메뉴
Guides agents on preparing datasets, running parameter-efficient LoRA fine-tuning, and tracking telemetry metrics using the mlxtune CLI on Apple Silicon. Use when an agent wants to fine-tune standard Gemma 4 base models on text instructions.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Guides agents on exporting fine-tuned Gemma models to multi-format mobile runtimes, comparing memory and hardware target characteristics, and implementing dynamic sandbox download workflows.
Guides agents on preparing multimodal datasets in interleaved ChatML format, executing audio/vision fine-tuning with mlx-vlm, targeting specific layers, and performing key-sanitized weights fusion to avoid silent blank weights loading.
Guides agents on fine-tuning Gemma 4 Quantization-Aware Training (QAT) unquantized checkpoints, maintaining mathematical alignment during LoRA, and exporting to GGUF using strict q4_0 parameters to avoid post-training quantization drift on edge devices.
| name | gemma-fine-tuning |
| description | Guides agents on preparing datasets, running parameter-efficient LoRA fine-tuning, and tracking telemetry metrics using the mlxtune CLI on Apple Silicon. Use when an agent wants to fine-tune standard Gemma 4 base models on text instructions. |
| compatibility | Requires Python 3.14+, uv, and Apple Silicon Mac |
This skill provides step-by-step instructions for AI coding assistants and agents to fine-tune standard Gemma 4 models on text datasets using the local mlxtune tool.
Download a high-precision or pre-quantized base model from Hugging Face:
uv run mlxtune download --model mlx-community/gemma-2-2b-it-4bit --dest ./model
Convert custom instruction-response pairs from Hugging Face or local data into the standard ChatML JSONL format:
uv run mlxtune prep --dataset yahma/alpaca-cleaned --dest ./data --samples 1000
This splits data into data/train.jsonl and data/valid.jsonl.
Run the parameter-efficient LoRA fine-tuning loop on Apple Metal GPU:
uv run mlxtune train --model ./model --data ./data --iters 200 --batch-size 2 --rank 8 --lr 1e-5
--iters: Total training steps.--batch-size: Batch size per step.--rank: LoRA adapter dimension (default 8).--lr: Learning rate (default 1e-5).Inspect the real-time logging output written to training_log.jsonl. The log format emits JSON objects representing training steps:
{"type": "train", "iter": 10, "loss": 2.214, "learning_rate": 1e-5, "peak_mem_gb": 10.4}
This log is read dynamically by the native macOS companion app MLXMonitor.
--batch-size to 1 or decrease --lora-layers from 16 to 8.uv run mlxtune clean before starting a new experiment to ensure older adapter weights or logs do not pollute the workspace.