Google Colab で Rinna-3.6B のLoRAファインチューニングを試す

# Googleドライブのマウント
from google.colab import drive
drive.mount("/content/drive")

# 作業フォルダへの移動
import os
os.makedirs("/content/drive/My Drive/work", exist_ok=True)
%cd '/content/drive/My Drive/work'

# 基本パラメータ
model_name = "rinna/japanese-gpt-neox-3.6b"
dataset = "kunishou/databricks-dolly-15k-ja"
peft_name = "lora-rinna-3.6b"
output_dir = "lora-rinna-3.6b-results"

# PEFTのインストール
!pip install -Uqq  git+https://github.com/huggingface/peft.git
!pip install -Uqq transformers datasets accelerate bitsandbytes
!pip install sentencepiece

from transformers import AutoTokenizer

# トークナイザーの準備
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)

# スペシャルトークンの確認
print(tokenizer.special_tokens_map)
print("bos_token :", tokenizer.bos_token, ",", tokenizer.bos_token_id)
print("eos_token :", tokenizer.eos_token, ",", tokenizer.eos_token_id)
print("unk_token :", tokenizer.unk_token, ",", tokenizer.unk_token_id)
print("pad_token :", tokenizer.pad_token, ",", tokenizer.pad_token_id)

{'bos_token': '<s>', 'eos_token': '</s>', 'unk_token': '[UNK]', 'sep_token': '[SEP]', 'pad_token': '[PAD]', 'cls_token': '[CLS]', 'mask_token': '[MASK]'}
bos_token : <s> , 2
eos_token : </s> , 3
unk_token : [UNK] , 1
pad_token : [PAD] , 0

CUTOFF_LEN = 256  # コンテキスト長

# トークナイズ
def tokenize(prompt, tokenizer):
    result = tokenizer(
        prompt,
        truncation=True,
        max_length=CUTOFF_LEN,
        padding=False,
    )
    return {
        "input_ids": result["input_ids"],
        "attention_mask": result["attention_mask"],
    }

# トークナイズの動作確認
tokenize("hi there", tokenizer)

{'input_ids': [3201, 634, 1304, 3], 'attention_mask': [1, 1, 1, 1]}

from datasets import load_dataset

# データセットの準備
data = load_dataset(dataset)

# データセットの確認
data["train"][5]

{'category': 'information_extraction',
 'instruction': 'ステイルメイトの時に、私の方が多くの駒を持っていたら、私の勝ちですか？',
 'output': 'いいえ。\nステイルメイトとは、引き分けた状態のことです。どちらがより多くの駒を捕獲したか、または優勢であるかは関係ない',
 'input': 'ステイルメイトとは、チェスにおいて、手番が回ってきたプレーヤーがチェックされておらず、合法的な手がない状態のことである。ステイルメイトの結果、引き分けとなる。終盤では、ステイルメイトは劣勢にあるプレイヤーが負けるのではなく、ゲームを引き分けることを可能にする戦術である[2]。より複雑なポジションでは、ステイルメイトはより稀で、通常は優勢側が不注意な場合にのみ成功する詐欺の形をとる[引用] ステイルメイトは終盤研究や他のチェスの問題においても共通のテーマである。\n\nステイルメイトが引き分けに統一されたのは19世紀である。それ以前は、ステイルメイトしているプレイヤーの勝利、引き分け、負けとみなされたり、反則となったり、ステイルメイトしているプレイヤーはターンを失うことになったりと、その扱いは様々であった。ステイルメイトのルールは、チェス以外のチャトランガ系ゲームごとに異なる。',
 'index': '5'}

# プロンプトテンプレートの準備
def generate_prompt(data_point):
    if data_point["input"]:
        result = f"""### 指示:
{data_point["instruction"]}

### 入力:
{data_point["input"]}

### 回答:
{data_point["output"]}"""
    else:
        result = f"""### 指示:
{data_point["instruction"]}

### 回答:
{data_point["output"]}"""

    # 改行→<NL>
    result = result.replace('\n', '<NL>')
    return result

# プロンプトテンプレートの確認
print(generate_prompt(data["train"][5]))

### 指示:<NL>ステイルメイトの時に、私の方が多くの駒を持っていたら、私の勝ちですか？<NL><NL>### 入力:<NL>ステイルメイトとは、チェスにおいて、手番が回ってきたプレーヤーがチェックされておらず、合法的な手がない状態のことである。ステイルメイトの結果、引き分けとなる。終盤では、ステイルメイトは劣勢にあるプレイヤーが負けるのではなく、ゲームを引き分けることを可能にする戦術である[2]。より複雑なポジションでは、ステイルメイトはより稀で、通常は優勢側が不注意な場合にのみ成功する詐欺の形をとる[引用] ステイルメイトは終盤研究や他のチェスの問題においても共通のテーマである。<NL><NL>ステイルメイトが引き分けに統一されたのは19世紀である。それ以前は、ステイルメイトしているプレイヤーの勝利、引き分け、負けとみなされたり、反則となったり、ステイルメイトしているプレイヤーはターンを失うことになったりと、その扱いは様々であった。ステイルメイトのルールは、チェス以外のチャトランガ系ゲームごとに異なる。<NL><NL>### 回答:<NL>いいえ。<NL>ステイルメイトとは、引き分けた状態のことです。どちらがより多くの駒を捕獲したか、または優勢であるかは関係ない

VAL_SET_SIZE = 2000

# 学習データと検証データの準備
train_val = data["train"].train_test_split(
    test_size=VAL_SET_SIZE, shuffle=True, seed=42
)
train_data = train_val["train"]
val_data = train_val["test"]
train_data = train_data.shuffle().map(lambda x: tokenize(generate_prompt(x), tokenizer))
val_data = val_data.shuffle().map(lambda x: tokenize(generate_prompt(x), tokenizer))

from transformers import AutoModelForCausalLM

# モデルの準備
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    load_in_8bit=True,
    device_map="auto",
)

from peft import LoraConfig, get_peft_model, prepare_model_for_int8_training, TaskType

# LoRAのパラメータ
lora_config = LoraConfig(
    r= 8, 
    lora_alpha=16,
    target_modules=["query_key_value"],
    lora_dropout=0.05,
    bias="none",
    task_type=TaskType.CAUSAL_LM
)

# モデルの前処理
model = prepare_model_for_int8_training(model)

# LoRAモデルの準備
model = get_peft_model(model, lora_config)

# 学習可能パラメータの確認
model.print_trainable_parameters()

trainable params: 3244032 || all params: 3610489344 || trainable%: 0.08985020286491116

import transformers
eval_steps = 200
save_steps = 200
logging_steps = 20

# トレーナーの準備
trainer = transformers.Trainer(
    model=model,
    train_dataset=train_data,
    eval_dataset=val_data,
    args=transformers.TrainingArguments(
        num_train_epochs=3,
        learning_rate=3e-4,
        logging_steps=logging_steps,
        evaluation_strategy="steps",
        save_strategy="steps",
        eval_steps=eval_steps,
        save_steps=save_steps,
        output_dir=output_dir,
        report_to="none",
        save_total_limit=3,
        push_to_hub=False,
        auto_find_batch_size=True
    ),
    data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False),
)

# 学習の実行
model.config.use_cache = False
trainer.train() 
model.config.use_cache = True

# LoRAモデルの保存
trainer.model.save_pretrained(peft_name)

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

# モデルの準備
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    load_in_8bit=True,
    device_map="auto",
)

# トークナイザーの準備
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)

# LoRAモデルの準備
model = PeftModel.from_pretrained(
    model, 
    peft_name, 
    device_map="auto"
)

# 評価モード
model.eval()

# プロンプトテンプレートの準備
def generate_prompt(data_point):
    if data_point["input"]:
        result = f"""### 指示:
{data_point["instruction"]}

### 入力:
{data_point["input"]}

### 回答:
"""
    else:
        result = f"""### 指示:
{data_point["instruction"]}

### 回答:
"""

    # 改行→<NL>
    result = result.replace('\n', '<NL>')
    return result

# テキスト生成関数の定義
def generate(instruction,input=None,maxTokens=256):
    # 推論
    prompt = generate_prompt({'instruction':instruction,'input':input})
    input_ids = tokenizer(prompt, 
        return_tensors="pt", 
        truncation=True, 
        add_special_tokens=False).input_ids.cuda()
    outputs = model.generate(
        input_ids=input_ids, 
        max_new_tokens=maxTokens, 
        do_sample=True,
        temperature=0.7, 
        top_p=0.75, 
        top_k=40,         
        no_repeat_ngram_size=2,
    )
    outputs = outputs[0].tolist()
    print(tokenizer.decode(outputs))

    # EOSトークンにヒットしたらデコード完了
    if tokenizer.eos_token_id in outputs:
        eos_index = outputs.index(tokenizer.eos_token_id)
        decoded = tokenizer.decode(outputs[:eos_index])

        # レスポンス内容のみ抽出
        sentinel = "### 回答:"
        sentinelLoc = decoded.find(sentinel)
        if sentinelLoc >= 0:
            result = decoded[sentinelLoc+len(sentinel):]
            print(result.replace("<NL>", "\n"))  # <NL>→改行
        else:
            print('Warning: Expected prompt template to be emitted.  Ignoring output.')
    else:
        print('Warning: no <eos> detected ignoring output')

generate("自然言語処理とは？")

人間が日常的に使っている自然な言語を理解し、自然に文章を生成する技術。

generate("日本の首都は？")

東京は、日本の首都です。

generate("まどか☆マギカで一番かわいいのは？")

暁美ほむらは、まどかマミの親友であり、魔女の使い魔であるキュウベエの力で、時間を遡り、自分の命を救ってくれた少女の命を救うために、魔法少女になる。  ほむほーむは「魔法少女まどか★マゥト」のアニメで人気となり、多くのファンに愛されました。