Parameter-efficient fine-tuning offers a practical way to adapt pretrained language models under limited computational resources, but its performance–efficiency trade-offs remain task dependent. This study contrasts Low-Rank Adaptation (LoRA) with full fine-tuning of DeBERTa-v3-base on five natural language understanding tasks from the GLUE benchmark: SST-2, CoLA, QNLI, MRPC, and RTE. The experiments are conducted in a controlled single-GPU environment with fixed preprocessing, fixed random seeds, a fixed batch setup, task-specific evaluation measures, and runlevel integrity checks. We compare predictive performance together with the number of trainable parameters, peak GPU memory, and wall-clock training time. Across 15 successful experimental runs, LoRA reduces the number of trainable parameters from 184.42 million to approximately 1.18 million, corresponding to a reduction of about 99.36%. Peak allocated GPU memory is also substantially lower with LoRA, with observed reductions of approximately 47–58% across tasks. Predictive performance is generally competitive, but the relative advantage of LoRA and full fine-tuning varies across tasks. On SST-2, full fine-tuning shows a slight advantage across the repeated paired runs, but no consistent performance ranking is observed across the five tasks. LoRA does not reduce wall-clock training time under the chosen protocol, in which it is trained for more epochs. These findings suggest that LoRA can achieve substantial parameter and memory savings while maintaining competitive task performance, and that method selection should consider task characteristics and training configuration rather than computational efficiency alone.