With the rise of social media, users increasingly share their medication experiences online, offering a new data source for real-time ADR detection. However, ADR detection on social media faces two key challenges: limited labeled data and an imbalance between positive and negative samples. While previous studies have explored solutions like transfer learning, multi-source data fusion, joint task training, and loss function optimization, these approaches can introduce noise, increase annotation costs, or complicate training complexity. Moreover, despite the promising zero-shot and few-shot capabilities of large language models (LLMs) in natural language processing tasks, their performance in social media-based ADR detection remains below that of smaller, fine-tuned models. To tackle these challenges, we propose the Bal-LLaMA framework, comprising three modules: a data augmentation module to balance positive and negative samples as well as mitigating the challenge posed by limited annotated data, an instruction data construction module tailored for social media ADR detection, and a QLoRA-based module for efficient parameter fine-tuning. Experimental results demonstrate that Bal-LLaMA significantly outperforms existing state-of-the-art models on various social media ADR detection datasets, confirming the effectiveness of our approach.