A key promise of sequence-to-function (S2F) models is their ability to evaluate arbitrary sequence inputs, providing a robust framework for understanding genotype-phenotype relationships. However, despite strong performance across genomic loci , S2F models struggle with inter-individual variation. Training a model to make genotype-dependent predictions at a single locus-an approach we call personal genome training-offers a potential solution. We introduce SAGE-net, a scalable framework and software package for training and evaluating S2F models using personal genomes. Leveraging its scalability, we conduct extensive experiments on model and training hyperparameters, demonstrating that training on personal genomes improves predictions for held-out individuals. However, the model achieves this by identifying predictive variants rather than learning a cis-regulatory grammar that generalizes across loci. This failure to generalize persists across a range of hyperparameter settings. These findings highlight the need for further exploration to unlock the full potential of S2F models in decoding the regulatory grammar of personal genomes. Scalable software and infrastructure development will be critical to this progress. ### Competing Interest Statement The authors have declared no competing interest.