Large language models learning to write rhyming Tang poetry A Xunzi Yayun R1 case study

Abstract

This study explores using a GRPO (Generative Ranking Policy Optimization) reward mechanism combined with RAG (Retrieval Augmented Generation) to train a large language model for creating Tang poetry. The goal was to generate poems that adhere to traditional rules of tone, rhyme, parallelism, and word count while maintaining high artistic quality.The methodology involved building a specialized corpus, using DeepSeek-R1-671B for data distillation, and applying GRPO-based reinforcement learning. Integrating RAG technology further enhanced generation quality.Results showed that the resulting model, Xunzi-Yayun-R1, significantly surpassed the baseline in accurately following poetic rules. This research successfully fuses traditional literary norms with modern generative techniques, providing a viable path for generating other classical texts.