qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % python3 -m venv mlx_env
qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % source mlx_env/bin/activate
(mlx_env) qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % git clone https://github.com/ggerganov/llama.cpp.git
(mlx_env) qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % pip install -r llama.cpp/requirements.txt
(mlx_env) qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % pip install torch sentencepiece
(mlx_env) qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % pip install -U mlx-lm

(mlx_env) qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % python -m mlx_lm lora \
    --model google/gemma-3-4b-it \
    --data ./data \
    --train \
    --iters 200 \
    --batch-size 1 \
    --learning-rate 1e-5 \
    --mask-prompt \
    --num-layers 8 \
    --adapter-path ./adapters

(mlx_env) qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % mlx_lm.fuse \
  --model google/gemma-3-4b-it \
  --adapter-path adapters \
  --save-path ./gemma3-fused-f16

(mlx_env) qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % python llama.cpp/convert_hf_to_gguf.py ./gemma3-fused-f16 \
  --outfile gemma3-custom.gguf \
  --outtype f16

(mlx_env) qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % nano Modelfile2
(mlx_env) qsnich-it@MacBook-Air-khxng-Qsnich-IT Finetune % ollama create gemma3-jarvic -f Modelfile2