System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5

Researchers have developed PoetryQwen, a specialized language model fine-tuned for classical Chinese poetry analysis, along with a new 49,404-pair dataset called CCPoetry-49K. The model achieves 9.7% performance improvement over baseline Qwen2.5, demonstrating the effectiveness of domain-specific optimization for nuanced linguistic tasks.

This research represents a meaningful advancement in domain-specialized AI development, showcasing how targeted datasets and fine-tuning techniques can dramatically improve language model performance on highly specialized tasks. The creation of CCPoetry-49K and PoetryQwen addresses a genuine gap in AI capabilities: while general-purpose LLMs have made impressive strides, they often struggle with culturally and linguistically nuanced domains like classical poetry that require simultaneous understanding of lexical meaning, semantic depth, and emotional resonance.

The decomposition of poetic appreciation into three distinct subtasks—term interpretation, semantic interpretation, and emotional inference—reflects sophisticated task design that acknowledges the multifaceted nature of poetry understanding. This methodological approach transcends the specific poetry domain and offers a template for tackling other culturally specific or emotionally complex linguistic challenges. The 9.7% improvement over the Qwen2.5-14B-Instruct baseline, while seemingly modest numerically, represents substantial gains in subjective quality for a task where precision matters significantly.

From an industry perspective, this work demonstrates that open-source LLM ecosystems thrive when researchers invest in domain specialization rather than pursuing only general-purpose scale. For developers working with Asian language models or cultural applications, this establishes that LoRA-based fine-tuning remains cost-effective for achieving meaningful performance gains without requiring massive computational resources. The public release of CCPoetry-49K benefits the research community by providing a benchmark for evaluating classical Chinese understanding and emotional intelligence in language models.

• →PoetryQwen achieves 0.757 score on CCL25-Eval Task 5, representing 9.7% improvement over Qwen2.5-14B baseline through LoRA fine-tuning
• →CCPoetry-49K dataset of 49,404 instruction-response pairs provides domain-specific training material for classical Chinese poetry tasks
• →Task decomposition into term, semantic, and emotional interpretation enables more nuanced approaches to complex linguistic understanding
• →Results validate that specialized fine-tuning on high-quality datasets outperforms general-purpose models for culturally-specific language tasks
• →Open-source LLM optimization for regional/cultural domains offers practical alternatives to scaling general models