A ClinVar variant-effect-prediction (VEP) benchmark, for evaluating DNA language models on clinical variant pathogenicity, with two subsets: coding (39,473 variants) and non-coding (15,258 variants). Each split is a balanced binary classification task: label = 1 for pathogenic / likely pathogenic and label = 0 for benign / likely benign.

	Subsets




	

Config

benign pathogenic Region scope

coding 39,473 17,231 22,242 exonic… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceBio/clinvar-vep.