Comparatively fine-tuning pretrained BERT models on downstream, text classification tasks with different architectural configurations in PyTorch.