見出し画像

【Open Assistant】英語解説を日本語で読む【2023年4月25日|@Testing AI】

Open Assistantの解説です。
公開日:2023年4月25日
※動画を再生してから、インタビューを読むのがオススメです。


Hello, everybody.

皆さん、こんにちは!

In this video, I'm going to talk about Open Assistant, which is very similar to ChatGPT.

このビデオでは、ChatGPTと非常によく似たOpen Assistantについてお話します。

However, it is truly open source.

しかし、これは本当にオープンソースです。

That means all its data sets, its models and its code is available for the public to see and contribute.

つまり、すべてのデータセット、モデル、コードが公開されており、一般の人が見て貢献できるようになっています。

If this is your first time to my channel, I recommend everybody to go to my playlist section to find the videos that are relevant to you.

私のチャンネルを初めてご覧になる方は、プレイリストのセクションで自分に関連するビデオを見つけることをお勧めします。

All right, let's get started.

では、さっそく始めましょう。

So what exactly is Open Assistant?

さて、Open Assistantとは一体何なのでしょうか?

According to their GitHub page, it's a project meant to give everybody access to a great chat based large language model.

GitHubのページによると、チャットベースの大規模言語モデルを誰でも利用できるようにするためのプロジェクトだそうです。

So just like a lot of AIR tools are using stable diffusion in the backend to generate AI images, Open Assistant is aiming to have a bunch of open source model available for other people to use.

多くのAIRツールがAI画像を生成するためにバックエンドで安定拡散を使用しているように、Open Assistantは他の人が使用できるオープンソースモデルを多数用意することを目指しています。

You can find information of all the models made available by Open Assistant in this Hugging Face page that I'll link to my description.

オープンアシスタントが公開しているすべてのモデルの情報は、私の説明にリンクしているこのHugging Faceのページで見ることができます。

If you want to learn more about how Open Assistant collected its data and how it came up with its model, I'll have all the relevant links including this research paper from Cornell linked to my description.

Open Assistantがどのようにデータを収集し、モデルを構築したかについて詳しく知りたい場合は、コーネル大学の研究論文を含む関連リンクをすべて説明にリンクさせます。

But this is more of a beginner's tutorial as promised in my thumbnail.

しかし、これはサムネイルで約束したように、どちらかというと初心者向けのチュートリアルなのです。

So I'm not going to overwhelm you with information.

だから、情報であなたを圧倒するつもりはありません。

So all you need to do here is to come to this link called chat in its useful link section and click on the button called chat.

そこで必要なことは、便利なリンク集にあるチャットというリンクのところに来て、チャットというボタンをクリックすることです。

This opens up the chat front end of Open Assistant, which is similar to ChatGPT.

これで、ChatGPTと同じようなOpen Assistantのチャットフロントエンドが開かれます。

And if you don't have an account, make sure to create an account for free and log in.

そして、もしアカウントを持っていないのであれば、必ず無料でアカウントを作成してログインしてください。

I'm already logged in as you see.

私はご覧の通り、すでにログインしています。

Now there's a bunch of information on this page.

さて、このページにはたくさんの情報が掲載されています。

The only two sections you need to worry about here is the dashboard and the chat.

ここで気になるのは、ダッシュボードとチャットの2つのセクションだけです。

By the way, if you prefer the screen to be more of a dark mode style, you can go on the right here and click on this button right here.

ところで、画面をダークモード風にしたい場合は、この右側にあるボタンをクリックします。

That changes the page to be more of a black and white style, which I prefer.

そうすると、ページが白黒のスタイルに変わり、私の好みとなります。

So I'm going to keep it like this.

私はこのスタイルを維持するつもりです。

So if you scroll down on this dashboard, you're going to see it says grab a task.

このダッシュボードを下にスクロールすると、「タスクを取得する」と表示されます。

And you might be wondering why do I need to do a task?

なぜタスクを実行する必要があるのか、不思議に思うかもしれません。

So Open Assistant is an open source project.

Open Assistantは、オープンソースのプロジェクトです。

That means it gathers inputs from users like you and me and improves upon its model.

つまり、あなたや私のようなユーザーから情報を集め、そのモデルを改良しているのです。

So if you are like me who is a bit hesitant on paying monthly subscriptions, you can contribute by providing your own responses in training the model.

ですから、もしあなたが私のように毎月の購読料を支払うことに少し躊躇しているのであれば、モデルを訓練するためにあなた自身の反応を提供することで貢献することができます。

If you look closely, you're going to see this one says 281 tasks available, 216 tasks available.

よく見ると、281のタスクが利用可能、216のタスクが利用可能、と書いてありますね。

And if you click on any one of them, you're going to see there's a bunch of things you can do and help out the model.

そのうちのどれかをクリックすると、モデルを助けるためにできることがたくさんあることがわかると思います。

For example, here someone asked this question and Open Assistant made this reply.

例えば、ここでは誰かがこのような質問をし、Open Assistantがこのような返答をしています。

And how you can contribute here is to answer the following questions based on the reply Open Assistant gave.

ここであなたができることは、Open Assistantの返答をもとに、以下の質問に答えることです。

Is this message spam?

このメッセージはスパムか?

Is it a bad reply?

悪い返事なのか?

Select any of these which apply to the highlighted message and rate the highlighted message.

これらのうち、ハイライトされたメッセージに該当するものを選び、ハイライトされたメッセージを評価してください。

Is it low quality?

低品質ですか?

Is it unhelpful?

役に立ちませんか?

So these are all information that trains and improves the responses of Open Assistant.

これらはすべて、Open Assistantの対応を訓練し、向上させるための情報なんですね。

And I think this is both a great way to contribute to this open source project and also learn about different kinds of prompts.

これは、このオープンソースプロジェクトに貢献すると同時に、さまざまな種類のプロンプトについて学ぶことができる素晴らしい方法だと思います。

That way you both improve your own prompt engineering skills while helping out this model.

そうすれば、このモデルを支援しながら、自分のプロンプトエンジニアリングスキルを向上させることができます。

By the way, if you're unclear on how to complete any of these tasks, you can just click on this button on the left called guidelines and that will give you all the information you need to participate in training the model.

ところで、これらのタスクのいずれかを完了する方法がわからない場合は、左側にある「ガイドライン」というボタンをクリックするだけで、モデルのトレーニングに参加するために必要な情報がすべて得られます。

All right, after playing around with a bunch of tasks, you're going to get an idea of how the assistant works.

さて、たくさんのタスクで遊んでいるうちに、アシスタントの仕組みがわかってきましたね。

And after that, all you do is to click this button called Try Our Assistant.

その後、「Try Our Assistant」というボタンをクリックすればよいのです。

And this opens up an interface which looks very similar to Playground AI where you can interact with the chatbot.

すると、Playground AIによく似たインターフェースが開き、チャットボットと対話することができます。

By the way, I recommend naming all of your chats.

ちなみに、チャットにはすべて名前をつけることをおすすめします。

You might see both of these stays untitled because I have not named them.

この2つのチャットが無題のままになっているのが見えるかもしれませんが、これは私が名前を付けていないためです。

So if you click this edit button right here, you can add a name.

そこで、この編集ボタンをクリックすると、名前を付けることができます。

So the first prompt I'm going to use is related to a buzzword I've been hearing a lot lately and that's called de-dollarization.

最初のプロンプトは、最近よく耳にする「デドラーライゼーション」という言葉に関連したものです。

Explain me de-dollarization in simple terms.

デドラーライゼーションについて簡単に説明してください。

Press enter.

エンターキーを押してください。

Okay, it says de-dollarization is a process in which countries reduce their dependence on the US dollar as a reserve currency or shift away from using the US dollars for international trade.

デドル化とは、国々が米ドルを準備通貨としての依存度を減らしたり、国際取引で米ドルを使用することから離れたりするプロセスだと言っています。

The US dollar has been used as a global currency since World War Two and it's used to settle most international transactions.

米ドルは第二次世界大戦以降、世界的な通貨として使用されており、ほとんどの国際取引の決済に使用されています。

If you look on the right, it says Model LLaMA 30B.

右を見ると、Model LLaMA 30Bと書かれています。

And there's no other options because right now that's the only model that's available on the front end for Open Assistant.

そして、今現在、Open Assistantのフロントエンドで利用できるのはこのモデルだけなので、他の選択肢はありません。

Also, I just saw that I was kind of highlighted on the first prompt and I renamed the second prompt to be decided for now.

また、最初のプロンプトで強調表示されていたのを確認したので、2番目のプロンプトの名前を変更して、とりあえず決定することにしました。

And what I'm going to do is I'm going to try to change the structure of this response.

そして、これからやることは、この回答の構造を変えてみることです。

I'm going to say change that to bullet points.

箇条書きにしようと思います。

All right, it has changed the entire response to bullet points.

よし、これでレス全体が箇条書きに変わりましたね。

I do notice that the actual bullet points are more like paragraphs so I don't know if that's intentional.

ただ、実際の箇条書きは段落に近いので、それが意図的なものかどうかは分かりませんが。

All right, now what I'm going to do is I'm going to try to change the language.

それでは、これから文言を変えてみようと思います。

I'm going to say change the language to Chinese.

言語を中国語に変更してみます。

Okay, it's starting to give me the response in Chinese.

さて、中国語で応答が返ってくるようになりました。

I see some question marks.

クエスチョンマークが見えますが、これは何でしょうか?

I'm not sure if that's expected or it's just my browser not rendering them properly.

これは予想通りなのか、それとも私のブラウザが正しくレンダリングできていないだけなのか、よくわかりません。

If you're Chinese speaking, maybe you can comment and say if this is a proper response.

もしあなたが中国語を話せるなら、これが適切な回答かどうか、コメントできるかもしれません。

I see there are some words still in English.

まだ英語の単語もあるんですね。

Okay, so it seems to have converted the whole thing to Chinese, but I still see some English letters right here.

なるほど、全体が中国語に変換されたようですが、まだここに英語の文字がありますね。

There's the word cleanse as well.

cleanseという単語もありますね。

So I'm not sure if that is expected.

ということは、想定内なのかな。

So now I'm going to say change the response to an English sonnet.

だから、今度は英語のソネットにレスを変えろと言うことです。

Dollar hegemony reigns supreme indeed.

ドルの覇権は確かに君臨している。

A fact that some nations view with dread, for though it serves as global currency, its impact can be quite capricious.

世界的な通貨でありながら、その影響力は非常に気まぐれであるため、一部の国々はこの事実を恐れている。

Each nation must maintain large hordes of greenbacks held in foreign shores to facilitate cross border dealings but lack control or its worth and measure.

各国は、国境を越えた取引を円滑に行うために、外国に大量のドル紙幣を保有しなければならないが、その価値や尺度をコントロールすることはできない。

China, Russia, both among the ranks that voice concerns about dollars instability, which threatens much greater risk and reward in today's tumultuous global scenery.

中国とロシアは、ドルの不安定性に懸念を示す国の中にあり、今日の激動する世界情勢において、はるかに大きなリスクと報酬が脅かされています。

And while some fears may seem far fetched, others are grounded in real politic.

そして、ある懸念は遠回しに見えるかもしれないが、他の懸念は現実の政治に根ざしている。

For what would come should America drift?

アメリカはどうなるのだろう?

These questions weigh heavily upon us all.

これらの疑問は、私たち全員に重くのしかかる。

I had to read the whole thing because that was actually pretty good.

私は、この本がとても良かったので、全部読みました。

And I'm pretty impressed by this sonnet actually.

そして、このソネットにはかなり感心させられました。

So I'm going to go ahead and give it a thumbs up.

だから、私はこの作品に親指を立てるつもりです。

For the last one, I'm actually going to give it a thumbs down because I see some English letters come up.

最後の1つは、英字がいくつか出てきたので、実は親指を立てることにします。

All right, for the next one, I'm going to do some programming.

さて、次はプログラミングをやってみます。

And this time that's to Python LangChain.

今回はPythonのLangChainを作るんだ。

So LangChain is a new Python library that came along and I'm trying to learn it.

LangChainはPythonの新しいライブラリで、私はそれを学ぼうとしています。

So what I'm going to do is explain by tons Lang chain library using some code.

そこで今回は、LangChainライブラリのトンを、いくつかのコードを使って説明しようと思います。

Okay, it says Python language has a vast collection of libraries that can be used to perform various tasks.

さて、Python言語には様々なタスクを実行するために使用できる膨大なライブラリのコレクションがあると書かれています。

And it actually gave me the math library which I did not request.

そして、私が要求していない数学ライブラリを実際に提供してくれました。

So I think it completely missed the point here and did not understand what LangChain means at all.

つまり、LangChainの意味を全く理解していない、完全にポイントを外しているのだと思います。

So if you don't know LangChain is a framework for developing applications powered by language models, and I was hoping Open Assistant could help me out there, but I don't think it understands LangChain yet, or maybe it's really new, and it's not in the system.

もし知らない場合、LangChainは言語モデルによって駆動されるアプリケーションを開発するためのフレームワークであり、私はOpen Assistantがそこで助けてくれることを期待していましたが、まだLangChainを理解していないかもしれませんし、それは本当に新しいので、システムには含まれていないと思います。

But I was thinking it could access the internet, but maybe I was wrong.

でも、インターネットにはアクセスできると思っていたのですが、もしかしたら間違っていたのかもしれません。

All right, let me start a new chat and do some actual programming.

よし、新しいチャットを立ち上げて、実際にプログラミングをやってみよう。

Give me the basics of JavaScript as a cheat sheet.

カンニングペーパーとしてJavaScriptの基本を教えてくれ。

So it says here's a brief overview of the basic concepts of JavaScript.

ということで、JavaScriptの基本的な概念について簡単に説明しますと書いてありますね。

Okay, it seems to have given me a lot of responses which are more informational and text based versus code examples because I was expecting cheat sheet to be something different.

どうやら、私がチートシートとは違うものを期待していたので、情報提供とテキストベースの回答が多く、コードの例が少ないようです。

But here it seems to be more of an article.

というのも、チートシートはもっと違うものだと思っていたのですが、ここではもっと記事的なものだったようです。

And it kind of stopped here.

そして、ここでちょっと止まってしまいました。

Let me see if it continues.

続きがあれば見てみます。

Please continue.

続けてください。

Okay, so I don't seem to have an option to let this specific chat continue.

このチャットを継続させるオプションはないようですね。

And it starts again from scratch.

そして、また一から始まります。

And this time it does a brand new reference sheet.

そして、今回は全く新しいリファレンスシートを作成します。

And I don't seem to have an option to stop this generation.

そして、この世代を止めるオプションはないようです。

And I don't know if that is by design, or I just don't know how to do it.

それがデザインによるものなのか、それとも私がやり方を知らないだけなのかは分かりませんが。

But I can't seem to stop this generation because I don't think it understands what a cheat sheet means here because it's giving me full scale definitions of things like big int in paragraphs which are written in English.

しかし、私はこれを止めることができないようです。なぜなら、チートシートという言葉の意味を理解していないと思われるためです。例えば、ビッグインティといったものの完全な定義を英語で書かれた段落で提供してくれています。

So I don't think it did the job of creating a cheat sheet that well.

つまり、この世代はチートシートを作るという仕事をあまりしていないと思うのです。

Let me try a different kind of cheat sheet because I do want to try generating some kind of cheat sheet.

では、別の種類のチートシートを試してみましょう。ある種のチートシートを生成してみたいのです。

So I'm going to do cheat sheet for regular expressions.

というわけで、正規表現のチートシートを作ってみようと思います。

If you don't know, by the way, regular expressions are a way to find patterns in different kinds of text and identify the patterns and scrape out different parts of text.

ちなみに、もし知らないのであれば、正規表現は様々な種類のテキストでパターンを見つけ、それらのパターンを識別し、テキストの異なる部分を取り出す方法です。

It has given me a lot of information, but I still don't think it understands what a cheat sheet means.

多くの情報を与えてくれたが、まだチートシートの意味を理解していないようだ。

And I think I'm going to let it go.

そして、私はそれを手放そうと思う。

Let me try a last example of implementing an API.

最後にAPIを実装する例をやってみます。

API use, give me a code example of using Java API.

APIの使用、Java APIを使用するコード例を教えてください。

Okay, it says this code reads the content of example.txt and prints out the output.

なるほど、このコードはexample.txtの内容を読み取り、出力を出力すると書いてありますね。

Okay, let me change that to give me a code example of using the stable diffusion API.

では、安定した拡散APIを使用するコード例を教えてくださいに変更してみます。

Here's an example of a Python script using the stable diffusion API says PIL.

安定した拡散APIを使用したPythonスクリプトの例として、PILと書かれています。

I think that's the stable diffusion API library.

これは安定型拡散APIライブラリだと思います。

It imports this method called image.

このimageというメソッドをインポートしています。

There's this URL from stable diffusion.

stable diffusionからこのようなURLがあります。

There's an API key and it opens the image.

APIキーがあり、画像を開く。

It converts the image into bytes for uploading.

アップロードするために画像をバイトに変換する。

It uploads the image to stable diffusion.

画像をstable diffusionにアップロードします。

It runs the diffusion process and it gets the image path from the JSON response.

拡散処理を実行し、JSONレスポンスから画像パスを取得する。

So I think this was a much better response than the ones I'm getting.

というわけで、これは私が得ているものよりもずっと良い反応だったと思います。

I think there's definitely some quirks and I might need to improve my own prompting anyways, but I definitely think there's a lot of potential here because this is open source, so it's going to improve over time.

確かにいくつかの癖があり、私自身のプロンプトを改善する必要があるかもしれませんが、これはオープンソースなので、時間とともに改善されることは間違いないと思います。

So I'm going to do a last query with something I was doing in my last few videos.

というわけで、最後のクエリは、私がここ数回のビデオでやっていたことをやってみようと思っています。

That is my newsletter, email newsletter for my AI YouTube channel.

それは、私のニュースレター、私のAI YouTubeチャンネルのメールマガジンです。

The latest and greatest from AI channel name.

AIチャンネル名から最新・最高のものをお届けします。

It's been a busy few weeks here.

ここ数週間は忙しかったですね。

We wanted to share some exciting updates.

エキサイティングな最新情報を共有したいと思います。

We hit a major milestone, 10,000 subscribers.

チャンネル登録者数10,000人という大きなマイルストーンを達成しました。

I'm not there yet.

まだ到達していませんが

Hopefully reaching there soon.

すぐにでも達成したい

Okay, the structure isn't that bad.

さて、構成はそれほど悪くはありません。

Let me update this email and say, add five AI art tools to that email.

このメールを更新して、AIアートツールを5つ追加してみましょう。

All right, I think I might have broken Open Assistant already or maybe it's my internet.

もうOpen Assistantが壊れてしまったのか、それとも私のインターネットのせいなのか。

I don't know.

よくわかりません。

Let me try that again.

もう一回やってみます。

Okay, so here's a revised version of that email newsletter, including information about five popular AI art tools.

では、そのメールマガジンの改訂版ですが、人気のある5つのAIアートツールについての情報が含まれています。

It mentions tally to deep art, photo, sketcher, stable diffusion, latent layers, and the email gets cut off again.

タリーからディープアート、フォト、スケッチャー、安定した拡散、潜像層について触れていて、またメールが切れてしまう。

And I don't know how to proceed when it gets cut off.

そして、切られたときにどう進めていいのかわからない。

And I think I just have to recreate the rest of the email myself.

そして、残りのメールを自分で作り直すしかないと思っています。

Let me do my last challenge and try to break it again.

前回のチャレンジで、もう一度壊してみようか。

This time I'm going to say, now change those five AI art tools to be ones which are discord only.

今回は、今度はその5つのAIアートツールを、ディスコード専用のものに変えてみてくださいということです。

So I wanted to mention only the tools which are available on discord.

だから、discordで使えるツールだけを挙げたいと思った。

For example, mid journey or blue below, like not Leonardo AI.

例えば、ミッドジャーニーとか、下のブルーとか、レオナルドAIじゃないみたいな。

So I want to see how it reacts.

だから、その反応を見てみたいんです。

Here's updated newsletter, including five AI art tools exclusive to discord platforms.

discordプラットフォーム限定の5つのAIアートツールを含む、アップデートされたニュースレターがこちらです。

Okay, so these are tools I haven't used and it never mentioned mid journey at all.

なるほど、これらは私が使ったことのないツールで、ミッドジャーニーについてはまったく触れられていない。

I wonder if that's because these tools are new and I wanted to mention new tools only.

これらのツールは新しいものなので、新しいツールだけに言及したかったからでしょうかね。

But I think I tested Open Assistant enough today and you got an idea of how it works.

でも、今日はOpen Assistantを十分に試したので、どのように機能するかはお分かりいただけたかと思います。

Let me know if you have been able to use it properly and it has helped you out in your own work.

もし、あなたがちゃんと使えて、自分の仕事に役立ったのなら教えてください。

And that's all I have for today's video.

今日の動画は以上です。

If you did enjoy it, make sure to click the like button and subscribe to my channel.

もし楽しんでいただけたなら、「いいね!」ボタンをクリックし、私のチャンネルを購読してください。

Till the next one, thank you so much.

次回まで、よろしくお願いします。

この記事が気に入ったらサポートをしてみませんか?