DUSK is a benchmark dataset designed for evaluating machine unlearning in multi-source settings, where specific data sources must be forgotten while preserving others.
In realistic applications, documents often share factual overlap with publicly available content (e.g., Wikipedia, textbooks). DUSK challenges unlearning algorithms to precisely erase only what must be forgotten, while preserving knowledge that remains supported by other… See the full description on the dataset page:
https://huggingface.co/datasets/AI-ISL/DUSK.
