Gemma 4をLoRAでファインチューニングする方法（ステップバイステップ）

# 仮想環境を作成
python -m venv gemma4-finetune
source gemma4-finetune/bin/activate

# Unslothをインストール（すべての依存関係を含む）
pip install unsloth

# 古いGPUでQLoRAを使う場合、追加で：
pip install bitsandbytes

import torch
print(f"CUDA available: {torch.cuda.is_available()}")
print(f"GPU: {torch.cuda.get_device_name(0)}")
print(f"VRAM: {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} GB")

{"messages": [{"role": "user", "content": "What's the return policy?"}, {"role": "assistant", "content": "Our return policy allows returns within 30 days of purchase with original receipt."}]}
{"messages": [{"role": "user", "content": "Do you ship internationally?"}, {"role": "assistant", "content": "Yes, we ship to 45 countries. Shipping takes 7-14 business days."}]}
{"messages": [{"role": "system", "content": "You are a medical coding assistant."}, {"role": "user", "content": "Code: Patient presents with acute bronchitis"}, {"role": "assistant", "content": "ICD-10: J20.9 - Acute bronchitis, unspecified"}]}

import json

def validate_jsonl(filepath):
    valid = 0
    errors = 0
    with open(filepath, 'r') as f:
        for i, line in enumerate(f, 1):
            try:
                data = json.loads(line)
                assert "messages" in data, "Missing 'messages' key"
                assert len(data["messages"]) >= 2, "Need at least 2 messages"
                valid += 1
            except (json.JSONDecodeError, AssertionError) as e:
                print(f"Line {i}: {e}")
                errors += 1
    print(f"\nValid: {valid}, Errors: {errors}")

validate_jsonl("training_data.jsonl")

from unsloth import FastLanguageModel
import torch

# QLoRA（4bit）でモデルをロード
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="google/gemma-4-e4b",
    max_seq_length=4096,
    dtype=None,           # 自動検出
    load_in_4bit=True,    # QLoRA
)

# LoRAアダプターを追加
model = FastLanguageModel.get_peft_model(
    model,
    r=16,                  # ランク — 高いほど容量増、VRAM増
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    lora_alpha=16,
    lora_dropout=0,        # Unsloth最適化 — 0を維持
    bias="none",
    use_gradient_checkpointing="unsloth",
)

# 訓練データをロード
from datasets import load_dataset
dataset = load_dataset("json", data_files="training_data.jsonl", split="train")

# 訓練引数
from trl import SFTTrainer
from transformers import TrainingArguments

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    max_seq_length=4096,
    args=TrainingArguments(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        warmup_steps=10,
        num_train_epochs=3,
        learning_rate=2e-4,
        fp16=not torch.cuda.is_bf16_supported(),
        bf16=torch.cuda.is_bf16_supported(),
        logging_steps=10,
        output_dir="outputs",
        optim="adamw_8bit",
        seed=42,
    ),
)

# 訓練
trainer.train()

# LoRAアダプターを保存
model.save_pretrained("gemma4-finetuned-lora")
tokenizer.save_pretrained("gemma4-finetuned-lora")

# アダプターをマージしてGGUFにエクスポート（Q4_K_M量子化）
model.save_pretrained_gguf(
    "gemma4-finetuned-gguf",
    tokenizer,
    quantization_method="q4_k_m",
)

# または複数の量子化レベルでエクスポート
for method in ["q4_k_m", "q5_k_m", "q8_0"]:
    model.save_pretrained_gguf(
        f"gemma4-finetuned-{method}",
        tokenizer,
        quantization_method=method,
    )

# Modelfile
FROM ./gemma4-finetuned-gguf/gemma4-finetuned-q4_k_m.gguf

PARAMETER temperature 0.7
PARAMETER top_p 0.9

SYSTEM "You are a helpful assistant fine-tuned for customer support."

# カスタムモデルを作成
ollama create my-gemma4 -f Modelfile

# テスト
ollama run my-gemma4 "What's your return policy?"

# リストに表示されることを確認
ollama list