Sample-efficient pretraining on a developmentally plausible corpus