大语言模型微调之道2——为什么要微调

宝玉的技术分享
24 Aug 202314:36

Summary

TLDR本课程介绍了为什么应该对大型语言模型(LLMs)进行微调,微调是什么,以及如何通过实验比较微调模型与非微调模型。微调是将通用模型专业化,使其更适合特定用途,如聊天或代码自动完成。与提示工程相比,微调可以处理更多数据,纠正错误信息,并减少幻觉现象。此外,微调有助于提高模型在特定领域的性能和一致性,增强隐私保护,降低成本,并提供更好的控制。课程还介绍了用于微调的不同技术,包括PyTorch、Hugging Face和Laminai库。

Takeaways

  • 📚 微调(Fine-tuning)是将通用模型(如GPT-3)转变为特定用途的模型,如聊天机器人或代码自动完成工具。
  • 👨‍⚕️ 微调模型类似于从全科医生(通用模型)到心脏病专家(特定用途模型)的转变,提供更深入的专业知识。
  • 🧠 微调使模型能够从大量数据中学习,而不仅仅是访问数据,从而提升其性能和专业性。
  • 💡 微调有助于模型提供更一致的输出和行为,减少模型的幻觉(hallucinations)问题。
  • 🚀 与提示工程(Prompt Engineering)相比,微调可以处理几乎无限量的数据,允许纠正模型之前的错误信息。
  • 💼 微调适用于企业级或特定领域的用例,适合生产环境使用。
  • 🔒 微调可以在私有环境中进行,有助于防止数据泄露和保护隐私。
  • 💰 微调可以提高成本透明度,对于大量使用的模型,可以降低每请求的成本。
  • ⏱️ 微调可以减少模型响应的延迟,对于需要快速响应的应用(如自动完成)尤其重要。
  • 🛡️ 微调允许为模型设置更多的安全防护,如自定义响应和内容过滤。
  • 📚 微调过程中可以使用多种技术,包括PyTorch、Hugging Face和Laminar等库。
  • 📈 微调模型在实际应用中的表现明显优于未微调模型,能够提供更准确和有用的信息。

Q & A

  • 为什么要进行模型微调?

    -模型微调是为了让通用模型(如GPT-3)专门化,适应特定的使用场景,比如将GPT-4转变为专门自动完成代码的GitHub Copilot。微调可以使模型从更多数据中学习,提升其在特定领域的专业性和一致性,减少错误信息的产生,并能更好地适应用户的需求。

  • 微调和提示工程(prompt engineering)有什么区别?

    -提示工程是通过精心设计的问题来引导模型产生期望的输出,不需要额外数据,适合快速开始和原型开发。而微调则是通过大量特定数据训练模型,使其在特定任务上表现得更好,但需要更多的数据和计算资源。

  • 微调模型有哪些优势?

    -微调模型可以提高性能,减少生成错误信息的情况,使模型在特定领域有更深入的专业知识,输出更一致,并且可以更好地进行内容审查。此外,微调可以在私有环境中进行,有助于保护数据隐私和防止数据泄露。

  • 微调模型有哪些潜在的缺点?

    -微调模型需要大量的高质量数据,存在前置计算成本,并且可能需要一定的技术知识来正确地准备和使用数据。与简单的提示工程相比,微调的门槛更高。

  • 微调模型适合哪些使用场景?

    -微调模型适合企业级或特定领域的使用场景,特别是当模型需要频繁使用或者需要处理大量请求时。微调可以使模型更加专业和稳定,适合生产环境。

  • 在微调过程中,隐私如何得到保护?

    -微调可以在私有云(VPC)或本地进行,这样可以防止数据泄露和第三方解决方案可能带来的数据安全问题。

  • 微调模型如何帮助降低成本?

    -通过微调一个较小的模型,可以降低每请求的成本,从而在大量使用模型时节省开支。此外,微调后的模型可以更好地控制成本,包括响应时间和吞吐量。

  • 在微调模型时,有哪些工具和库可以使用?

    -可以使用包括PyTorch、Hugging Face和Laminai(Llama)库在内的多种工具和库。PyTorch是最低级别的接口,Hugging Face提供了更高级别的接口,而Laminai则提供了一个非常高级的接口,可以用很少的代码训练模型。

  • 微调模型时,如何处理模型的自动补全问题?

    -在微调模型时,可以通过添加指令标签来告知模型具体的指令和边界,从而避免模型继续自动补全无关的内容。

  • 在比较微调模型和非微调模型时,有哪些明显的差异?

    -微调模型在处理特定任务时,如训练狗坐下的指令,能够提供更详细、更准确的指导。相比之下,非微调模型可能无法理解或正确响应特定的指令。

  • 微调模型在对话中的表现如何?

    -微调模型在对话中能够更好地理解上下文和问题,提供连贯和相关的回答。而非微调模型可能无法进行有效的对话,其回答可能不相关或者缺乏连贯性。

Outlines

00:00

📚 了解为什么需要微调AI模型

本段落介绍了微调AI模型的重要性和基本概念。微调是将通用模型(如GPT-3)转化为特定用途的模型(如聊天机器人或GitHub Copilot),以提高其在特定任务上的表现。通过微调,模型能够从更多数据中学习,从而变得更专业,类似于从全科医生到专科医生的转变。此外,微调有助于模型提供更一致的输出,减少错误信息,并能够根据新数据更新知识库。

05:01

🔍 微调与提示工程的比较

这一段讨论了微调和提示工程(prompt engineering)的区别和各自的优缺点。提示工程不需要额外数据,成本低,易于开始,适合快速原型制作和通用场景。而微调需要更多高质量数据,有前期计算成本,可能需要技术知识,但可以处理大量数据,纠正错误信息,适用于企业级或特定领域的使用场景。此外,微调有助于提高模型的性能、一致性和隐私性,同时降低成本和提高对模型的控制力。

10:04

🧠 微调模型的实际应用示例

本段落通过实际的编程示例展示了微调模型和非微调模型在实际应用中的差异。通过比较两个模型对同一问题的回答,明显可以看出微调模型在理解指令、提供相关回答和进行对话方面表现更佳。微调模型能够根据指令提供具体的训练狗狗坐下的步骤,而未微调的模型则无法给出有用的回答。此外,微调模型在处理Mars话题和Taylor Swift的搜索查询时也显示出更好的理解和回答能力。

Mindmap

Keywords

💡fine-tuning

Fine-tuning是一种机器学习技术,指的是对已经训练好的通用模型进行再训练,使其更好地适应特定的应用场景。在视频中,通过将通用模型GPT-3转化为专门用于聊天的ChatGPT或GitHub Copilot,展示了fine-tuning的实际应用。这使得模型能够处理特定领域的问题,比如皮肤科医生针对皮肤问题提供专业意见一样,fine-tuned模型能够针对特定输入给出更精确的输出。

💡prompt engineering

Prompt engineering是指对大型语言模型的输入进行精心设计,以引导模型产生期望的输出。在视频中,它被描述为一种不需要额外数据即可开始的技术,通过编辑查询来改变模型的输出结果。这种方法的优点在于无需额外成本和技术知识,但可能存在信息错误和不一致性的问题。

💡general-purpose models

通用模型(General-Purpose Models)是指那些未经特定领域训练的机器学习模型,它们可以处理广泛的任务,但可能在特定任务上的表现不如专门针对该任务训练过的模型。在视频中,将通用模型比喻为全科医生,它们可以处理一般性的问题,但在特定领域(如心脏问题或皮肤问题)上可能不够专业。

💡specialized models

专门模型(Specialized Models)是指那些经过特定领域数据训练的机器学习模型,它们在某一特定任务上的表现优于通用模型。在视频中,专门模型被比喻为专科医生,如心脏病专家或皮肤科医生,他们能够更深入地处理特定的健康问题。

💡data

在机器学习中,数据是用于训练模型的输入信息集合。在视频中,数据被用来区分fine-tuning和prompt engineering两种方法:fine-tuning允许模型从大量数据中学习,而prompt engineering则不需要额外数据即可开始。数据的质量和数量对模型的性能有重要影响。

💡hallucinations

在机器学习模型中,幻觉(Hallucinations)是指模型生成的不真实或不准确的信息。这通常发生在模型对输入数据的理解不足或训练数据中存在错误时。在视频中,通过fine-tuning可以减少模型的幻觉现象,提高其输出的准确性和一致性。

💡retrieval augmented generation (RAG)

检索增强生成(Retrieval Augmented Generation, RAG)是一种结合了数据检索和文本生成的技术,它允许模型在生成回答时参考特定的数据集。在视频中,RAG被提及作为一种方法,可以在prompt engineering中选择性地使用数据,以改善模型的输出。

💡cost

成本(Cost)在视频中指的是使用机器学习模型时所涉及的经济开销。对于fine-tuning,需要更多的数据和计算资源,因此会有更高的前期计算成本。而对于prompt engineering,由于不需要额外的数据,所以前期成本较低。

💡privacy

隐私(Privacy)是指保护个人或组织数据不被未经授权的访问和使用。在视频中,fine-tuning自己的模型可以在本地或私有云中进行,从而避免数据泄露和第三方解决方案可能带来的数据安全风险。

💡moderation

内容审核(Moderation)是指对生成的内容进行管理和控制,以确保其符合特定的标准或规定。在视频中,通过fine-tuning,可以使模型更好地进行内容审核,比如对不当内容做出响应或拒绝回答某些问题。

💡libraries

库(Libraries)是指为软件开发提供预先编写好的代码集合,以简化和加速开发过程。在视频中,提到了三个用于fine-tuning的Python库:PyTorch、Hugging Face和Laminar(Llama)。这些库提供了不同层次的接口,使开发者能够更容易地训练和使用模型。

Highlights

本课程将学习为什么应该对大型语言模型(LLMs)进行微调,以及微调究竟是什么。

微调是将通用模型如GPT-3专业化,以适应特定的使用案例,例如聊天机器人或GitHub Copilot自动完成代码。

微调的模型可以处理比提示更多的数据,从而从这些数据中学习并提升自身能力。

微调有助于模型提供更一致的输出和行为,减少模型的幻觉问题。

与提示工程相比,微调可以适应几乎无限量的数据,允许纠正模型之前学习的错误信息。

微调后,模型在特定领域内的性能得到提升,可以更加专业和一致。

微调可以在私有环境中进行,有助于防止数据泄露和数据泄露。

微调可以降低成本,提高对成本的控制,包括响应时间和成本。

微调有助于提高模型的监管能力,为模型提供定制化响应。

使用微调的模型可以在对话中更好地进行轮次转换,提供更有用的回复。

微调模型能够针对特定指令给出清晰、具体的指导,而不是简单的重复或无关信息。

微调模型在处理查询时能够提供更准确的信息,如关于泰勒·斯威夫特最好的朋友的问题。

微调模型在模拟对话时能够更好地理解上下文并给出相关回复。

微调模型可以通过特定的指令标签来控制模型的输出,避免不必要的自动完成。

在比较微调模型和非微调模型时,微调模型在执行特定任务时表现更佳。

微调模型能够根据用户的输入提供更加详细和有用的步骤指导。

微调模型在处理复杂查询时能够提供更加精准和相关的信息。

微调模型在对话中能够更好地捕捉用户意图,提供更加个性化的回复。

Transcripts

play00:01

in this lesson you'll get to learn why

play00:03

you should fine tune what fine tuning

play00:05

really even is compare it to prompt

play00:08

engineering and go through a lab where

play00:10

you get to compare a fine-tuned model to

play00:13

a non-fine-tuned model

play00:14

cool let's get started

play00:17

all right so why should you fine-tune

play00:20

llms

play00:21

well before we jump into why let's talk

play00:24

about what fine-tuning really is so what

play00:27

fine tuning is is taking these general

play00:30

purpose models like gpd3 and

play00:33

specializing them into something like

play00:34

chat GPT the specific chat use case to

play00:37

make it chat well or using gpt4 and

play00:40

turning that into a specialized GitHub

play00:43

co-pilot use case to autocomplete code

play00:45

an analogy I like to make is a PCP a

play00:49

primary care physician is like your

play00:51

general purpose model you go to your PCP

play00:54

every year for a general checkup but a

play00:57

fine tune or specialized model is like a

play01:00

cardiologist or dermatologist a doctor

play01:02

that has a specific specialty and can

play01:04

actually take care of your heart

play01:06

problems or skin problems in much more

play01:08

depth so what fine tuning actually does

play01:11

for your model is that it makes it

play01:13

possible for you to give it a lot more

play01:15

data than what fits into the prompt so

play01:18

that your model can learn from that data

play01:20

rather than just get access to it and

play01:22

from that learning process it's able to

play01:23

upgrade itself from that PCP into

play01:27

something more specialized like a

play01:29

dermatologist so you can see in this

play01:30

figure you might have some symptoms that

play01:32

you input into the model like skin

play01:34

irritation redness itching and the base

play01:37

model which is the general purpose model

play01:39

might just say this is probably acne a

play01:43

model that is fine-tuned on Dermatology

play01:45

data however by taking the same symptoms

play01:48

and be able to give you a much clearer

play01:50

more specific diagnosis in addition to

play01:53

learning new information

play01:55

fine-tuning can also help steer the

play01:57

model to more consistent outputs or more

play02:00

consistent behavior for example you can

play02:01

see the base model here when you ask it

play02:04

what's your first name it might respond

play02:06

with what's your last name because it's

play02:08

seen so much survey data out there of

play02:10

different questions so it doesn't even

play02:13

know that it's supposed to answer that

play02:14

question but a fine-tuned model by

play02:16

contrast when you ask it what's your

play02:18

first name would be able to respond

play02:20

clearly my first name is Sharon this bot

play02:23

was probably trained on Me In addition

play02:26

to steering the model to more consistent

play02:28

outputs or behavior fine tuning can help

play02:31

the model reduce hallucinations which is

play02:34

a common problem where the model makes

play02:35

stuff up maybe it will say my first name

play02:38

is Bob when this was trained on my data

play02:42

and my name is definitely not Bob

play02:43

overall fine tuning enables you to

play02:46

customize the model to a specific use

play02:48

case in the fine-tuning process which

play02:50

will go into far more detail later is

play02:52

very similar to the model's earlier

play02:54

training recipe

play02:56

so now to compare it with something that

play02:58

you're probably a little bit more

play02:59

familiar with which is prompt

play03:01

engineering this is something that

play03:03

you've already been doing for a while

play03:05

with large language models but maybe

play03:07

even for over the past decade with

play03:09

Google which is just putting a query in

play03:11

editing the query to change the results

play03:13

that you see so there are a lot of Pros

play03:15

to prompting one is that you really

play03:17

don't need any data to get started you

play03:18

can just start chatting with the model

play03:20

there's a smaller upfront cost so you

play03:23

don't really need to think about cost

play03:24

since every single time you ping the

play03:27

model it's not that expensive

play03:29

and you don't really need technical

play03:31

knowledge to get started you just need

play03:33

to know how to send a text message

play03:35

what's cool is that there are now

play03:37

methods you can use such as retrieval

play03:40

augmented generation or rag to connect

play03:42

more of your data to it to selectively

play03:44

choose what kind of data goes into the

play03:45

prompt

play03:47

now of course if you have more than a

play03:49

little bit of data then it might not fit

play03:51

into the prompt so you can't use that

play03:53

much data oftentimes when you do try to

play03:56

fit in a ton of data unfortunately it

play03:59

will forget a lot of that data there are

play04:01

issues with hallucination which is when

play04:03

the model does make stuff up and it's

play04:05

hard to correct that incorrect

play04:06

information that it's already learned so

play04:08

while using retrieval augmented

play04:10

generation can be great to connect your

play04:12

data it will also often miss the right

play04:14

data get the incorrect data and cause

play04:17

the model to Output the wrong thing fine

play04:20

tuning is kind of the opposite of

play04:23

prompting so you can actually fit in

play04:24

almost an unlimited amount of data which

play04:26

is nice because the model gets to learn

play04:29

new information on that data as a result

play04:33

you can correct that incorrect

play04:34

information that it may have learned

play04:36

before or even put in recent information

play04:38

that it hadn't learned about previously

play04:41

there's less cost afterwards if you do

play04:43

fine tune a smaller model and this is

play04:46

particularly relevant if you expect to

play04:48

hit the model A lot of times so have a

play04:50

lot of either throughput or you expect

play04:53

it to just handle a larger load

play04:57

and also retrieval augmented generation

play04:59

can be used here too I think sometimes

play05:01

people think it's a separate thing but

play05:02

actually you can use it for both cases

play05:04

so you can actually connect it with far

play05:06

more data as well even after it's

play05:08

learned all this information

play05:10

there are cons however you need more

play05:12

data and that data has to be higher

play05:14

quality to get started there is an

play05:17

upfront compute cost as well so it's not

play05:20

free necessarily it's not just a couple

play05:22

dollars just to get started of course

play05:24

there are now free tools out there to

play05:26

get started but there is compute

play05:28

involved in making this happen far more

play05:30

than just prompting

play05:32

and oftentimes you need some technical

play05:34

knowledge to get the data in the right

play05:36

place

play05:37

um and that that's especially you know

play05:39

surrounding this data piece and you know

play05:41

there are more and more tools now that's

play05:43

making this far easier but you still

play05:45

need some understanding of that data and

play05:48

uh you don't you don't have to be just

play05:50

anyone who can send a text message

play05:52

necessarily

play05:53

so finally what that means is for

play05:55

prompting you know that's great for

play05:57

generic use cases it's great for

play06:00

different side projects and prototypes

play06:01

it's great to just get started really

play06:03

really fast meanwhile fine tuning is

play06:06

great for more Enterprise or domain

play06:07

specific use cases and for production

play06:09

usage and we'll also talk about how it's

play06:12

useful for privacy in this next section

play06:14

which is the benefits of fine-tuning

play06:16

your own llm so if you have your own llm

play06:20

that you fine-tuned one benefit you get

play06:23

is around performance so this can stop

play06:26

the llm from making stuff up especially

play06:29

around your domain it can have far more

play06:31

expertise in that domain it can be far

play06:33

more consistent so sometimes these

play06:35

models will just produce you know

play06:37

something really great today but then

play06:39

tomorrow you hit it and it isn't

play06:40

consistent anymore it's not giving you

play06:42

that great output anymore and so this is

play06:44

one way to actually make it far more

play06:46

consistent and reliable and and you can

play06:48

also have it be better at moderating if

play06:51

you've played a lot with Chachi BT you

play06:53

might have seen charging Beauty to say

play06:54

I'm sorry I I can't respond to that and

play06:57

you can actually get it to say the same

play06:59

thing or something different that's

play07:01

related to your company or use case to

play07:04

help the person chatting with it stay on

play07:05

track and again so now I want to touch

play07:07

on privacy when you fine-tune your own

play07:10

llm this can happen in your VPC or on

play07:13

premise this prevents data leakage and

play07:16

data breaches that might happen on off

play07:18

the shelf third-party Solutions and so

play07:21

this is one way to keep that data safe

play07:23

that you've been collecting for a while

play07:25

that might be the last few days it might

play07:28

be the last couple decades as well

play07:30

another reason you might want to

play07:31

fine-tune your own llm is around cost so

play07:35

one is just cost transparency maybe you

play07:38

have a lot of people using your model

play07:41

and you actually want to lower the cost

play07:43

per request then fine-tuning a smaller

play07:46

llm can actually help you do that and

play07:48

overall you have greater control over

play07:50

cost and a couple other factors as well

play07:52

that includes uptime and also latency

play07:55

you can greatly reduce the latency for

play07:57

certain applications like autocomplete

play07:59

you might need latency that is sub 200

play08:02

milliseconds so that it is not

play08:04

perceivable by the person doing

play08:06

autocomplete you probably don't want

play08:07

autocomplete to happen across 30 seconds

play08:09

which is currently the case with running

play08:11

gpd4 sometimes and finally in moderation

play08:14

we talked about that a little bit here

play08:16

already but basically if you want the

play08:18

model to say I'm sorry to certain things

play08:20

or to say I don't know to certain things

play08:23

or even to have a custom response this

play08:26

is one way to actually provide those

play08:28

guard rails to the model and what's

play08:30

really cool is you actually get to see

play08:33

an example of that in the notebooks all

play08:36

right so across all of these different

play08:37

Labs you'll be using a lot of different

play08:40

Technologies to fine-tune so there are

play08:43

three python libraries one is pi torch

play08:45

developed by meta this is the lowest

play08:48

level interface that you'll see and then

play08:50

there's a great Library by hugging face

play08:52

on top of Pi torch and a lot of the

play08:54

great work that's been done and it's a

play08:57

much higher level you can import data

play08:59

sets and train models very easily and

play09:02

then finally you'll see the laminai

play09:04

library which I've been developing with

play09:06

my team and we call it the Llama library

play09:09

for all the great llamas out there and

play09:11

this is an even higher level interface

play09:13

where you can train models with just

play09:16

three lines of code all right so let's

play09:18

hop over to the notebooks and see some

play09:19

fine-tuned models in action

play09:21

okay so we're going to compare a

play09:24

fine-tuned model with a non-fine-tuned

play09:27

model so first we're importing from the

play09:29

Llama Library again this is from lamini

play09:31

the basic model Runner and all this

play09:34

class does is it helps us run open

play09:37

source models so these are hosted open

play09:38

source models on gpus to run run them

play09:41

really efficiently and the first model

play09:43

you can run here is the Llama 2 model

play09:46

which is very popular right now and this

play09:49

one is not fine-tuned so we're going to

play09:51

just instantiate it based on this is its

play09:53

hugging face name and we're gonna say

play09:56

tell me how to train my dog to sit so

play09:59

it's just you know really really simple

play10:01

here into the non-fine-tuned model

play10:04

we're going to get the output out

play10:06

and let's print

play10:10

non-tuned

play10:11

output and see

play10:14

okay so we asked it tell me how to train

play10:17

my dog to sit it said period and then

play10:19

tell me how to train my dog to say tell

play10:21

me how to teach by dog to Cub and tell

play10:23

me how to get my dog to heal so clearly

play10:25

this is very similar to the what's your

play10:28

first name what's your last name answer

play10:29

this model has not been told or trained

play10:33

to actually respond to that command

play10:37

so maybe a bit of a disaster but let's

play10:40

keep looking so maybe we can ask it

play10:42

um what do you think of Mars

play10:47

so now you know at least it's responding

play10:50

to the question but it's not great

play10:51

responses I think it's a great Planet I

play10:53

think it's a good Planet I think it'll

play10:54

be a great Planet so it keeps going

play10:57

um very philosophical potentially even

play11:00

existential if you keep reading

play11:03

all right what about something like a

play11:05

Google search query like Taylor Swift's

play11:07

best friend let's see what that actually

play11:09

says

play11:11

all right well uh it doesn't quite get

play11:14

Taylor Swift's best friend but

play11:16

um it did say that it's a huge Taylor

play11:19

Swift fan

play11:21

um

play11:22

all right let's keep exploring maybe

play11:24

something that's a conversation to see

play11:26

if it can do turns in a conversation

play11:27

like Chachi PT so this is uh agent for

play11:30

an Amazon delivery order

play11:33

okay so uh at least it's doing the

play11:36

different customer agent turns here uh

play11:38

but it isn't quite getting getting

play11:40

anything out of it this is not something

play11:42

usable for any kind of like fake turns

play11:44

or or help with making an auto agent all

play11:47

right so you've seen enough of that

play11:49

let's actually compare this to llama 2

play11:51

that has been fine-tuned to actually

play11:53

chat

play11:54

so I'm going to instantiate the fine

play11:56

tune model notice that this name all

play11:58

that's different is this chat here and

play12:01

then I'm going to let this fine tune

play12:02

model

play12:04

do the same thing so tell me how to

play12:06

train my dog to sit I'm going to print

play12:08

that

play12:11

okay very interesting so you can

play12:14

immediately tell a difference so tell me

play12:15

how to train my dog to say it's still

play12:17

trying to auto complete that so tell me

play12:19

how to train my dog to sit on command

play12:20

but then it actually goes through almost

play12:23

a step-by-step guide of what to do to

play12:26

train my dog to sit

play12:28

cool so that's much much better

play12:31

and the way to actually quote unquote

play12:33

get rid of this extra autocomplete thing

play12:37

is actually to inform the model that you

play12:40

want instructions so I'm actually

play12:42

putting these instruction tags here this

play12:43

was used for llama too you can use

play12:45

something different when you fine-tune

play12:46

your own model but this helps with

play12:48

telling the model hey these are my

play12:50

instructions and these are the

play12:51

boundaries I'm done with giving this

play12:53

instruction stop stop continuing to give

play12:55

me an instruction so here you can see

play12:57

that it doesn't autocomplete that on

play12:59

command thing and just to compare just

play13:01

to be fair we can use the see what the

play13:03

non-fine-toed model actually says

play13:05

um great it just repeats the same uh

play13:08

same thing or something very similar so

play13:11

um not quite right cool let's keep going

play13:13

down so what do you think of Mars this

play13:15

model

play13:17

oh it's a fascinating planet let's

play13:19

capture the imagination of humans for

play13:21

centuries Okay cool so something that's

play13:23

much better out uh output here what

play13:25

about Taylor Swift's best friend let's

play13:27

see how this does

play13:29

okay this one's pretty cute it has a few

play13:32

candidates for who uh Taylor Swift's

play13:35

best friend actually is

play13:38

let's take a look at these turns from

play13:40

the Amazon Delivery Agent okay it says I

play13:44

see can you provide me with your order

play13:46

number this is much much better it's

play13:48

interesting because I down here it also

play13:50

summarizes what's going on

play13:52

which you know may or may not be

play13:53

something that you would want and that

play13:54

would be something you can fine tune

play13:55

away and now I'm curious Which hat GPD

play13:58

would say for tell me how to train my

play14:00

dog to sit

play14:02

okay so it gives you know different

play14:04

steps as well great all right feel free

play14:06

to use Chachi BT or any other model to

play14:09

to see what else they can each do uh and

play14:12

compare the results but it's pretty

play14:14

clear I think that the ones that have

play14:16

been fine-tuned including Chachi PT and

play14:19

this llama2 chat llm they're clearly

play14:22

better than the one that was not

play14:23

fine-tuned

play14:24

now in the next lesson we're going to

play14:26

see where fine tuning fits in in the

play14:29

whole training process so you'll get to

play14:31

see the first step in in how to even get

play14:33

here with this fine-tuned model

Rate This

5.0 / 5 (0 votes)

Related Tags
AI微调模型性能一致性隐私保护数据泄露成本效益个性化企业应用开发工具编程辅助
Do you need a summary in English?