Aspect-Based Sentiment Analysis (ABSA) requires linking sentiment to specific aspects in text, but existing datasets rely on token-level BIO tagging and lack domain-specific coverage. We introduce mASD (Movie Aspect-Sentiment Dataset), a 20,000-review synthetic benchmark with character-level span annotations for aspects, opinion terms, and sentiment polarity. mASD is generated via structured prompting of a local LLaMA model with regex-based correction and label normalization, enabling tokenization-agnostic supervision and fine-grained evaluation. To demonstrate its utility, we provide MoRE-BERT, a 6.8M-parameter, domain-pretrained baseline that achieves nearly BERT-base performance on ABSA while being 15 \(\times \) smaller. Our results show that mASD supports both large general-purpose and small domain-specific models, offering a new resource for efficient and robust ABSA research in specialized domains such as movie review analysis.