I'm Megha.

我叫梅加

Today I'm going to be talking about large language models.

今天我要說的是大型語言模型。

Don't know what those are?

不知道這些是什麼？

Me either.

我也是。

Just kidding.

開個玩笑

I actually know what I'm talking about.

其實我知道自己在說什麼。

I'm a customer engineer here at Google Cloud, and today I'm going to teach you everything you need to know about LLMs.

我是 Google Cloud 的客戶工程師，今天我將向你傳授關於 LLM 的一切知識。

That's short for large language models.

這是大型語言模型的簡稱。

In this course, you're going to learn to define large language models, describe LLM use cases, explain prompt tuning, and describe Google's generative AI development tools.

在本課程中，您將學習如何定義大型語言模型、描述 LLM 用例、解釋提示調整並介紹 Google 的生成式人工智能開發工具。

Let's get into it.

讓我們開始吧。

Large language models, or LLMs, are a subset of deep learning.

大型語言模型或 LLM 是深度學習的一個子集。

To find out more about deep learning, check out our Introduction to Generative AI course video.

要了解有關深度學習的更多資訊，請觀看我們的 "生成式人工智能入門 "課程視頻。

LLMs and generative AI intersect and they are both a part of deep learning.

LLM 和生成式人工智能相互交叉，它們都是深度學習的一部分。

Another area of AI you may be hearing a lot about is generative AI.

您可能經常聽到的另一個人工智能領域是生成式人工智能。

This is a type of artificial intelligence that can produce new content including text, images, audio, and synthetic data.

這是一種人工智能，可以生成新的內容，包括文本、影像、音頻和合成數據。

All right, back to LLMs.

好了，說回法學碩士。

So what are large language models?

那麼，什麼是大型語言模型呢？

Large language models refer to large, general purpose language models that can be pre-trained and then fine-tuned for specific purposes.

大型語言模型指的是大型通用語言模型，可以進行預訓練，然後針對特定目的進行微調。

What do pre-trained and fine-tuned mean?

預訓練和微調是什麼意思？

Great questions.

問得好

Let's dive in.

讓我們深入瞭解一下。

Imagine training a dog.

想象一下訓練一條狗。

Often you train your dog basic commands such as sit, come, down, and stay.

通常情況下，您會訓練狗狗基本的指令，如坐、來、趴下和停留。

These commands are normally sufficient for everyday life and help your dog become a good canine citizen.

這些指令通常足以應付日常生活，並能幫助您的愛犬成為良好的犬類公民。

Good boy.

好孩子

But if you need special service dogs such as a police dog, a guide dog, or a hunting dog, you add special trainings, right?

但如果你需要特殊的服務犬，如警犬、導盲犬或獵犬，你就需要增加特殊訓練，對嗎？

A similar idea applies to large language models.

類似的想法也適用於大型語言模型。

These models are trained for general purposes to solve common language problems such as text classification, question answering, document summarization, and text generation across industries.

這些模型經過訓練後可用於解決常見的語言問題，如文本分類、問題解答、文檔摘要和跨行業文本生成。

The models can then be tailored to solve specific problems in different fields such as retail, finance, and entertainment using a relatively small size of field datasets.

然後，就可以利用相對較小的現場數據集對這些模型進行定製，以解決零售、金融和娛樂等不同領域的具體問題。

So now that you've got that down, let's further break down the concept into three major features of large language models.

既然你已經明白了這一點，讓我們進一步將這一概念分解為大型語言模型的三大特徵。

We'll start with the word large.

我們從 "大 "字開始。

Large indicates two meanings.

大表示兩種含義。

First is the enormous size of the training dataset, sometimes at the petabyte scale.

首先是訓練數據集的巨大規模，有時甚至達到 PB 級。

Second, it refers to the parameter count.

其次，它指的是參數計數。

In machine learning, parameters are often called hyperparameters.

在機器學習中，參數通常被稱為超參數。

Parameters are basically the memories and the knowledge the machine learned from the model training.

參數基本上是機器從模型訓練中學到的記憶和知識。

Parameters define the skill of a model in solving a problem such as predicting text.

參數定義了模型解決文本預測等問題的技能。

So that's why we use the word large.

所以，我們才用 "大 "這個字。

What about general purpose?

通用性如何？

General purpose is when the models are sufficient to solve common problems.

通用型是指模型足以解決常見問題。

Two reasons led to this idea.

產生這一想法有兩個原因。

First is the commonality of human language regardless of the specific tasks.

首先是人類語言的共性，無論具體任務如何。

And second is the resource restriction.

其次是資源限制。

Only certain organizations have the capability to train such large language models with huge datasets and a tremendous number of parameters.

只有某些機構有能力利用龐大的數據集和大量的參數來訓練這種大型語言模型。

How about letting them create fundamental language models for others to use?

讓他們創建基本語言模型供他人使用如何？

So this leaves us with our last terms, pre-trained and fine-tuned, which mean to pre-train a large model for a general purpose with a large dataset and then fine-tune it for specific aims with a much smaller dataset.

這就剩下了最後一個術語，預訓練和微調，這意味著使用大型數據集為通用目的預訓練一個大型模型，然後使用小得多的數據集為特定目的對其進行微調。

So now that we've nailed down the definition of what large language models LLMs are, we can move on to describing LLM use cases.

既然我們已經明確了大型語言模型 LLM 的定義，那麼接下來就可以描述 LLM 的使用案例了。

The benefits of using large language models are straightforward.

使用大型語言模型的好處顯而易見。

First, a single model can be used for different tasks.

首先，一個模型可用於不同的任務。

This is a dream come true.

這真是夢想成真。

These large language models that are trained with petabytes of data and generate billions of parameters are smart enough to solve different tasks, including language translation, sentence completion, text classification, question answering, and more.

這些大型語言模型使用 PB 級數據進行訓練，可生成數十億個參數，足以解決不同的任務，包括語言翻譯、句子補全、文本分類、問題解答等。

Second, large language models require minimal field training data when you tailor them to solve a specific problem.

其次，大型語言模型為解決特定問題而量身定製時，需要的現場訓練數據極少。

Large language models obtain decent performance even with little domain training data.

即使只有很少的領域訓練數據，大型語言模型也能獲得不錯的性能。

In other words, they can be used for few-shot or even zero-shot scenarios.

換句話說，它們可以用於少發甚至零發的情況。

In machine learning, few-shot refers to training a model with minimal data, and zero-shot implies that a model can recognize things that have not explicitly been taught in the training before.

在機器學習中，"少數據 "指的是用最少的數據來訓練模型，而 "零數據 "指的是模型能夠識別之前訓練中沒有明確教授的事物。

Third, the performance of large language models is continuously growing when you add more data and parameters.

第三，當您添加更多數據和參數時，大型語言模型的性能會不斷提高。

Let's take POM as an example.

讓我們以 POM 為例。

In April 2022, Google released POM, short for Pathways Language Model, a 540 billion parameter model that achieves a state-of-the-art performance across multiple language tasks.

2022 年 4 月，谷歌發佈了 POM（Pathways Language Model 的縮寫），這是一個 5400 億參數的模型，在多個語言任務中實現了最先進的性能。

POM is a dense decoder-only transformer model.

POM 是一個僅有密集解碼器的變壓器模型。

It leverages a new pathway system which enabled Google to efficiently train a single model across multiple TPU v4 pods.

它利用新的路徑系統，使谷歌能夠在多個 TPU v4 pod 上高效地訓練一個模型。

Pathways is a new AI architecture that will handle many tasks at once, learn new tasks quickly, and reflect a better understanding of the world.

Pathways 是一種全新的人工智能架構，可以同時處理多項任務，快速學習新任務，並反映出對世界的更好理解。

The system enables POM to orchestrate distributed computation for accelerators, but I'm getting ahead of myself.

該系統使 POM 能夠為加速器協調分佈式計算，但我說得太多了。

I previously mentioned that POM is a transformer model.

我之前提到過，POM 是一種變壓器模型。

Let me explain what that means.

讓我來解釋一下這意味著什麼。

A transformer model consists of an encoder and a decoder.

變壓器模型由編碼器和解碼器組成。

The encoder encodes the input sequence and passes it to the decoder, which learns how to decode the representations for a relevant task.

編碼器對輸入序列進行編碼，然後將其傳遞給解碼器，解碼器學習如何對相關任務的表徵進行解碼。

We've come a long way from traditional programming to neural networks to generative models.

從傳統編程到神經網絡，再到生成模型，我們已經走過了漫長的道路。

In traditional programming, we used to have to hard code the rules for distinguishing a cat.

在傳統編程中，我們過去不得不硬編碼區分貓的規則。

Type, animal, legs 4, ears 2, fur yes, likes, yarn and catnip.

類型，動物，4 條腿，2 只耳朵，有毛，喜歡毛線和貓薄荷。

In the wave of neural networks, we could give the network pictures of cats and dogs and ask, is this a cat?

在神經網絡浪潮中，我們可以給網絡提供貓和狗的圖片，然後問：這是貓嗎？

And they would predict a cat.

他們會預測一隻貓。

What's really cool is that in the generative wave, we as users can generate our own content, whether it be text, images, audio, video, or more.

最酷的是，在生成浪潮中，作為用戶，我們可以生成自己的內容，無論是文字、圖片、音頻、視頻還是其他內容。

For example, models like POM, or pathways language model, or Lambda, language model for dialogue applications, ingest very, very large data from multiple sources across the internet, and build foundation language models we can use simply by asking a question, whether typing it into a prompt or verbally talking into the prompt itself.

例如，像 POM 這樣的模型，或路徑語言模型，或用於對話應用的語言模型 Lambda，可以從互聯網上的多個來源獲取非常非常龐大的數據，並建立基礎語言模型，我們只需提出一個問題，無論是將問題輸入到提示符中，還是對著提示符本身進行口頭交談，就可以使用這些模型。

So when you ask it, what's a cat?

那麼，當你問它，貓是什麼？

It can give you everything it has learned about a cat.

它可以向你提供關於貓的一切資訊。

Let's compare LLM development using pre-trained models with traditional ML development.

讓我們將使用預訓練模型的 LLM 開發與傳統的 ML 開發進行比較。

First, with LLM development, you don't need to be an expert.

首先，在 LLM 開發方面，您不需要成為專家。

You don't need training examples, and there is no need to train a model.

您不需要訓練示例，也不需要訓練模型。

All you need to do is think about prompt design, which is a process of creating a prompt that is clear, concise, and informative.

您所需要做的就是思考提示語的設計，這是一個創建清晰、簡潔、資訊豐富的提示語的過程。

It is an important part of natural language processing, or NLP for short.

它是自然語言處理（簡稱 NLP）的重要組成部分。

In traditional machine learning, you need expertise, training examples, compute time, and hardware.

在傳統的機器學習中，你需要專業知識、訓練示例、計算時間和硬件。

That's a lot more requirements than LLM development.

這比法律碩士的發展要求要高得多。

Let's take a look at an example of a text generation use case to really drive the point home.

讓我們來看一個文本生成使用案例的例子，以便真正理解這一點。

Question answering, or QA, is a subfield of natural language processing that deals with the task of automatically answering questions posed in natural language.

問題解答或 QA 是自然語言處理的一個子領域，主要處理自動回答用自然語言提出的問題的任務。

QA systems are typically trained on a large amount of text and code, and they are able to answer a wide range of questions, including factual, definitional, and opinion-based questions.

質量保證系統通常會在大量文本和代碼的基礎上進行訓練，能夠回答各種問題，包括事實性問題、定義性問題和觀點性問題。

The key here is that you needed domain knowledge to develop these question answering models.

關鍵在於，開發這些問題解答模型需要領域知識。

Let's make this clear with a real-world example.

讓我們用一個真實的例子來說明這一點。

Domain knowledge is required to develop a question answering model for customer IT support, or healthcare, or supply chain.

開發客戶 IT 支持、醫療保健或供應鏈問題解答模型需要領域知識。

But using generative QA, the model generates free text directly based on the context.

但使用生成式質量保證時，模型會根據上下文直接生成自由文本。

There's no need for domain knowledge.

無需領域知識。

Let me show you a few examples of how cool this is.

讓我舉例說明這有多酷。

Let's look at three questions given to Gemini, a large language model chatbot developed by Google AI.

讓我們來看看由谷歌人工智能開發的大型語言模型哈拉機器人雙子座提出的三個問題。

Question one.

問題一

This year's sales are $100,000.

今年的銷售額為 100 000 美元。

Expenses are $60,000.

支出為 60 000 美元。

How much is net profit?

淨利潤是多少？

Gemini first shares how net profit is calculated, then performs the calculation.

雙子座首先介紹淨利潤的計算方法，然後進行計算。

Then Gemini provides the definition of net profit.

然後，雙子座提供了淨利潤的定義。

Here's another question.

還有一個問題。

Inventory on hand is 6,000 units.

庫存為 6000 件。

A new order requires 8,000 units.

一個新訂單需要 8000 個組織、部門。

How many units do I need to fill to complete the order?

我需要填寫多少單元才能完成訂單？

Again, Gemini answers the question by performing the calculation.

雙子座通過計算再次回答了問題。

And our last example.

最後一個例子

We have 1,000 sensors in 10 geographic regions.

我們在 10 個地區擁有 1,000 個傳感器。

How many sensors do we have on average in each region?

每個地區平均有多少個傳感器？

Gemini answers the question with an example on how to solve the problem and some additional context.

雙子座在回答問題時，舉例說明了如何解決問題，並提供了一些額外的背景資料。

So how is that?

怎麼樣？

In each of our questions, a desired response was obtained.

在我們提出的每一個問題中，都得到了理想的答覆。

This is due to prompt design.

這是因為設計及時。

Fancy.

花哨

Prompt design and prompt engineering are two closely related concepts in natural language processing.

提示設計和提示工程是自然語言處理中兩個密切相關的概念。

Both involve the process of creating a prompt that is clear, concise, and informative.

二者都涉及到創建一個清晰、簡潔、內容豐富的提示的過程。

But there are some key differences between the two.

但兩者之間也有一些主要區別。

Prompt design is the process of creating a prompt that is tailored to the specific task the system is being asked to perform.

提示設計是針對系統需要執行的特定任務創建提示的過程。

For example, if the system is being asked to translate a text from English to French, the prompt should be written in English and should specify that the translation should be in French.

例如，如果要求系統將文本從英文翻譯成法文，則提示語應以英文書寫，並指明翻譯應為法文。

Prompt engineering is the process of creating a prompt that is designed to improve performance.

提示工程是創建旨在提高績效的提示的過程。

This may involve using domain-specific knowledge, providing examples of the desired output, or using keywords that are known to be effective for the specific system.

這可能涉及使用特定領域的知識、提供所需輸出的示例或使用已知對特定系統有效的關鍵詞。

In general, prompt design is a more general concept while prompt engineering is a more specialized concept.

一般來說，及時設計是一個更為寬泛的概念，而及時工程則是一個更為專業的概念。

Prompt design is essential while prompt engineering is only necessary for systems that require a high degree of accuracy or performance.

及時的設計是必不可少的，而及時的工程設計只適用於對精確度或性能要求較高的系統。

There are three kinds of large language models.

大型語言模型有三種。

Generic language models, instruction-tuned, and dialogue-tuned.

通用語言模型、教學調整和對話調整。

Each needs prompting in a different way.

每個人都需要不同的提示方式。

Let's start with generic language models.

讓我們從通用語言模型開始。

Generic language models predict the next word based on the language in the training data.

通用語言模型根據訓練數據中的語言預測下一個單詞。

Here is a generic language model.

下面是一個通用語言模型。

In this example, the cat sat on.

在這個例子中，貓坐在上面。

The next word should be the, and you can see that the is most likely the next word.

下一個單詞應該是the，可以看出the很可能是下一個單詞。

Think of this model type as an autocomplete in search.

將這種模型類型視為搜索中的自動完成。

Next, we have instruction-tuned models.

其次是指令調整模型。

This type of model is trained to predict a response to the instructions given in the input.

這類模型經過訓練，可預測對輸入指令的響應。

For example, summarize a text of x.

例如，總結 x 的文本。

Generate a poem in the style of x.

以 x 的風格創作一首詩。

Give me a list of keywords based on semantic similarity for x.

給我一份基於 x 語義相似性的關鍵詞列表。

In this example, classify text into neutral, negative, or positive.

在本例中，請將文本分為中性、負面或正面。

And finally, we have dialogue-tuned models.

最後，我們還有經過對話調整的模型。

This model is trained to have a dialogue by the next response.

該模型經過訓練，可以在下一個迴應之前進行對話。

Dialogue-tuned models are a special case of instruction-tuned where requests are typically framed as questions to a chatbot.

對話調整模型是指令調整模型的一種特例，在這種模型中，請求通常是以向哈拉機器人提問的形式提出的。

Dialogue-tuning is expected to be in the context of a longer back-and-forth conversation and typically works better with natural question-like phrasings.

對話調整需要在較長時間的前後對話中進行，通常使用自然的問句效果會更好。

Chain of thought reasoning is the observation that models are better at getting the right answer when they first output text that explains the reason for the answer.

思維鏈推理是指，當模型首先輸出解釋答案原因的文本時，它們能更好地得到正確答案。

Let's look at the question.

讓我們來看看這個問題。

Roger has five tennis balls.

羅傑有五個網球。

He buys two more cans of tennis balls.

他又買了兩罐網球。

Each can has three tennis balls.

每個罐子裡有三個網球。

How many tennis balls does he have now?

他現在有多少個網球？

This question is posed initially with no response.

最初提出這個問題時，沒有得到任何答覆。

The model is less likely to get the correct answer directly.

模型直接得到正確答案的可能性較小。

However, by the time the second question is asked, the output is more likely to end with the correct answer.

然而，當提出第二個問題時，輸出更有可能以正確答案結束。

But there is a catch.

但有一個問題。

There's always a that can do everything has practical limitations.

總有一款產品能做到的一切都有其實際侷限性。

But task-specific tuning can make NLMs more reliable.

但是，針對具體任務的調整可以使無噪聲監聽器更加可靠。

Vertex AI provides task-specific foundation models.

Vertex AI 提供針對特定任務的基礎模型。

Let's get into how you can tune with some real-world examples.

讓我們通過一些實際例子來了解如何進行調整。

Let's say you have a use case where you need to gather how your customers are feeling about your product or service.

假設您有一個使用案例，需要收集客戶對產品或服務的感受。

You can use a sentiment analysis task model.

您可以使用情感分析任務模型。

Same for vision tasks.

視覺任務也是如此。

If you need to perform occupancy analytics, there is a task-specific model for your use case.

如果您需要執行佔用分析，則可根據您的使用案例選擇特定的任務模型。

Tuning a model enables you to customize the model response based on examples of the tasks that you want the model to perform.

調整模型可讓您根據希望模型執行的任務示例自定義模型響應。

It is essentially the process of adapting a model to a new domain or a set of custom use cases by training the model on new data.

其本質是通過在新數據上訓練模型，使模型適應新領域或自定義用例集的過程。

For example, we may collect training data and tune the model specifically for the legal or you can also further tune the model by fine-tuning, where you bring your own data set and retrain the model by tuning every weight in the LLM.

例如，我們可以收集訓練數據，並專門針對法律問題調整模型；您也可以通過微調進一步調整模型，即您可以自帶數據集，通過調整 LLM 中的每個權重來重新訓練模型。

This requires a big training job and hosting your own fine-tuned model.

這需要進行大量的培訓工作，並主持自己的微調模型。

Here's an example of a medical foundation model trained on healthcare data.

下面是一個利用醫療保健數據訓練醫學基礎模型的例子。

The tasks include question answering, image analysis, finding similar patients, etc.

任務包括問題解答、影像分析、尋找類似病人等。

Fine-tuning is expensive and not realistic in many cases.

微調成本高昂，而且在很多情況下並不現實。

So are there more efficient methods of tuning?

那麼，有沒有更有效的調整方法呢？

Yes.

是的。

Parameter-efficient tuning methods, PETM, are methods for tuning a large language model on your own custom data without duplicating the model.

參數高效調優方法（PETM）是在不復制模型的情況下，根據自己的自定義數據對大型語言模型進行調優的方法。

The base model itself is not altered.

基本型號本身沒有改變。

Instead, a small number of add-on layers are tuned, which can be swapped in and out at inference time.

取而代之的是對少量附加層進行調整，這些附加層可以在推理時調入或調出。

I'm going to tell you about three other ways Google Cloud can help you get more out of your LLMs.

下面我將向您介紹 Google Cloud 可以幫助您從法律碩士課程中獲得更多收穫的其他三種方式。

The first is Generative AI Studio.

第一個是 Generative AI Studio。

Generative AI Studio lets you quickly explore and customize generative AI models that you can leverage in your applications on Google Cloud.

Generative AI Studio 可讓您快速探索和定製生成式人工智能模型，並在 Google Cloud 上的應用程序中加以利用。

Generative AI Studio helps developers create and deploy generative AI models by providing a variety of tools and resources that make it easy to get started.

生成式人工智能工作室通過提供各種工具和資源，幫助開發人員創建和部署生成式人工智能模型，讓他們輕鬆上手。

For example, there is a library of pre-trained models, a tool for fine-tuning models, a tool for deploying models to production, and a community forum for developers to share ideas and collaborate.

例如，有一個預訓練模型庫、一個用於微調模型的工具、一個用於將模型部署到生產中的工具，以及一個供開發人員分享想法和開展合作的社區論壇。

Next, we have Vertex AI, which is particularly helpful for those of you who don't have much coding experience.

接下來是頂點人工智能（Vertex AI），它對那些沒有太多編碼經驗的人特別有幫助。

You can build generative AI search and conversations for customers and employees with Vertex AI Search and Conversation, formerly GenAI App Builder.

您可以使用 Vertex AI Search and Conversation（前身為 GenAI App Builder）為客戶和員工創建生成式人工智能搜索和對話。

Build with little or no coding and no prior machine learning experience.

只需很少或無需編碼，也無需任何機器學習經驗即可構建。

Vertex AI can help you create your own chat bots, digital assistants, custom search engines, knowledge bases, training applications, and more.