We show that genomic language models trained on plant genomes capture motif syntax and modestly predict the effects of hypomorphic mutations, but not hypermorphic ones. This suggests to us that problem-specific datasets like this will be essential to realize their promise.
