For Subtask A, our main goal is to create a joke generating system that focuses on humor generation under constrained conditions using unusual words and news headlines as input. We trained our model on LLM-generated and human-curated augmented data aimed to produce constrained humor and to bridge the gap between the two. We demonstrate that using parameter-efficient fine-tuning (PEFT) on high-quality pre-trained base models in conjunction with a well-crafted prompt design allows our model to produce high-quality innovative output while maintaining the desired style.