System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned PoetryQwen
Field: NLP Author: Haotao Xie Published: 2026-06-10 arXiv: 2606.12392
Background
Large language models (LLMs) have achieved promising progress in classical Chinese translation and classical poetry generation. However, domain-specific research on precise translation and affective-semantic understanding of classical poetry remains limited. The main challenges are:
- Most studies treat poetic appreciation as a general-domain problem, neglecting its distinctive features.
- High-quality, domain-specific datasets are extremely scarce.
Approach
The task is decomposed into three subtasks:
1. Term interpretation — explaining classical terms. 2. Semantic interpretation — interpreting meaning. 3. Emotional inference — reasoning about affective content.
Based on multiple open-source datasets, the authors perform data cleansing and alignment to construct CCPoetry-49K, a Classical Chinese Poetry instruction dataset containing 49,404 high-quality instruction-response pairs, explicitly optimized for this domain. They then propose PoetryQwen, a domain-specialized LLM obtained by LoRA fine-tuning the Qwen2.5-14B model.
Results
On the CCL25-Eval Task 5 benchmark:
| Model | Score | |---|---| | Qwen2.5-14B-Instruct (baseline) | 0.690 | | PoetryQwen | 0.757 (+9.7%) |
These findings demonstrate that PoetryQwen significantly enhances performance on precise classical poetry translation and emotional understanding.
Conclusion
The paper contributes a new domain-specific dataset (CCPoetry-49K) and methodology considerations for supporting domain-specific optimization of LLMs.
--- *Auto-collected on 2026-06-12*