Qwen2.5-0.5B-Vafin-Style-GGUF

Qwen2.5-0.5B-Yury-Vafin-Style

This is a fine-tuned version of Qwen/Qwen2.5-0.5B-Instruct trained on the text corpus of the famous Russian writer and blogger Yury Vafin. The model captures his unique absurdist, melancholic, and tragicomic prose style.

🚀 Project Purpose

This is a pet project created for fine-tuning practice, exploring QLoRA, and deploying small language models (SLMs).

📊 Dataset & Training Details

• Base Model: Qwen/Qwen2.5-0.5B-Instruct
• Method: QLoRA (4-bit quantization via bitsandbytes)
• Dataset: ~286 structured chunks (extracted from FB2 using a custom BeautifulSoup parser)
• Hardware: Trained on a single NVIDIA T4 GPU (Google Colab)
• Training Epochs: 3
• Final Train Loss: 2.74

🔧 Recommended Inference Settings

To avoid repetitive loops (common for 0.5B models), use the following parameters in your API request (e.g., via llama-cpp-python / Postman):

• Temperature: 0.85
• Frequency Penalty: 1.15
• Presence Penalty: 1.0
• Top-P: 0.9