This guide will help you fine-tune the Qwen3-Coder model (a Qwen-based language model optimized for code generation) to support the Amber programming language.

The goal is to train the model with hundreds of Amber scripting examples, enabling it to:

• Generate Amber code
• Convert Bash or Python scripts to Amber

This guide assumes basic knowledge of Linux, Python, and machine learning.

Fine-tuning a large model requires significant resources:

• CPU / RAM
• NVIDIA GPU (minimum 8GB VRAM, recommended 24GB+)

To reduce memory requirements, this guide uses PEFT (Parameter-Efficient Fine-Tuning) with LoRA.

No explanation needed

Run install.sh.

The previous script does it!

If authentication is required:

huggingface-cli login

Run:

source ./amber-finetune-env/bin/activate
./prepare_dataset.py /Amber/src/tests/stdlib/ /Amber/src/tests/translating/

Run training:

python train.py

Note: Training may take hours or days depending on dataset and GPU. Monitor with nvidia-smi or htop.

Run:

python test.py