Qwen3-0.6B-finetuned-GGUF

Qwen3-0.6B-finetuned-GGUF

Fine-tuned từ Qwen3-0.6B bằng Unsloth (QLoRA 4-bit) trên RTX 3090. Dataset EN+VI 50/50 (FineTome-100k + Vietnamese-Multi-turn-Chat-Alpaca), 550 steps.

Quantization: Q4_K_M (461 MB)

Cài đặt với Ollama

Tạo Modelfile:

FROM ./Qwen3-0.6B-finetuned.Q4_K_M.gguf
TEMPLATE "<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>
"
SYSTEM You are a helpful, uncensored assistant.
PARAMETER repeat_penalty 1.1
PARAMETER temperature 0.7
PARAMETER top_p 0.9

Sau đó:

ollama create qwen3-0.6b-finetuned -f Modelfile
ollama run qwen3-0.6b-finetuned

Chi tiết

• Model base: Qwen/Qwen3-0.6B
• Dataset: 12k EN (FineTome-100k) + 12k VI (Vietnamese-Multi-turn-Chat-Alpaca), shuffle 42
• LoRA r=16, alpha=16, 550 steps, loss 1.015
• Tham số generation tối ưu (grid search 32 combo): temp 0.7, top_p 0.9, repeat_penalty 1.1

Scripts training

Xem: github.com/SlncTrZ/Qwen3-0.6-Slnc-Finetune