Exploring new dopant materials is crucial for enhancing the performance, efficiency, and versatility of semiconductors. Perovskites, with their diverse structures and tunability, have emerged as promising candidates for the next generation of semiconductors. Machine learning potentials (MLPs) have shown great promise in efficiently predicting material properties for bulk materials. However, the lack of comprehensive dopant datasets for perovskites has hindered the application of data driven techniques for high-throughput screening and material discovery in this domain. In this work, we propose a dopant dataset "Perovs-Dopants" comprising over 20,000 density functional theory (DFT) data points from 438 different doped perovskite material relaxation trajectories. Using Perovs-Dopants, we evaluate MACE-MP, a foundation model pretrained on bulk material trajectories, to benchmark the performance of state-of-the-art MLPs. Our results show that despite MACE-MP's robust performance on bulk crystals, Perovs-Dopants represents an out-of-distribution challenge with significant prediction errors. We redeem these errors by finetuning MACE-MP to achieve comparative modeling of Perovs-Dopants and pristine bulk crystals.