Qwen3.5-0.8B-mlx-LoRA

Qwen3.5-0.8B-mlx-LoRA (Vietnamese)

This model is a LoRA adapter for Qwen3.5-0.8B-4bit, fine-tuned on the Vietnamese Alpaca dataset using the MLX framework on Apple Silicon.

Model Details

Model Description

• Developed by: baeGil
• Model type: LoRA Adapter
• Language(s) (NLP): Vietnamese (vi)
• License: Apache 2.0
• Finetuned from model: mlx-community/Qwen3.5-0.8B-4bit

Model Sources

• Repository: https://huggingface.co/baeGil/Qwen3.5-0.8B-mlx-LoRA

Uses

Direct Use

This model is intended for chat and assistant tasks in Vietnamese. It can be used to answer general questions, assist in writing, and provide information in Vietnamese.

Out-of-Scope Use

The model should not be used for generating harmful, illegal, or unethical content. It is not suitable for high-stakes medical or legal advice.

Training Details

Training Data

The model was fine-tuned on the 5CD-AI/Vietnamese-alpaca-gpt4-gg-translated dataset, which contains 500 samples of instruction-following data in Vietnamese.

Training Procedure

• Framework: MLX
• Iterations: 50 (experimented with multiple ranks)
• Batch Size: 1
• Learning Rate: 2e-5
• Hardware: Apple M2 Pro (16GB RAM)

Training Hyperparameters

• LoRA Rank: 16 (Optimal)
• LoRA Alpha: 32
• Max Sequence Length: 256

Evaluation

Results

Based on local evaluation on a test set:

• Test Loss: 1.7527
• Perplexity: 5.7702

How to Get Started with the Model

To generate text using this adapter with mlx-vlm:

python -m mlx_vlm.generate \
  --model-path mlx-community/Qwen3.5-0.8B-4bit \
  --adapter-path baeGil/Qwen3.5-0.8B-mlx-LoRA \
  --prompt "### Instruction:\nCách nấu phở bò truyền thống?\n\n### Response:\n" \
  --max-tokens 256

Environmental Impact

• Hardware Type: Apple M2 Pro
• Time used: ~12 minutes
• Carbon Emitted: Negligible due to high efficiency of Apple Silicon.

Model Card Contact

[baeGil]

Github Repository

2A202600038-DoHaiNam-Day21