The burgeoning field of large language models (LLMs) holds tremendous promise for healthcare, particularly in the realms of medication guidance and adverse drug reaction prediction. Despite this, extant LLMs grapple with managing intricate polypharmacy scenarios. To address these limitations, we introduce ShennongGPT, a cutting-edge LLM, expressly tailored for robust medication guidance and adverse drug reaction forecasting. Our model employs a novel two-stage training strategy: initial learning from distilled drug databases for foundational knowledge on drug interactions, followed by simulation of human-like decision-making processes through the use of real-world patient data, enhancing the relevance and applicability of its guidance. This two-fold approach empowers ShennongGPT to excel in predicting potential adverse drug reactions and offering personalised medication advice, thereby significantly enhancing medication safety and the overall quality of healthcare services. Rigorous evaluations by healthcare professionals and AI experts highlight the superiority of Shennong GPT, which outperforms existing general and specialty LLMs.