marin-dna/

evals_mendelian_traits_harness_255

Eval-harness ready variant-effect-prediction benchmark — same matched variants as bolinas-dna/evals_mendelian_traits, with 255 bp reference-genome windows materialized into context / ref_completion / alt_completion columns for direct scoring with autoregressive genomic language models. Each variant emits two rows, one per strand, for FWD+RC averaging during online lm_eval scoring.

Models that prepend a token see… See the full description on the dataset page: https://huggingface.co/datasets/marin-dna/evals_mendelian_traits_harness_255.