peterbuitho/VietPoet-Qwen3.5-4B
peterbuitho/VietPoet-Qwen3.5-4B is a 4.5 billion parameter Qwen3.5-4B model fine-tuned by peterbuitho specifically for generating Vietnamese lục bát poems. This model excels at producing poems that adhere to the complex structural and tonal rules of lục bát, achieving high rule scores when used with a line-by-line sampler. It is primarily designed for creative text generation in Vietnamese, focusing on a specific poetic form.
Loading preview...
Overview
peterbuitho/VietPoet-Qwen3.5-4B is a specialized 4.5 billion parameter model, fine-tuned from Qwen3.5-4B, dedicated to generating Vietnamese lục bát poems. It leverages QLoRA training on a corpus of 8,000 filtered Vietnamese poems to master the intricate rules of this poetic form. The model is provided in a 16-bit (bf16) safetensors format, suitable for vLLM or transformers, with a GGUF version also available for LM Studio/llama.cpp.
Key Capabilities
- Lục Bát Poem Generation: Specifically trained to compose Vietnamese lục bát poems.
- Rule Adherence: When combined with a line-by-line sampler and rule checker (from
github.com/peterbuitho/ThoLucBat), the model achieves a 98% rate of fully valid poems, significantly improving rule adherence compared to the raw model. - Form over Meaning: While highly accurate in poetic form (rhyme, length, tone), the generated poems' meaning can sometimes be loose or off-topic, as training focused on form rather than deep semantic coherence with diverse topics.
Training Details
The model was fine-tuned using QLoRA (4-bit base) with specific parameters including r=16, alpha=16, lr 2e-4, and 2 epochs, utilizing Unsloth. The training data consisted of a filtered subset of the phamson02/vietnamese-poetry-corpus.
Prompt Format
It uses the Qwen chat format, with the sampler appending poem lines to the assistant's turn. Users can provide prompts requesting lục bát poems on various topics.