In this study, we examine sentiment swapping in Arabic dialects utilizing large language models (LLMs), which is a task that requires the reversal of sentiment polarity while preserving meaning and dialect. Five Arabic sentiment datasets were filtered and validated to construct a high-quality dialectal corpus, and multiple 7B to 13B models were evaluated under zero- and five-shot prompts. The results indicate that few-shot examples consistently improve performance across all models. Qwen2.5-7B-Instruct achieved the strongest sentiment inversion, whereas ALLaM-7B-Instruct-preview demonstrated the best dialect preservation and semantic similarity. Overall performance remained moderate, highlighting the difficulty of polarity-controlled rewriting in dialectal Arabic. This study provides a benchmark for future Arabic generative natural language processing research, as well as a carefully selected sentiment-swapped dataset.