shikunpunk/Qwen2.5-3B-Haizi-CN
shikunpunk/Qwen2.5-3B-Haizi-CN is a 3.1 billion parameter causal language model, fine-tuned from shikunpunk/Qwen2.5-3B-Haizi, which is based on Qwen2.5-3B-Instruct. This model is specifically optimized for generating Chinese poetry in the style of Haizi, with a unique constraint to prevent the inclusion of any English words or phrases. It excels at producing pure Chinese poetic output, making it suitable for creative writing applications requiring strict linguistic purity.
Loading preview...
Overview
shikunpunk/Qwen2.5-3B-Haizi-CN is a specialized 3.1 billion parameter language model, derived from the Qwen2.5-3B-Instruct architecture and further fine-tuned from shikunpunk/Qwen2.5-3B-Haizi. Its primary distinction lies in its rigorous enforcement of pure Chinese output, specifically for poetry generation. This model addresses an issue present in its predecessor, where English words or phrases could occasionally appear in generated text, disrupting the intended Chinese poetic experience.
Key Differentiators
- Pure Chinese Output: Through targeted SFT fine-tuning with 21 pure Chinese constraint samples, the model achieves a 0.0% English混入率 (English mixing rate) in its poetic generations, compared to 58.0% in the original Haizi model.
- Haizi Poetic Style: Retains the distinctive poetic style of Haizi, focusing on modern Chinese poetry.
- Efficient Fine-tuning: Achieved its specialized capability with only 3 rounds of SFT using QLoRA 4-bit NF4, demonstrating effective adaptation with minimal data (155 samples).
Use Cases
- Chinese Poetry Generation: Ideal for creating modern Chinese poetry, particularly in the style of Haizi.
- Linguistic Purity: Suitable for applications where strict adherence to a single language (Chinese) is critical, without any foreign language interference.
- Creative Writing Tools: Can be integrated into tools for writers and artists focusing on Chinese literary creation.