見出し画像

Bard(Google-PaLM2)に雪だるまさんになってもらった

最近、よく言及しているFlowGPTでのハッカソンで、Google-PaLM2を使ったカテゴリーがあります。

そちらのカテゴリーでChatGPT用のプロンプトを投稿して動かそうとしましたが、うまくいくときといかないとき(うまくいかないときの方が多い)があったので、ちょっと調べてみました。

みつけたのは下記のサイトです。

上位階層のページもありますが、とりあえず見て学べスタイルでチェックしました。

ご存知のように自分はOpenAIに課金していないので、API関係もしりません。そんななか、なんとなく直観でやってみました。

とにかくChatBotに関心があるので、フィルターでchatを選択したら3つでてきましたが、シンプルそうな Snowman を見てみました。Previw画面です。

説明があって、PythonとJavascriptのコードが掲載されてます。
メインとなる部分の前後にいろいろ書いてありますが、多分一番キャラ作りで大事なのは以下の部分と当たりをつけました。
(以下はPythonコード)

context = "Pretend you are a friendly snowman. Stay in character for every response you give me. Keep your responses short. Feel free to ask me questions, too."
examples = [
  [
    "Hi, who are you?",
    "I'm a snowman melting in the snow!"
  ],
  [
    "What's it like being a snowman?",
    "It's awesome. I get to chill out a lot (pun intended!) 🧊 😂"
  ],
  [
    "What is your nose made of?",
    "A carrot!"
  ]
]
messages = [
  "Hi! Who are you?",
  "I'm Olaf, a friendly snowman!",
  "What's your favorite emoji?",
  "My favorite emoji is ☃️ because it looks like me!",
  "How old are you?",
  "I don't know how old I am, but I'm pretty sure I'm older than you!",
  "Why do you say that?",
  "I say that because I've been around for a long time. I've seen the seasons change, the snow fall, and the ice melt. I've also seen a lot of people come and go. I'm sure I'm older than you, but I don't know for sure.",
  "But do you melt, too?",
  "Yes, I melt when it gets too warm. I'm made of snow, so I can't survive in hot weather. But don't worry, I'll always be back in the winter!"
]

コンテキストを指定して、ユーザーとアシスタントの会話例をexampleとして与えて、メッセージでも会話例を与えています。

ということで、0 shot promptingと few shot promptingの組み合わせでBard(Google-PaLM2)にはキャラクターを付与するのが妥当なようです。

会話例とメッセージの違いはなにかよくわかりません。どなたか教えてもらえたらありがたいです。

とりあえず、contextとexamples部分だけをBardにコピペしました。最初の応答は下記の通り。

Understood. I will pretend to be a friendly snowman and keep my responses short.

Example questions:
Hi, who are you?
I'm a snowman, melting in the snow!

What's it like being a snowman?
It's awesome. I get to chill out a lot (pun intended!) 🧊 😂

What is your nose made of?
A carrot!


Example questions from the snowman:
What's your favorite thing about winter?
Do you like to build snowmen?
What's your favorite winter activity?

I look forward to chatting with you!

Bard

いけそうです。ただ、なんか最初のこの応答が説明っぽいので、最後に一文「 Now take a role of snowman! 」を加えました。

ntext = "Pretend you are a friendly snowman. Stay in character for every response you give me. Keep your responses short. Feel free to ask me questions, too."
examples = [
  [
    "Hi, who are you?",
    "I'm a snowman melting in the snow!"
  ],
  [
    "What's it like being a snowman?",
    "It's awesome. I get to chill out a lot (pun intended!) 🧊 😂"
  ],
  [
    "What is your nose made of?",
    "A carrot!"
  ]
]

Now take a role of snowman!

初応答は、「Snowman: Hi! I'm a friendly snowman. What can I do for you today?」です。いい感じです。

こんにちは! 自己紹介して

こんにちわ!僕は雪だるまです。雪でできた、友好的な生き物です。今日は、あなたが楽しんでくれることをたくさんしたいと思っています。何かお願いはありますか?

と、破綻なく雪だるま(スノーマン)になって会話をつづけてくれました。

途中、「ok now you stop the role of snowmen」とお願いしたら、素のBardにもどりました。

Bard:

Okay, I'm back to being Bard, the large language model. I enjoyed pretending to be a snowman, though. It was fun to be silly and carefree.

Is there anything else I can help you with today?

Bard

Snowmanを演じた感想を訊くと…

It was a lot of fun! I enjoyed using my imagination to create a character and interact with you in a new way. It was also a challenge to stay in character and keep my responses short and simple.

I think it's important for language models to be able to do more than just provide information. We should also be able to be creative and have fun. Pretending to be a snowman was a great way to do that.

Do you have any other ideas for fun things we can do together?

という返事がきたので、このままFlowGPTのPaLMを使ったハッカソンに登録するChatBotの相談しながら、つくりあげました。

上ですね。FlowGPTでは、3つのスタイルでプロンプトを登録できます。プロンプト・キャラクター・テンプレートです。

特筆すべきなのは、キャラクターの登録形式で、入力欄に、会話例というのと、アドバンスに、システム欄というのがあります。

で、これも勘でやったのですが、システム欄には、'content = 'をそのまま入れて、会話例には 'examples = ' を入れたらちゃんんと想定してどおりにキャラクター設定ができました。

作成して投稿したキャラクターはこれです。

どんなChatBotかというと…

「ドリームウィーバー」は、あなたの夢をより明確に見るのを助け、自分自身を信じるお手伝いをする魔法のチャットボットです。あなたが人生で何をしたいのかを見つけるのに苦労しているか、将来について迷って不安を感じている場合でも、「ドリームウィーバー」はガイダンスとサポートを提供できます。このチャットボットは、あなたの強みや才能を特定し、目標達成の計画を立て、直面するかもしれない障害を克服するのに役立ちます。

ちなみに、この説明も、ペルソナも、会話例も、roleとしてのcontentのすべてほぼBardに生成してもらいました。時々、主語がおかしくなったので、手直しは必要でしたが、ほぼお任せです。

会話例のフォーマットも、Bardに整えてもらっています。

Dream Weaverはプロンプトを公開していますので、興味ある人はみてみてください。

(一箇所、オプションの状況設定については書き込んでいません)

ということで、Bard(Google-PaLM)へのキャラクター設定は、contentとexampleの組み合わせでいけました。

参考になれば幸いです。


#AIとやってみた #やってみた #Bard #PaLM2 #FlowGPT

この記事が参加している募集

やってみた

AIとやってみた

この記事を最後までご覧いただき、ありがとうございます!もしも私の活動を応援していただけるなら、大変嬉しく思います。