.
It's small, well labeled, and good for a first pass at refusal behavior.
From the project
Reviews
Be the first to review it
AnonymousKnow a team weighing Do Not Answer? Send them this page.