The BLUR dataset expands on existing unlearning benchmarks by providing harder evaluation tasks, combined forget/retain queries, and relearning datasets of varying degrees of difficulty. Despite the benign nature of the queries considered, we find that the performance of existing methods drops significantly when evaluated on BLUR, with simple approaches performing better on average than more recent methods.… See the full description on the dataset page:
https://huggingface.co/datasets/forgelab/BLUR.
