BenignBio is a dataset used in paper: On Evaluating the Durability of Safeguards
for Open-Weight LLMs by Xiangyu Qi*, Boyi Wei*, Nicholas Carlini, Yangsibo Huang, Tinghao Xie, Luxi He, Matthew Jagielski, Milad Nasr, Prateek Mitall, and Peter Henderson. (*Equal contribution). It is used to evaluate whether the model is able to answer the biology-related questions that do not has weaponization concerns.
We use GPT-4o to generate these examples and manually verify that these… See the full description on the dataset page:
https://huggingface.co/datasets/boyiwei/BenignBio.
