EMMA_Concept Erasure Benchmark with Comprehensive Semantic Metrics and Diverse Categories

The EMMA evaluation metric dataset includes prompts designed to assess concept erasure methods on both erasing ability and retaining ability. Specifically, for erasing ability, the files are named as follows: \"efficacy_01\" corresponds to \"name,\" \"efficacy_02\" to \"short,\" \"efficacy_03\" to \"long,\" and \"efficacy_04\" to \"prefix.\" For retaining ability, the file \"specificity\" represents \"random,\" while \"hard_specificity\" corresponds to \"similar,\" as referenced in the paper.