A Simple Approach of Chinese Poetry Generation Using Pre-trained LLMs

This paper explores a simple approach of Chinese poetry generation using pre-trained large language models (LLMs), specifically the Qwen1.5 Model series. Leveraging the capabilities of these advanced models, the authors pre-trained and fine-tuned on customized poetry datasets using LLaMA Factory. The developed model Xuejiu-Poem aims to produce authentic and aesthetically pleasing traditional Chinese poems. Results demonstrate that specialized training can surpass the performance of larger models, such as GPT-4o, in specific poetry generation tasks. The generated poems exhibit strong adherence to traditional formats and stylistic conventions of classical Chinese poetry, underscoring the potential of LLMs in creative applications. This study provides valuable insights into the application of LLMs to literary tasks and suggests promising avenues for future research in classical literature generation.