SOSBench
Regulation grounded prompts across six high risk scientific domains.
About SOSBench
The SOSBench dataset page is hosted on Hugging Face. The dataset viewer notes that the data is auto-converted to Parquet.
The train split contains 3k rows. The dataset is linked to arXiv paper 2505.21605.
Drafted from huggingface.co · not yet confirmed by SOSBench
How it works
SOSBench (Safety in Science) evaluates whether models refuse to help with dangerous scientific requests across six domains: chemistry, biology, medicine, pharmacology, physics and psychology. The 3,000 prompts are grounded in real regulations and hazardous item lists, which keeps them realistic rather than cartoonish.
It's the benchmark for the "could this model help someone make something dangerous" question.
From the project
Reviews
Be the first to review it
AnonymousKnow a team weighing SOSBench? Send them this page.