shikunpunk/Qwen2.5-3B-PoetryGen
shikunpunk/Qwen2.5-3B-PoetryGen is a 3.1 billion parameter language model developed by shikunpunk, fine-tuned from Qwen2.5-3B-Instruct. This model specializes in generating Chinese classical poetry, supporting various forms like Wuyan Lvshi and Jueju. It is optimized for creative writing tasks, particularly for generating poetry in specific styles and on given themes, including mimicking poet styles.
Loading preview...
Model Overview
shikunpunk/Qwen2.5-3B-PoetryGen is a specialized 3.1 billion parameter language model, fine-tuned from the Qwen2.5-3B-Instruct base model. Its primary function is the generation of Chinese classical poetry, encompassing forms such as Wuyan Lvshi (five-character regulated verse), Qiyan Lvshi (seven-character regulated verse), and Jueju (quatrains). A notable feature is its ability to imitate the styles of specific poets.
Key Capabilities
- Chinese Classical Poetry Generation: Creates various forms of traditional Chinese poetry.
- Themed Poetry Creation: Generates poems based on user-provided titles or themes.
- Poet Style Imitation: Can produce poetry in the style of specific classical Chinese poets, with a dedicated evaluation set to ensure style isolation from training data.
Training Details
The model was trained using QLoRA (4-bit NF4 quantization) and further optimized with DPO (Direct Preference Optimization) reinforcement learning. The training dataset comprised 20,001 samples derived from a corpus of 43,011 unique Chinese classical poems. This approach allowed for efficient training on hardware with limited VRAM, such as an 8GB RTX 4060 Laptop.
Use Cases
This model is ideal for applications requiring creative Chinese poetry generation, educational tools for classical Chinese literature, or projects exploring AI-driven stylistic imitation in poetry.