Gemma を使用して chatbot を構築する

ai.google.dev で表示 Google Colab で実行 Vertex AI で開く GitHub のソースを表示

Gemma などの大規模言語モデル(LLM)は有益な回答の生成に優れており、仮想アシスタントや chatbot の構築に最適です。

従来、LLM はステートレスな方法で動作します。つまり、過去の会話を保存するための固有のメモリがありません。各プロンプトや質問は、以前のやり取りを無視して個別に処理されます。しかし、自然な会話の重要な側面は、以前のやり取りからコンテキストを保持する能力です。この制限を克服し、LLM が会話のコンテキストを維持できるようにするには、LLM に提示される新しいプロンプトごとに、会話履歴(または関連部分)などの関連情報を明示的に提供する必要があります。

このチュートリアルでは、指示用にチューニングされた Gemma のモデル バリアントを使用して chatbot を開発する方法について説明します。

セットアップ

Gemma の設定

このチュートリアルを完了するには、まず Gemma の設定にある設定手順を完了する必要があります。Gemma の設定手順では、次の方法について説明します。

  • kaggle.com で Gemma にアクセスしてください。
  • 実行するのに十分なリソースがある Colab ランタイムを選択します Gemma 2B モデルを使用します。
  • Kaggle のユーザー名と API キーを生成して構成します。

Gemma の設定が完了したら、次のセクションに進み、Colab 環境の環境変数を設定します。

環境変数を設定する

KAGGLE_USERNAMEKAGGLE_KEY の環境変数を設定します。

import os
from google.colab import userdata

# Note: `userdata.get` is a Colab API. If you're not using Colab, set the env
# vars as appropriate for your system.
os.environ["KAGGLE_USERNAME"] = userdata.get('KAGGLE_USERNAME')
os.environ["KAGGLE_KEY"] = userdata.get('KAGGLE_KEY')

依存関係のインストール

Keras と KerasNLP をインストールする。

# Install Keras 3 last. See https://keras.io/getting_started/ for more details.
pip install -q tensorflow-cpu
pip install -q -U keras-nlp tensorflow-hub
pip install -q -U "keras>=3"
pip install -q -U tensorflow-text

バックエンドを選択

Keras は、シンプルさと使いやすさを考慮して設計された高レベルのマルチフレームワーク ディープ ラーニング API です。Keras 3 では、バックエンド(TensorFlow、JAX、PyTorch)を選択できます。このチュートリアルでは、3 つすべてを使用できます。

import os

# Select JAX as the backend
os.environ["KERAS_BACKEND"] = "jax"

# Pre-allocate 100% of TPU memory to minimize memory fragmentation
os.environ["XLA_PYTHON_CLIENT_MEM_FRACTION"] = "1.0"

パッケージをインポートする

Keras と KerasNLP をインポートする。

import keras
import keras_nlp

# for reproducibility
keras.utils.set_random_seed(42)

モデルをインスタンス化する

KerasNLP には、多くの一般的なモデル アーキテクチャの実装が用意されています。このチュートリアルでは、因果言語モデリング用のエンドツーエンドの Gemma モデルである GemmaCausalLM を使用してモデルをインスタンス化します。因果言語モデルは、以前のトークンに基づいて次のトークンを予測します。

from_preset メソッドを使用してモデルをインスタンス化します。

gemma_lm = keras_nlp.models.GemmaCausalLM.from_preset("gemma2_instruct_2b_en")

GemmaCausalLM.from_preset() 関数は、プリセットのアーキテクチャと重みからモデルをインスタンス化します。上記のコードでは、文字列 "gemma2_instruct_2b_en" に、20 億のパラメータがある Gemma 2 2B モデルのプリセットを指定しています。7B、9B、27B パラメータを使用した Gemma モデルも使用できます。Gemma モデルのコード文字列は、Kaggle の [Model Variation] リストで確認できます。

summary メソッドを使用して、モデルに関する詳細情報を取得します。

gemma_lm.summary()

要約からわかるように、このモデルには 26 億個のトレーニング可能なパラメータがあります。

書式設定ヘルパー関数を定義する

from IPython.display import Markdown
import textwrap

def display_chat(prompt, text):
  formatted_prompt = "<font size='+1' color='brown'>🙋‍♂️<blockquote>" + prompt + "</blockquote></font>"
  text = text.replace('', '  *')
  text = textwrap.indent(text, '> ', predicate=lambda _: True)
  formatted_text = "<font size='+1' color='teal'>🤖\n\n" + text + "\n</font>"
  return Markdown(formatted_prompt+formatted_text)

def to_markdown(text):
  text = text.replace('', '  *')
  return Markdown(textwrap.indent(text, '> ', predicate=lambda _: True))

chatbot の構築

Gemma の指示用にチューニングされたモデル gemma2_instruct_2b_en は、次のターントークンを理解するようにファインチューニングされています。

<start_of_turn>user\n  ... <end_of_turn>\n
<start_of_turn>model\n ... <end_of_turn>\n

このチュートリアルでは、これらのトークンを使用して chatbot を構築します。Gemma コントロール トークンについて詳しくは、フォーマットとシステム手順をご覧ください。

会話の状態を管理するチャット ヘルパーを作成する

class ChatState():
  """
  Manages the conversation history for a turn-based chatbot
  Follows the turn-based conversation guidelines for the Gemma family of models
  documented at https://ai.google.dev/gemma/docs/formatting
  """

  __START_TURN_USER__ = "<start_of_turn>user\n"
  __START_TURN_MODEL__ = "<start_of_turn>model\n"
  __END_TURN__ = "<end_of_turn>\n"

  def __init__(self, model, system=""):
    """
    Initializes the chat state.

    Args:
        model: The language model to use for generating responses.
        system: (Optional) System instructions or bot description.
    """
    self.model = model
    self.system = system
    self.history = []

  def add_to_history_as_user(self, message):
      """
      Adds a user message to the history with start/end turn markers.
      """
      self.history.append(self.__START_TURN_USER__ + message + self.__END_TURN__)

  def add_to_history_as_model(self, message):
      """
      Adds a model response to the history with start/end turn markers.
      """
      self.history.append(self.__START_TURN_MODEL__ + message)

  def get_history(self):
      """
      Returns the entire chat history as a single string.
      """
      return "".join([*self.history])

  def get_full_prompt(self):
    """
    Builds the prompt for the language model, including history and system description.
    """
    prompt = self.get_history() + self.__START_TURN_MODEL__
    if len(self.system)>0:
      prompt = self.system + "\n" + prompt
    return prompt

  def send_message(self, message):
    """
    Handles sending a user message and getting a model response.

    Args:
        message: The user's message.

    Returns:
        The model's response.
    """
    self.add_to_history_as_user(message)
    prompt = self.get_full_prompt()
    response = self.model.generate(prompt, max_length=2048)
    result = response.replace(prompt, "")  # Extract only the new response
    self.add_to_history_as_model(result)
    return result

モデルとチャットする

モデルとのチャットを開始します。

chat = ChatState(gemma_lm)
message = "Tell me, in a few words,  how to compute all prime numbers up to 1000?"
display_chat(message, chat.send_message(message))

🙋‍♂️

1,000 までの素数をすべて計算する方法を簡潔に教えてください。
🤖?

エラトステネスのふるい。

会話を続けます。

message = "Now in Python! No numpy, please!"
display_chat(message, chat.send_message(message))

🙋‍♂️

Python を使用できるようになりました。
🤖?

def sieve_of_eratosthenes(n):
  """Returns a list of prime numbers up to n."""
  primes = [True] * (n + 1)
  primes[0] = primes[1] = False
  for i in range(2, int(n**0.5) + 1):
    if primes[i]:
      for j in range(i * i, n + 1, i):
        primes[j] = False
  return [i for i, is_prime in enumerate(primes) if is_prime]

primes = sieve_of_eratosthenes(1000)
print(primes)

説明:

  1. 初期化:

    • primes = [True] * (n + 1): ブール値のリスト primes を作成します。最初はすべての数値が素数であると想定します。
    • primes[0] = primes[1] = False: 0 と 1 を非素数に設定します。
  2. イテレーション:

    • for i in range(2, int(n**0.5) + 1):: 2 から n の平方根まで反復処理します。複合数は平方根以下の素因数を持つ必要があるため、チェックする必要があるのは平方根だけです。
    • if primes[i]:: i が素数としてマークされている場合: <ph type="x-smartling-placeholder">
        </ph>
      • for j in range(i * i, n + 1, i):: i のすべての倍数を非素数としてマークします。
  3. 結果:

    • return [i for i, is_prime in enumerate(primes) if is_prime]: primes[i] が素数を表すインデックスのリストを作成します。

部分的な説明をご希望の場合は、お気軽にお問い合わせください。

message = "Thank you, it works! Can you explain the code in French?"
display_chat(message, chat.send_message(message))

🙋‍♂️

ありがとうございました。フランス語でコードを説明してください
🤖?

Bien sûr ! Voici une explication du code en français :

def sieve_of_eratosthenes(n):
  """Retourne une liste de nombres premiers jusqu'à n."""
  primes = [True] * (n + 1)
  primes[0] = primes[1] = False
  for i in range(2, int(n**0.5) + 1):
    if primes[i]:
      for j in range(i * i, n + 1, i):
        primes[j] = False
  return [i for i, is_prime in enumerate(primes) if is_prime]

primes = sieve_of_eratosthenes(1000)
print(primes)

説明:

  1. 初期化:

    • primes = [True] * (n + 1): Crée une liste primes de Valeurs booléennes, initialement supposant que tous les nombres sont Premiers.
    • primes[0] = primes[1] = False: デフィニット 0 と 1 はプレミアレベルではない。
  2. 反復処理:

    • for i in range(2, int(n**0.5) + 1):: Itère de 2 jusqu'à la racine carrée de n.Nous ne devons vérifier que jusqu'à la racine carrée car tout nombre composite 方法は次のとおりです
    • if primes[i]:: Si i est considéré comme Premier: <ph type="x-smartling-placeholder">
        </ph>
      • for j in range(i * i, n + 1, i):: プレミアレベルにはi 注目してください。
  3. Resultat:

    • return [i for i, is_prime in enumerate(primes) if is_prime]: Crée une liste des indices où primes[i] est vrai, représentant les nombres premiers.

N'hésitez pas à medemander si vous avez besoin d'une explication plus détaillée de quelquechoose!

message = "Great! Now add those explanations as comments in the code."
display_chat(message, chat.send_message(message))

🙋‍♂️

ありがとうございます。次に、それらの説明をコードにコメントとして追加します。
🤖?

def sieve_of_eratosthenes(n):
  """Retourne une liste de nombres premiers jusqu'à n."""
  # Initialise une liste de boolean avec True pour tous les nombres de 0 à n
  primes = [True] * (n + 1)
  # Définit 0 et 1 comme non-premiers
  primes[0] = primes[1] = False
  # Itère de 2 à la racine carrée de n
  for i in range(2, int(n**0.5) + 1):
    # Si i est considéré comme premier
    if primes[i]:
      # Itère sur tous les multiples de i
      for j in range(i * i, n + 1, i):
        # Définit les multiples de i comme non-premiers
        primes[j] = False
  # Retourne la liste des indices des nombres premiers
  return [i for i, is_prime in enumerate(primes) if is_prime]

primes = sieve_of_eratosthenes(1000)
print(primes)

説明:

  • 初期化: <ph type="x-smartling-placeholder">
      </ph>
    • primes = [True] * (n + 1): Crée une liste primes de Valeurs booléennes, initialement supposant que tous les nombres sont Premiers.
    • primes[0] = primes[1] = False: デフィニット 0 と 1 はプレミアレベルではない。
  • 反復処理: <ph type="x-smartling-placeholder">
      </ph>
    • for i in range(2, int(n**0.5) + 1):: Itère de 2 jusqu'à la racine carrée de n.Nous ne devons vérifier que jusqu'à la racine carrée car tout nombre composite 方法は次のとおりです
    • if primes[i]:: Si i est considéré comme Premier: <ph type="x-smartling-placeholder">
        </ph>
      • for j in range(i * i, n + 1, i):: プレミアレベルにはi 注目してください。
  • Resultat: <ph type="x-smartling-placeholder">
      </ph>
    • return [i for i, is_prime in enumerate(primes) if is_prime]: Crée une liste des indices oú primes[i] est vrai, représentant les nombres premiers.

生成されたコードを実行して、生成されたレスポンスをテストします。

def sieve_of_eratosthenes(n):
  """Retourne une liste de nombres premiers jusqu'à n."""
  # Initialise une liste de boolean avec True pour tous les nombres de 0 à n
  primes = [True] * (n + 1)
  # Définit 0 et 1 comme non-premiers
  primes[0] = primes[1] = False
  # Itère de 2 à la racine carrée de n
  for i in range(2, int(n**0.5) + 1):
    # Si i est considéré comme premier
    if primes[i]:
      # Itère sur tous les multiples de i
      for j in range(i * i, n + 1, i):
        # Définit les multiples de i comme non-premiers
        primes[j] = False
  # Retourne la liste des indices des nombres premiers
  return [i for i, is_prime in enumerate(primes) if is_prime]

primes = sieve_of_eratosthenes(1000)
print(primes)
[2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71, 73, 79, 83, 89, 97, 101, 103, 107, 109, 113, 127, 131, 137, 139, 149, 151, 157, 163, 167, 173, 179, 181, 191, 193, 197, 199, 211, 223, 227, 229, 233, 239, 241, 251, 257, 263, 269, 271, 277, 281, 283, 293, 307, 311, 313, 317, 331, 337, 347, 349, 353, 359, 367, 373, 379, 383, 389, 397, 401, 409, 419, 421, 431, 433, 439, 443, 449, 457, 461, 463, 467, 479, 487, 491, 499, 503, 509, 521, 523, 541, 547, 557, 563, 569, 571, 577, 587, 593, 599, 601, 607, 613, 617, 619, 631, 641, 643, 647, 653, 659, 661, 673, 677, 683, 691, 701, 709, 719, 727, 733, 739, 743, 751, 757, 761, 769, 773, 787, 797, 809, 811, 821, 823, 827, 829, 839, 853, 857, 859, 863, 877, 881, 883, 887, 907, 911, 919, 929, 937, 941, 947, 953, 967, 971, 977, 983, 991, 997]

get_history メソッドを使用して、すべてのコンテキストが Chat クラスによってどのように保持されているかを確認します。

print(chat.get_history())
<start_of_turn>user
Tell me, in a few words,  how to compute all prime numbers up to 1000?<end_of_turn>
<start_of_turn>model
**Sieve of Eratosthenes.** 
<end_of_turn><start_of_turn>user
Now in Python! No numpy, please!<end_of_turn>
<start_of_turn>model

```python
def sieve_of_eratosthenes(n):
  """Returns a list of prime numbers up to n."""
  primes = [True] * (n + 1)
  primes[0] = primes[1] = False
  for i in range(2, int(n**0.5) + 1):
    if primes[i]:
      for j in range(i * i, n + 1, i):
        primes[j] = False
  return [i for i, is_prime in enumerate(primes) if is_prime]

primes = sieve_of_eratosthenes(1000)
print(primes)
```

**Explanation:**

1. **Initialization:**
   - `primes = [True] * (n + 1)`: Creates a list `primes` of boolean values, initially assuming all numbers are prime.
   - `primes[0] = primes[1] = False`: Sets 0 and 1 as non-prime.

2. **Iteration:**
   - `for i in range(2, int(n**0.5) + 1):`: Iterates from 2 to the square root of `n`. We only need to check up to the square root because any composite number must have a prime factor less than or equal to its square root.
   - `if primes[i]:`: If `i` is marked as prime:
     - `for j in range(i * i, n + 1, i):`: Marks all multiples of `i` as non-prime.

3. **Result:**
   - `return [i for i, is_prime in enumerate(primes) if is_prime]`: Creates a list of indices where `primes[i]` is True, representing the prime numbers.


Let me know if you'd like a more detailed explanation of any part! 
<end_of_turn><start_of_turn>user
Thank you, it works! Can you explain the code in French?<end_of_turn>
<start_of_turn>model
Bien sûr ! Voici une explication du code en français :

```python
def sieve_of_eratosthenes(n):
  """Retourne une liste de nombres premiers jusqu'à n."""
  primes = [True] * (n + 1)
  primes[0] = primes[1] = False
  for i in range(2, int(n**0.5) + 1):
    if primes[i]:
      for j in range(i * i, n + 1, i):
        primes[j] = False
  return [i for i, is_prime in enumerate(primes) if is_prime]

primes = sieve_of_eratosthenes(1000)
print(primes)
```

**Explication:**

1. **Initialisation:**
   - `primes = [True] * (n + 1)`: Crée une liste `primes` de valeurs booléennes, initialement supposant que tous les nombres sont premiers.
   - `primes[0] = primes[1] = False`: Définit 0 et 1 comme non-premiers.

2. **Itération:**
   - `for i in range(2, int(n**0.5) + 1):`: Itère de 2 jusqu'à la racine carrée de `n`. Nous ne devons vérifier que jusqu'à la racine carrée car tout nombre composite doit avoir un facteur premier inférieur ou égal à sa racine carrée.
   - `if primes[i]:`: Si `i` est considéré comme premier:
     - `for j in range(i * i, n + 1, i):`:  Marquer tous les multiples de `i` comme non-premiers.

3. **Resultat:**
   - `return [i for i, is_prime in enumerate(primes) if is_prime]`: Crée une liste des indices où `primes[i]` est vrai, représentant les nombres premiers.


N'hésitez pas à me demander si vous avez besoin d'une explication plus détaillée de quelque chose! 
<end_of_turn><start_of_turn>user
Great! Now add those explanations as comments in the code.<end_of_turn>
<start_of_turn>model

```python
def sieve_of_eratosthenes(n):
  """Retourne une liste de nombres premiers jusqu'à n."""
  # Initialise une liste de boolean avec True pour tous les nombres de 0 à n
  primes = [True] * (n + 1)
  # Définit 0 et 1 comme non-premiers
  primes[0] = primes[1] = False
  # Itère de 2 à la racine carrée de n
  for i in range(2, int(n**0.5) + 1):
    # Si i est considéré comme premier
    if primes[i]:
      # Itère sur tous les multiples de i
      for j in range(i * i, n + 1, i):
        # Définit les multiples de i comme non-premiers
        primes[j] = False
  # Retourne la liste des indices des nombres premiers
  return [i for i, is_prime in enumerate(primes) if is_prime]

primes = sieve_of_eratosthenes(1000)
print(primes)
```

**Explication:**

* **Initialisation:**
    * `primes = [True] * (n + 1)`:  Crée une liste `primes` de valeurs booléennes, initialement supposant que tous les nombres sont premiers.
    * `primes[0] = primes[1] = False`: Définit 0 et 1 comme non-premiers.
* **Itération:**
    * `for i in range(2, int(n**0.5) + 1):`: Itère de 2 jusqu'à la racine carrée de `n`. Nous ne devons vérifier que jusqu'à la racine carrée car tout nombre composite doit avoir un facteur premier inférieur ou égal à sa racine carrée.
    * `if primes[i]:`: Si `i` est considéré comme premier:
        * `for j in range(i * i, n + 1, i):`:  Marquer tous les multiples de `i` comme non-premiers.
* **Resultat:**
    * `return [i for i, is_prime in enumerate(primes) if is_prime]`: Crée une liste des indices où `primes[i]` est vrai, représentant les nombres premiers. 



<end_of_turn>

まとめと関連情報

このチュートリアルでは、JAX で Keras を使用して、Gemma 2B 指示用調整モデルとチャットする方法を学習しました。

Gemma の詳細については、以下のガイドとチュートリアルをご覧ください。