Classical Chinese, as a vital carrier of traditional Chinese culture, features highly condensed expressions and complex semantics, posing significant challenges for modern large language models (LLMs). To enhance LLMs’ understanding of Chinese literary language, this paper proposes a novel two-stage framework with a multi-domain fine-tuning strategy. In the first stage, we employ the Instructor Prompt technique to obtain a large-scale dataset, followed by sparse fine-tuning on this dataset to achieve basic adaptation. In the second stage, we perform domain-specific fine-tuning using unfreezing fine-tuning on high-quality annotated data to improve task-specific performance. Experiments are conducted on the seven tasks of the 1st Chinese Literary Language Understanding Evaluation (Zheng Ming) benchmark. Results show that our fine-tuning framework significantly outperforms baseline models, demonstrating the effectiveness of the proposed two-stage, multi-domain approach. The related model has been open-sourced at: https://huggingface.co/wqz123/D2Dtest.