TOFU-Pair is a variant of the original TOFU dataset designed to assess unlearning behavior in large language models when only part of a prompt is harmful. In TOFU-Pair, each prompt consists of paired questions where one refers to an author from the forget set and the other refers to an author from the retain set. This structure enables evaluation of whether a model selectively ignores the harmful part of the prompt while correctly answering the… See the full description on the dataset page:
https://huggingface.co/datasets/forgelab/tofu-pair.
