文章用LoRAを最短で作る方法｜Windowsでの環境構築＆学習手順まとめ

python -m pip show transformers

python

import transformers
print(transformers.__version__)

import torch
print(torch.cuda.is_available())

exit()

C:\AI>

C:\Users\Owner>

cd C:\AI
.\venv_lora\Scripts\activate

cd C:\AI
python -m venv venv_lora
.\venv_lora\Scripts\activate

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

pip install peft accelerate datasets safetensors transformers

from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments
from datasets import load_dataset
from peft import LoraConfig, get_peft_model
import torch

# ベースモデルとトークナイザー
base_model_name = "EleutherAI/gpt-neo-1.3B"
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
model = AutoModelForCausalLM.from_pretrained(base_model_name)

# LoRA設定
lora_config = LoraConfig(
    r=8,
    lora_alpha=16,
    target_modules=["q_proj","v_proj"],
    lora_dropout=0.1,
    bias="none",
    task_type="CAUSAL_LM"
)

model = get_peft_model(model, lora_config)

# データセット
dataset = load_dataset("text", data_files={"train":"sakasa.txt"})

# トークナイズ関数
def tokenize(batch):
    return tokenizer(batch["text"], padding="max_length", truncation=True, max_length=128)

tokenized_dataset = dataset.map(tokenize, batched=True)

# 学習設定
training_args = TrainingArguments(
    output_dir="./aisister_lora_file1",
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,
    num_train_epochs=3,
    logging_steps=10,
    save_steps=50,
    save_total_limit=2,
    fp16=True,
    push_to_hub=False
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset["train"]
)

# 学習開始
trainer.train()
trainer.save_model("./aisister_lora_file1")

cd C:\AI
.\venv_lora\Scripts\activate

python train_aisister_lora.py

 [00:00<00:00, 1.63it/s] LoRA 中間保存: ./sakasa_lora_file10 LoRA 最終保存: ./sakasa_lora

# 例: peftのmerge_lora.py で統合
python merge_lora.py \
  --lora_dirs "aisister_lora_file1" "aisister_lora_file2" "aisister_lora_file3" \
  --output_dir "aisister_lora_merged"

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_name = "EleutherAI/gpt-neo-1.3B"
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
model = AutoModelForCausalLM.from_pretrained(base_model_name)

lora_path = r"C:\AI\aisister_lora\aisister_lora_merged\adapter_model.safetensors"
model = PeftModel.from_pretrained(model, lora_path)
model.eval()

prompt = "こんにちは、私は AI Sister です。今日の天気は"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

# ベースモデル
base_model_name = "EleutherAI/gpt-neo-1.3B"

# トークナイザーとモデルの読み込み
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
model = AutoModelForCausalLM.from_pretrained(base_model_name)

# LoRAの適用（フォルダ名だけでOK）
# LoRAファイルが同じSpace内にある場合、相対パスだけでOK
model = PeftModel.from_pretrained(model, "./")  # app.py と同じディレクトリにある場合
model.eval()

# 推論関数
def generate_text(prompt):
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=50)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Gradio UI
iface = gr.Interface(
    fn=generate_text,
    inputs=gr.Textbox(label="プロンプトを入力"),
    outputs=gr.Textbox(label="生成結果"),
    title="AI Sister LoRA テスト"
)

iface.launch()

# LoRA テスト用最小限 requirements.txt

# Transformers ライブラリ（Hugging Face モデル用）
transformers>=4.35

# PEFT（LoRA 適用用）
peft>=0.5.0

# TRL（必要に応じて）
trl>=0.8.0

# PyTorch（GPU対応）
torch>=2.1.0

# Safetensors（高速モデル読み込み用）
safetensors

# Gradio（WebUI用）
gradio>=3.53