【初学者向け】無料版 Google Colab でローカルLLM をファインチューニングする

Python

Copy

import os
!pip install --no-deps bitsandbytes accelerate xformers==0.0.29.post3 peft trl==0.15.2 triton cut_cross_entropy unsloth_zoo torchao==0.17.0
!pip install sentencepiece protobuf "datasets>=3.4.1" huggingface_hub hf_transfer
!pip install --no-deps unsloth

Python

Copy

from unsloth import FastLanguageModel
import torch
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Meta-Llama-3.1-8B",
    max_seq_length = 2048, # モデルが一度に処理できる最大トークン数
    dtype = None, # モデルの計算に使うデータの型(None: GPUに最適な型を自動判別)
    load_in_4bit = True, # 4bit量子化
)

Python

Copy

model = FastLanguageModel.get_peft_model(
    model,
    r = 16, # LoRA のランク。8~64が推奨値。
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",], # ファインチューニングの対象とする層
    lora_alpha = 16, # スケーリング係数。学習内容を元のモデルにどれくらい強く反映させるかを決める
    lora_dropout = 0, # 過学習を防ぐためのドロップアウト率
    bias = "none",
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
)

Python

Copy

from datasets import load_dataset
ds = load_dataset("amaydle/npc-dialogue", split="train")

Python

Copy

prompt_format = """Below is a conversation with a fictional character. Assume the persona described in the biography and reply appropriately.

### Biography/Context:
{}

### Instruction/Query:
{}

### Response:
{}"""

EOS_TOKEN = tokenizer.eos_token # 文の終わりを示す特殊トークン

def formatting_prompts_func(examples):
        biographies = examples["Biography"]
        queries     = examples["Query"]
        responses   = examples["Response"]
        texts = []
        for biography, query, response in zip(biographies, queries, responses):
            text = prompt_format.format(biography, query, response) + EOS_TOKEN
            texts.append(text)
        return { "text" : texts, }

# データセット全体に一括適用（map関数）
ds = ds.map(formatting_prompts_func, batched = True)

Python

Copy

from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = ds,
    dataset_text_field = "text",
    max_seq_length = 2048,
    dataset_num_proc = 2,
    packing = False, # Can make training 5x faster for short sequences.
    args = TrainingArguments(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 4,
        warmup_steps = 5,
        # num_train_epochs = 1, # Set this for 1 full training run.
        max_steps = 500,
        learning_rate = 2e-4,
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
        report_to = "none", # Use this for WandB etc
    ),
)
trainer_stats = trainer.train()

Python

Copy

FastLanguageModel.for_inference(model)
inputs = tokenizer([prompt_text], return_tensors="pt").to("cuda")
text_streamer = TextStreamer(tokenizer, skip_prompt=True)
_ = model.generate(
    **inputs,
    streamer=text_streamer,
    max_new_tokens=2048,
    use_cache=True,
    temperature=0.7,
    do_sample=True
)

Python

Copy

# ローカル（Colabのストレージ内）にLoRAのアダプターを保存
model.save_pretrained("anime_lora_model")
tokenizer.save_pretrained("anime_lora_model")

# もしくは、Hugging Face Hubに直接アップロード（Write権限のトークンが必要）
# model.push_to_hub("your_username/anime_lora_model", token="YOUR_HF_TOKEN")
# tokenizer.push_to_hub("your_username/anime_lora_model", token="YOUR_HF_TOKEN")