Transcripts
1. Introduction to the finetuning course: We all know that touch of
beauty is an impressive, powerful, artificial
intelligent chatbot. But have you ever wondered if
you can use juncture APTs, impressive abilities to build your own project
that can help you in your Rick based on your data, such as building
a chocolate that helps with your customer
service department. If you ever did, then you
are in the right place. I am money, I'm homeless. And artificial
intelligence specialist. In this course, I will guide you to understand what fine-tuning is to do the whole process until you can create
your own types. The first module, we will
learn what fine-tuning is, why is it important? And what types of projects
do people use fine-tuning. In this second module, we will be working
on data preparation. How to prepare our data in a format that actually
can be given to a pre-trained model
and work on it to fine tune our own model
in our specific domain. In the third model,
we will configure the fine-tuning process
and we will reach to a level of which
we can test and evaluate the performance
of our model. By the end of this course, you will have gained
valuable skills about fine tuning and you will be ready to implement
your own project. So let's get started.
2. Understanding GPT Models and Fine-Tuning for Specific Domains: Welcome to the first
lesson in this course. In our lesson today we will
speak about GPT models. What are they and what
are their limitations? And why would we think of
fine tuning our own models, which will leave us with an
understanding of fine tuning. What does it mean and where
does it actually kept? Gpt-3 models is a large, powerful families
of language models. They all can deal
with text, prompts, questions, any
information according to what they were trained for. They apply the natural
language processing, which is a subfield of artificial intelligence
that is about computers. How do they deal, interpret, and generate human language? These pre-trained models
previously trained by open here, they learned how to deal
with human language, which makes them an excellent start for many applications. But we need to adapt
and to meet our own. Although there's very
powerful and impressive, but they have limitations. And those limitations
are the things that will make us think about
fine tuning our own. One of the limitations
is that they have an upper bound on quality and how a person
needs to do lots of prompt engineering in order
to receive a quick response. Another limitation
is that there is a maximum size for the front. So you have to work on the
prompt and a good way. And you can't provide
the model with much examples in order to
receive a good response. Other limitations
that we can speak about is the coast and literacy. We know that GPT charges
for tokenization. So it's not a smart
choice to think about depending on them forever
and keeping for it. So this takes us to start
thinking of fine tuning. And why would we fine
tune our own models and train them to work and
our specific domains. But what is fine tuning? Fine tuning is a technique
to use pre-trained models, adapt them to do
our order tasks. But first, we need to depend on the auto by removing
the final layer, feeding the motor with our
own prepared datasets, which will actually
create a model that knows exactly what
outputs should it give.
3. Mastering Fine-tuning for NLP Applications: Fine-tuning is a powerful
technique for adapting pre-trained models to
specific NLP applications. In this lesson, we'll
cover the key concepts, the steps you need to follow
in order to be successfully ready to find tune them before deciding to
fine tune your model. And you need to understand the
problem you're working on. What is the task in hand and the requirements
of this task, such as the speed, the accuracy. You also need to know
the characteristics of the dataset
that you're using, such as the size, the domain, and what type of
data are you using? By answering these questions, you are totally ready
to determine if fine-tuning is the solution
to your problem or not. Once you have decided
to find tune a model, you need to choose
the pre-trained model you want to depend on. There are many types
of pre-trained models, each has strengths
and reconsider. Some of them were trained
on data or text generation. Others great for coding. So you have to choose the
model that was trained on data that is similar
to the data you're using. Or in this specific domain. This will make fine
tuning or effective. Oxford you choose the
pre-trained model you have to start breaking
on preparing your data. Preparing the data includes
changing the format of the data to a format that the pre-trained model
can understand, such as tokenization
and encoding. If you have your data ready, this is the key for you to
start doing the fine tuning. And that means to
continue training the pre-trained model on the
dataset that you prepared. Once you have fine
tune your model, you have to evaluate the
performance of the model. You have to try the model, give it input, see what
outputs does it give you. This will help you to do some error analysis to know where your models struggling and how to improve its performance. So after understanding your
problem, preparing your data, choosing the pre-trained model, tuning the model
and evaluating it. You will be in your
way to mastering the fine tuning for
NLP applications.
4. Introduction to OpenAI Playground: Hello and welcome
to a new lesson. In our lesson today
we are going to be previewing Open
AI's. Click it on. Playground is an online, easy to interface that
allows users to experiment and test APIs capabilities of the chat models they created. The thing is that whenever we
want to use the data down, you don't have to have previous
knowledge in programming. Any user or any person
who doesn't know anything about programming can
actually use the Play-Doh. Because the playground
is an interface that allows you to test the already pre-trained
models done by OpenAI. Or you can actually use it to test your own fine-tuned models. So in order to check what is
this playground, Let's go. We're gonna go to a website
called OpenAI playground. Once you type
opening the grounds, it actually is going
to take you to open a nice page where you can
generate your API keys. Here you can present
your API keys or you can actually create a new
secret if the icky, remember these API
keys are the ones that we use to fine
tune our own models, to build our own
chatbots and so on. And whenever you
generate a new one, you have to copy it and keep
a copy of it because it is a secret key that doesn't appear to you anymore over here. So this is the page where we can actually create an API
keys and take them. But in this page we want to go to the playground
at the top part. Once I click on the
playground at the top, this is the playground, the one that we have to
preview its features today. First of all, the Brigham
is a testing tool for models for so on the
right side over here, you get to choose from
this drop-down list, which model are you using? Here I have the different Open AI's models,
Da Vinci query, Babbage and Adele, and they have different
specifications. We will be talking
about them later. I can get which one
of them to test. Once you choose the model, you have this box over here. This is the input
and output books. It is the place
where you get to ask a question such as your prompt, such as wiped me cars names, and sending the prompt. And in the same place you get to receive the risk bonds
coming back from Copenhagen. So this is an input and
output box where your prompt appears on the response to that prompt appears
as the thing is, there are some other
settings that you can work on in order
to control this bond. E.g. let's talk about the
temperature parameter. The temperature is a
parameter that varies 0-1. If you set it to zero, means you don't want
the model to be creative by responding
back to you. If you put it to one, it means that you
want the model to be much creative and to give you various answers every time I try to give creative
answers as well. If you want to talk about the
maximum length parameter, we're actually talking
about the tokenization. And this is a number
that can actually be the specific amount that you want the model to
respond back to you. It's not just about
the response, It's also including the prompt. So let's change this
parameter to true. E.g. refresh. Send you a prompt again. And the answer you'll receive
is very short because this altogether is around
or about two tokens. But give it more,
let's say tan and you submit and you get
a longer answer because this answer over here
with the prompt is around ten tokens get
the stop sequence is a variable that allows us to
make our foreign tuned model understand when to stop
responding to any prompt. This is a character that
we will decide or we will use whenever we're testing
our fine tuned modals. So these are the settings I can change over
here and there are some other features as well on the playground and
I can break e.g. any prompt generated
can be saved. It's also sharable,
so this is good. You can share it with people. You can copy it, take a copy of it and keep
it wherever you want, e.g. on your documents and so on. Another thing that I
want to talk about is the bar above over here. You'll have different
tabs over here. And these are very
helpful because you can find lots of things
about documentation. And it opened his eyes
blog and other examples of prompts that you can
use to test GP T-Mobile's. If you go to the
documentation section, we will leave the playground, this prompt going to go away. But you will find lots
of descriptive folders, files about whatever
you may need more other tutorials on
the tokenization pricing about fine-tuning and any
other document that you may need that can help heal dealing with Open
AI's to be team. So that's it for the playground. This is our playground. Again, it is an easy interface
that allows users or developers to test the
capabilities of GPT models.
5. Maximizing Model Accuracy with Data Preparation and Formatting: Hello and welcome
to another lesson. In our lesson today
we are going to be speaking about the benefits of preparing data and formatting it before using it to
fine tune a model. So first of all, we
have to know that preparing data is crucial
for fine tuning. Why? Because it affects the accuracy and the efficiency of the model. So first, formatted data
is used in a good way. So this makes the model can actually generalize
patterns in a good way, which means our data has
to be labeled encoded unstructured before given
to the pre-trained model, as well as
well-formatted data is actually a good starting point because it reduces the
errors and it makes sure that the model does not
learn biased information. This well-formatted
dataset also reduces the time and effort for
processing and cleaning data. Data actually comes in
different texts or formats. Using different formats means that you have to ensure that this data is prepared
in a way where it is suitable to be
fed to the pre-trained. So the data comes in
different formats. And different formats
can be used for machine learning or natural
language processing projects. But first, according
to the data you have, you have to know how many steps coming ahead to format it. If you have a row text, you need to know that
for this row texts cannot be given to the model, but yet there are some
preparation tips that we can do to prepare it to be
given to a pre-trained model. Another type of
data is CSV files, which is known for spreadsheets. This also has to be prepared in a way where the model
can understand it. The last format that
I'm going to speak about is the tastes and plot. And this is the Final Four
lot that we will break and prepare on any
type of data in order to be ready
to give the data directly to be used for fine
tuning a pre-trained model. So in this lesson, we spoke about the
benefits of formatting and preparing data before
fine tuning a model, we also spoke about the
types of data formats that can be used to
fine tune a model. And in the upcoming
list of lessons, we'll be talking
about techniques and ideas of power to prepare
data in the best way.
6. Fine-tuning a GPT Model: Understanding Data Formats: Hello and welcome
to a new essence. In our lesson today
we're going to be speaking about the format of the data that we will use for fine tuning it TP, T mode. First of all, we will
be talking about preparing our different data. We will be previewing
different datasets earned. We will be discussing what strengths and
weaknesses do we have them? Then? We also are going to be
mentioning tips about how to have a good data that is ready for fine
tuning a model. And finally, there is something
that I want to explain by the end of this lesson is
adding suffixes to data. I'm gonna go to my Google Drive. I have here a set of or a group of datasets
that I prepared for you. I want to preview with our
look at them and see what are the strengths and
weaknesses in such dataset. E.g. I. Have this first file
is called Arduino. Arduino is a specific
component, a microcontroller. So let's imagine or
making a chart mode. This chatbot is to answer
any question about Arduino. So the thing is, I need to
have a look at my data. First of all, as you can see, I have a group of
questions over here. On the left. I have answers to
these questions. And these are
different questions. What is Arduino what type of license does the Arduino
hardware product? Hughes and soma, want
to show you that it's sometimes a student
with a different way, like what is the
Arduino projects? So what is Arduino? What
is the Arduino project? These two answers, they may have the same answer or a close
answer that is acceptable. But if you look here, when did the Arduino
Project began? The Arduino Project began 2005. This is a waste of time for the model to read
each of the data. So I have a duplicate here. So first of all, I should
actually remove this row. So you have to preview your data and you have to
make sure that you have as single question of which there is a related
answer, a single answer. Sometimes I may use
the same answer, but for sure I have to feed my modal with different
questions structures. So using two sentences
with the same meaning, but they're written in a
different way, is acceptable. But choosing the same or giving the same answer to two different
prompts is not accepted. Okay, Let's preview
another dataset. I'm gonna go to this one. This one is about earthquakes. It's also a group of
questions about earthquakes. This dataset can be used in e.g. if you are a teacher or an instructor and you're
building a tool to help your students in order to ask any question about this
unit or something. So I have various group of questions on each of which
has a different answer. And as you can see, it
is acceptable to have a very long answer related to
one question that is okay. But again, as you can see, I have here e.g. only 25 values and this is the next step I'm
going to speak about. We should not train or fine tune a model according to only
25 different inputs. And results are prompts
and completions. We should actually give
the model at least around 200 different prompts and completions in order
to make sure that the final stage or the fine tuning stages actually
happening in a good way. Good. So this is a good dataset yet it doesn't have
much information. Let's preview another one. I'm going to go with this one,
the mental disorder data. This is also an initiative
of questions and answers, and it is actually created in a smart way that I'm
gonna be sharing with you in the
upcoming classes. First of all, what is
the mental disorder? And this is the
answer over here. So this is a set of questions with a set of officers.
Okay. That is good. But if I go down,
I have empty rows and I shouldn't actually have
integrals in my dataset, so these integrals
should be deleted. Another thing is, what are the
causes of mental disorder? This is a question. And what causes mental disorder? As I said, these two questions are written in a different way, but they both give
the same answer. That's okay because
they're actually taking the same answer. Good, because I have different questions with
this e-mails or this can help the modal understand
and link between that. Sometimes the question can
be asked in various ways. Good. Other things I want
to pay attention to is that if you can see here, these are sets of questions
that are actually repeated. I have a set of
questions over here. And I may have this same questions but with
different ways of typing. These actually have to be read in a proper way
because these were prepared with an online tool that generates
questions and answers. So such data lakes go down. If I scroll down, I can
see that I have around 138 rows removing the empty row. So 150, I can say that
it says acceptable. So such dataset is
acceptable to be used for fine tuning and model. Now let's preview the final
one I've prepared for you, which is the one over here. This is different. Why? Because it's not in the shape
of questions and answers. It's actually built as prompts,
topics, different topics. And here on the right side
of tweets of the person, of a specific person. Now, such data can be used, e.g. to mimic someone's way
of responding to people. If you can pay attention
the importance of staying hydrated
during exercise. And on the right side I have
his comment or his tweet. Just finished a long
run and being awesome, Remember to stay hydrated
while waking up. Good tips of improving
your sleep habits. I also have here,
struggling to fall asleep, try reading a book
before, and so on. So these are different
tweets in different topics. And these such data
can be used to build a chat bot that
can actually speak like natural people
in a friendly way. You can see here it says, Guys, I highly recommended and he is giving some
exclamatory sentences. Sometimes he starts
my asking you a question and then giving
his opinion about it. So good. Now, as you guys saw in the previous
datasets, there were, there was something that I didn't really pay attention
to or talk about. Looking here at each sentence
in my dataset over here, I said we have a prompt. This is the prompt or the
question that I expect my user, the one who's going
to use my application to type into the model. And this is the answer that I expect my Moodle to
give back to them. As you can see here, at
the end of each prompt, we included a backslash
end up flushing, and this is a suffix. Suffixes are supposed
to be added to my data before giving
my data to the model. This will make the
pre-trained model understand that this is
where the sentence or the prompt and this is where completion or mine
respond to the sentence start. And also the completion should also include
a suffix as well. Two different suffixes. Suffixes can be a backslash n, It's a can also be a word such as writing the word end, e.g. in the capital
letter or anything, it has to be repeated
in each grant, as well as I have
to make sure that this specifics of eggs is not included in my dataset
anywhere around the data. But now, going back to
this, the tweets data, let's think about how do I add the suffixes and
prefixes in one-click. I can go with the following. I can take this, paste it on the second column. Getting back to
the first column, this is the first
prompt that I have. I can go with this equals to. So we're applying somehow
a function, this cell. And I want you to add to it
between double quotation because I'm adding some
strings, e.g. backslash. And I can close and send me
whatever capital or small n. It's according to what is the
suffix that we want to use, then you can send this. And as you can
see, it's actually suggesting to auto-fill
the rest of them. Do you want from me to also add this backslash into
every so over here, I can go with yes, no, by actually created as set of bronze with
the suffix that I need. So I'm going to
copy this law in. I'm going to go back
to the first one, but I'm going to
paste as values only. Then I can remove this
column from here. So now I have my
dataset prepared with the same suffix added to
each one of them, earn. I can also do this scene
to this dataset over here. That's it for today. So to wrap up, we just previewed a
group of datasets. We discussed how do I
use these datasets to different applications using
GPT models or fine tuning. We saw that we should remove
duplicates in curios, we should add suffixes. So this is what we said or we spoke about
in the previous lessons. We said that whenever I want
to use data to fine tune, that data has to be labeled
structured and encoded. And this is the final shape
of the data that I'm going to use for mine point tonight. But the thing is, before I
use it as an Excel file, I have to change its
former two adjacent pipe. And that is the thing
that we're going to explain on the upcoming lessons. So stay tuned and thank
you for watching.
7. Data Cleaning with Python: Removing Missing Values and Duplicates: Hello and welcome
to another lesson. In our lesson today we're
gonna be working on our data. We want to make sure we have clean data to be used
for fine tuning. The thing is today is one of the techniques that you can use Python code for
data preparation. So what we're working on
is we have a CSV file, which is a dataset, and we're going to
read this slide. Then we want to remove
any missing values or any duplicates
because in fine tuning, duplicates are actually bought. Then we want to save the new
dataset into a new file. So this is our plan
for today's lesson. So first I'm gonna go
to a new web browser, and I'm going to go to
something called Google Colab. Google Colab is an online
tool that allows you to run and execute your
Python code easily. So this is where we're
gonna be working. This is the main
page that bulk bins whenever you go to Google Colab, you have to make sure you're
logged into your own e-mail, the one that is
linked to your drive, because this is where
we're going to keep our detailed and use it for the drive earned
when we wanna do is press to fall and
start a new York, we have to make sure that
the data you want to record is on your drive. So I'm gonna go to my
drive, my Google Drive, and make sure that I
prepared my datasets. So going to the drive takes me to show you that I created a
file that is called dataset. Inside this line, I pasted
or I uploaded my data. And if I want to preview
the data that I have, you can see that I have a
dataset as an Excel sheet, a CSV file that has some
information about computers and what is their speed or
other information that I need, such as the prize and
wherever I can find them, the country, the city, and some other information. Maybe I want to build
a model that uses this information in
order to be able to protect a computer's
priced according to the internal memory and the
speed and the CPU will, and any other details that I need to know
about the computer. Good. So this is my dataset
that we're going to work. I put it in my drive, in a file, in a folder called data set inside a file that is
called hardware software. Now here I'm minimally new. Note, fill in Google Colab. Now the thing is first of all, you have to make sure that the Colab is connected to your. I'm gonna go to the
left over here, clicking on this icon over here, I can preview my files. You can see that it is
connecting to something, but it's not actually
connected to my droid. So clicking on the
third button over here, I can actually mount Molly
drive to the global color. So I'm going to request
mountain with Google Drive. Sometimes it may be autoscaled, full permissions and so on. Connect to Google Drive. If you didn't connect before, maybe Osman you for access. So you can actually
get access stone. It's supposed to
appear over here. And as you guys can
see here it is. We're ready to go.
Now we're going to start typing some Python code in order to be able to read
the file and recorded. First thing is that we need
to import the pandas package. So we're going to work
on import pandas as pd. So the thing is we're trying to import the pandas package, the one that is going to allow
us to deal with CSV files. On the band, I click trunk and I'm waiting for
partners to be imported. Good. Now I'm ready to use
Pandas whenever I want. The faint. And the next thing is to start by adding a
second code line. And now the thing is we want to decide the path of our file. The part of the CSP found
that we added to our dry. I'm gonna go click on
the drop over here. And then I'm gonna
go to my drive. And then I'm gonna go look
for the folder that I call dataset that I have
my file inside. There it is. And once I open it, we can see there is my file. I will click on it and
ask for copy path. This is the path of the file. So I'm gonna go here
and say path is equal to, I'm hoping condition. And paste your path over here
and make sure that you had pasted correctly since it
did give me this chick sign. So two means that my file
is ready to be written out. What do we want to? We're
going to add another code line now we're going to start
to try to read our four. So the thing is,
we're going to go with defining a
function is equal to PD dot read
underscore csv, fine. And this CSP file
can be found King. The path that I
already wrote above. I'm ready to belong and
to run like cooked. Good. No, I can actually read my data to make sure that
I can read my Anita, let's go with the following. Df, head. And I can use this function
to present or to show the first five rows of
my dataset or future. Let's run it again. And as you can see, I have access toy dataset
and I will present the first five values
or rows in it. So I can see that this is a
computer that is entailed and this is the speed and any other information
that I need to know. Nope. What boast won't. Thing is now we want to start doing the steps we agreed on. First of all, you
want to start by removing the missing values. So going back to our callout, we want to start a new code. Now, what do we want it to? We want to start by dropping
the missing values. So we're going to go
with the UDF dot, drop. Any heard. We're going to open
a bracket around. We can give a variable
that is cooled in place. True in place means
you want to replace the values in the
dataset you read and you upload it over here
by the new values. Yes, I do. So I'm gonna
go with the following. In-place equals true. Or will know I'm going
to run my cooked. Good. Now, this specific code removes any missing
values from my dataset. So if there were any
missing value that may affect the functioning
of my trained model, then I want to remove them so I can use this function
to remove them. Now the next step is to add another coat on.
What do we wanna do? We want to actually
remove the duplicates. As we said, training
a model with duplicate data actually affects the performance in the modal. Enlighten me. Good. So we're gonna go again with df dot drop on his
care duplicates. Here it says, you can
actually just collect on it. Earned. Again, we're gonna go with
in-place equals true. We have double bonds
state over here. And I'm going to run my code. Good. So the second
statement over here, just to remove any
duplicates from the dataset that we were using? No. Our next step is
to actually save our new dataset into a new file. So I'm going to open a new code. Now. Let's save this dataset
two hour drive. So I'm gonna go
with the following, the doped to on your scores, CSP, the function I created, I'm going to save it as a
CSV file that it opened. Rock kits are gonna go
with naming the new files. So I'm going to open
a single mutation. And in silencing,
Fisher will go with, I'm going to go with formatted
date underscore data set. And this is the name of the
no-fault that are creating. And it's going to
be as CSV file. And I need to also
add an index on make sure the index
variable is false. Hair and on my
statement is ready. I'm going to run it
and there we go. So what did this CO2, there's code actually
import the pandas package. The pandas package helps
me to deal with detail. I identified the part
of the file which I kept on my drive after
I mounted to My Drive. And then I started
by reading a file, presenting some of the
values and his bio, I dropped the unwanted
or missing values. I remove the duplicates
and I finally saved. My thoughts are
owned by the way, to show you this
nurse in unison call, I'm gonna go to my
Drive over here below, and I can find that
there is a flight over here called
formatted data set. So this is the new
dataset I can tell loaded than perfume on data and make sure
everything is one. So this was actually one of the ideas that people
may use in order to prepare their datasets before they use them
for fine-tuning. Thank you for watching.
8. Generating Questions and Answers from Text for Fine-tuning AI Models: Hello and welcome
to a new lesson. In our lesson today, we're gonna be speaking
about preparing questions and answers
from a fixed point. As we spoke before
in previous lessons, we said that sometimes
we may need to use text, just throw text in order
to fine tune a model. But we can't actually send this row text the way
you test to the model, we need to do so for plotting, the format is it can be
divided into two stages. Principle, preparing
the data as prompts and completions with the suffixes
included in the data. And the second one is to change this data two adjacent per lot. And then we can start doing
the fine-tuning process. So for today we're going
to speak about this step. How can I prepared questions
and answers, the text? What I'm going to actually share a good tool with
you that you can use. So as you can see here, I
have an essay about animals. This is around 500
words is saying. And as you can see, this essay is divided into different paragraphs or things. I can actually copy
analytics all of it and take it to an
online tool where I can use this text to generate questions and
answers to be able to train my model according
to those questions and answers as prompts
and completions. So the thing is
we're going to go to the width browser and
we're going to type drain Pali and QA generated. This is our tool, the one that I'm going to be
sharing with you for today. So this is, this is a tool
that was created by OpenAI. And even you can use this
tool without anything. We're just gonna be paying
for the tokenization, the one that is already touched, GPT or TBT is actually
charging for. So look here before we start, we need to enter our API key. Because we're actually using an ATI module to
actually do this. We're, let's reach here, woke up extremely
equals two nodes, which in a renter. And it can actually help you by signing up for logging in
if you have an account. But let's say I don't
have an account. I can go with
bringing my API key. And just by adding more API key here and
verifying the key, I can actually start
using the tool. So I have only the icky
over here prepared. I'm going to copy it. I'm going to go
back to my website. I'm going to paste it. I'm going to try
to verify my key. And it says that
thing Q UK has been verified and stored
locally in your browser. Good. I'm gonna go with, okay, and let's not say VB icky. Now I'm ready to
go. Look at this. This is so easy and smooth
and it's actually helpful. It says, enter the sticks out. You want to generate questions
and answers two over here. So I'm gonna go back to
the text I prepared. I'm going to copy my text. New ladder, how long it is. Once I have void, it's ready. I'm going to go
back and paste it. Over here. You have to give the modal or at the program an idea about how was my
paragraphs separated. So it says, what's the paragraph separated with a single return? Or let's say a blank
space or something. So I'm gonna say, I'm gonna keep it
as singular turn. Or maybe I can just go
with the following, giving a blank space
at the beginning or the end of each paragraph. And this will actually make wine data more clear and
prepare it to be used. Good. So that's it, Good. Now I'm gonna go with
reviewing my paragraph. And here the tool just divided my paragraphs
into different groups. Okay? Now, there is something that I need
to explain here. Generate three sets of up to five questions
of each paragraph. So if I asked the tool
to do this for me, I'm actually asking
it to generic three different sets of five
questions of each paragraph. So by multiplying,
I can know that I'm getting for each paragraph three sets its moons that my questions are
going to be repeated, but maybe in a different way, but they're actually going to be supplied with the same answer. So let's try with asking for two sets of five questions
for each paragraph. Now, I have small paragraphs, so I'm gonna go with e.g. three questions for
each paragraph. Good. Now I have here the
estimated cost. And this is of course
the coast that was that I'm gonna be
charged for by OpenAI. I can check my balance
for my API account, which actually in the beginning, whenever you started
using the API services, you have around $18 for free. So that will be enough to
use this for a long time because they're charging a
very low amount on this. Good. So I'm gonna
go with change text. If I wanted to remove this text, I can go with change text
or start generating. It is going to take some
time, but believe me, it does worth, and as you
can see, it's giving me. This is the first set, three questions about
the first paragraph. And this is another set
off the first paragraph, but with different questions. What are the benefits
of animals to humans? What do animals offer asked these two questions
are actually alike and they may have the same or somehow some of the
ulcer may be included. So that is good. It's
actually giving me different structures
to the same question, different sentences, which is actually smart because this can actually make the modal understand more about
how can I ask a question and walked exactly
should I respond if the question
looked like this? So the model is learning
and building on generalizing patterns
in an effective way. Now I'm the only thing that I
need to do as I can go with copying my answer and
questions this way, taking the copy of them, e.g. going to start a
new Google Sheet. And I can actually just paste my questions over
there in a new sheet. So I'm gonna go with
the first row over here now and I'm gonna paste. And as you can see, let's give some space over here. I have groups of questions
and their answers. The only thing that
I need to do as we discussed previously
in other lessons, is all I can remove
unwanted rolls into rows. I can also read my data
to make sure I don't have a sub two
duplicates in my data. And I can also add suffixes as the thing that we explained
in previous lessons. Good. So now I have a list of questions and
answers to my topic. Some questions are repeated
but in a different way, which is more formally model. And this was a very nice
and easy tool to use. Now after you prepare your data or your suffixes and
everything is fine, you can download
this Excel sheet and then we will learn in the upcoming lessons how to use this Excel sheet to change its format to adjacent Vermont, where we are actually
going to use this data in that format
to fine tune our model. And that's it for today.
Thank you for watching.
9. Using ChatGPT to generate Python code for data manipulation: Hello and welcome
to a new lesson. In our lesson today
we're going to learn how to work on our data in a smart way using one of the steps that I
spoke about previously. This step is that we want to use GPT to generate for
us a Python code. This Python code is
going to be used in order to work with
a big dataset and change this dataset into
the format of brands and completions of which
you will have to use these pumps and completion
to fine tune your model. So what we're gonna do is
that we will go to chat GPT, we'll do some
prompt engineering. We have to describe
the features and the requirements of
our projects and our dataset in a good
way in order to get a good Python code
that can change all of this dataset into
the shape of prompts and completions and all of
this with one-click on. So I went to chat TBT and
I wrote to the following. First of all, I want
to use a Python code. And this Python code is to be, to convert a prepared dataset to be carved into the shape
of Bronson completions. So I told you that GPT than I need a Python code
for the following, I enrolled every
tip that I need to mention to touch it before
it generates Michael, such as the data, where is it going to be
low default, which file? And maybe tussle important import some necessary libraries. It will take the first
value in each row. It will include this first
value into a prompt, which is the name of the movie. The rest of the
data in each row, which is the rest of the details mentioned in my CSV file about each movie is gonna be prepared in an
appropriate completion. And I'm just including some of the data and not everything
from the dataset, e.g. each movie has details released data director,
actor, and general. That's the information
that I need to include in my prompts
and completion. I also want to judge
anybody to add less. As such as new columns
in the original dataset. These columns will include
generated prompts, engineering hated
completions, two new columns that will be
added by the Python code. Then I have to save my
data or the new DataFrame. And another file, this file
is the one that I'm going to take and use to do the
factoring process. I also had to mention subtle
details about my data. E.g. what is my
detail about e.g. a. Group of obese? These are the features of the movies
that I included in my dataset. So I rotate every feature that was included in my
dataset for the movie. I also gave examples, the name of the movie, the other features on
the Ruby description. So whenever I send
this to chat GPT chat, typically you will be
able to understand my dataset in a very good way. It will know exactly what do
I need my Python code to do. Once I sent this to jot GPT, I received the following. Let's have a preview
on the code. First of all, it's imports
the pandas package. This package helps us deal
with data and CSV files. In Python. It's also going
to import OpenAI. It's going to ask
for the API key. This can help me to communicate and complete the
fine tuning process, like sending some requests and receiving
responses from OpenAI. Good. Now the first thing is that it's going to create a new DataFrame. It's like this data frame. It's going to fetch more
data for all my CSV. But remember, this is the name of the file that two-half
to place over here. What did you name the font that you put inside your drive? Or maybe rather than writing
the name of the file, we can write the
path of the file. Where did I actually
save this one in later. Then it creates the
two lists I asked for, the prompts and the completion. These two lists are to be appended each time like
code reads my data, it's going to take
the name of them. It will be added to
the prompt section. It's going to take
the other features at them to the
completion section. So these are the new dataset, less that will be added to my
DataFrame by the way, good. Now in this for loop over here, it's gonna go read every row, create the prompts
and the completions. And in these two
lines over here, it's going to append
whatever new prompt it's generated to the prompts list. And it's also going to add
whatever new concretions it created to the
list of completions. Finally, when these
two lists star ready, it's going to send these
last two the DataFrame. And it's going to save
the modified DataFrame, that new CSV file in my drive under the
name of repaired data, and that is the CSP file. And once we execute this
code in the next lesson, we will be able to see the new DataFrame that
was generated by my code. So that's it. So I actually
find this helpful, asking shut UP teacher generate
for obstacles that will work the best with our code
is actually so helpful.
10. Running and Executing Python Code on Google Colaboratory: Hello, welcome to
another lesson. In the previous lesson, we
previewed the Python code. That code was built for
us to use to prepare prompts and completions
from a big CSV data file. In this lesson, we will
be running and executing this code step-by-step
until we make sure that we get
prompt syncopations ready to be used in
fine tuning our model. This is the Python code that
was prepared by tattoo PT, according to the prompts engineering we did in
the previous lesson, we previewed the code we so its terms and what
things do we need to change in order
to make sure that this code meets the
requirements of our dataset. Now, in this lesson we're
going to execute the code. So I'm going to copy the
code the way it has. I'm going to go to
the environment which allows me to execute
a Python code. For today's lesson,
we're gonna be using something called Google Colab. We spoke about it in
our previous lessons, but let's remind you of
what Google Colab is. Now once I go to
the Google Colab, it is call up lottery that allows you to run
your Python code editor. And it was done by Google. Once I go into the Colab, it's going to suggest for me my previous projects
or previous notebooks. This is what we call the
files on Google Colab. Other thing is I can just start a new notebook
to work on it. So this is the interface or the shape of the notebook
whenever you open it. And as you can see, there are some things
that we can do. We can change the name of the notebook According to
what we're going to do. We can also add codes or texts to this notable before running and
executing the code. And so what is low
to our Google Drive? Why? Because we want to read
the CSV file as on our. So as you can see, this is my Drive over here. And in my drive I created
a folder called data set, and I included the datasets
that I may need to use. And this is the one we
will be working on. Let's have a preview on
the dataset over here. Let me open it. So what if Google sheets
that have thing is, as you can see on
the first column, we have the movie's name and
then the world earnings, the general at the Director
and the release date, the actors and other
details about the movie. This is the Z to set I'm
going to be working on, I'm going to go back
to my color up. I'm not going to need
this folder over here, the one that I opened
up as an Excel file. So I'm just going to remove it. I need a CSV file to be able to read it with getting
MAC two minute. First of all, I have
to make sure I can access this on
using Google Colab. So I'm gonna go to the files
to the left over here, as you can see, connecting to runtime to enable the fires. And so one, once it's activated, I'll have to do something
called Mounting. Have to mount the Colab toy. And this is by clicking
on the icon over here. Once you click on it, it's going to be mounting. And it's sometimes it may be asking for permissions
of your days. It's Martina to my Google Drive. Once it's learned to date
and everything is calling, you will see that mine Google Drive file
appears over here. So here it says, there, it says, and I can actually enter my
drive, found this place. And Chico, for the
availability of the fun that I added going to the
folder called data set, I can find a new dataset
I included in that bond. And this is actually the CSV file that we
will be working on. The Python code here, copy it and take it
to our Google Colab, add it as a text. So first of all, I'm
going to keep this takes out the top because I want this text to be there for
me whenever I needed. Good. There we go. So this is it this is
the text I prepared. Why did I take it here? Because I want to do
the steps one by one. So this can be my lead
or it can actually help me if I'm not familiar with using the Python
code in a good way, okay, Now, first of all, I have to add a part of the code which
is called the path. This is the path of the
file that I want to read. So I'm going to start
by creating a variable and then this is
the path and what is the path of the file
that I want to read? I'll go to my file here, click on it and copy the path, and paste it over here
in-between double quotations. There we go. And then I can run to make sure that there is no mistakes
in this sentence. That's all good. So let's preview what
we're going to work on. First of all, we have to import
the pandas package and we also have to import the
OpenAI bucket as well. We also have to clarify
our API key and so on. Now let's start
executing our code. First of all, organic go with installing with the
open API library. Pre wanted to use it, and it's going to take some
time for installments. So let's go with
the second part, which is importing the
pandas package from Python. So I'm gonna go with
import those paths PD. And I'm actually creating
a function called bd. Whenever I want to use
the pandas package, I'll be using the p.ball so
long as I run this code. Now, the second thing is
to read my file over here. To read my file over here, I should go with
this and creating a new DataFrame by reading
the previous DataFrame. So I have to include
the name of the file or the path of the file
where I can find it. So let's add another
coat pay status. And rather than writing the
name of the file over here, I'm gonna go with paths. The path that I declared
above over here, which is the path of the file. And we're going to
run this code now. So I can actually read the data in the CSV
file over here. Good. Now, after I read my file, I'm going to start taking the data and adding
it to new less. Now I'm gonna go with creating
the two lists over here, the generated font and
generated coefficient list. So in these two statements, I will be generating
two lists in order to append to them whenever
it is sustainable, creates a new prompt
or near completion. So let's go at another code
and go with the following. Now I have two lists. The first model called
prompt and completions, and they're ready to
be taking values. Now we have the for-loop on this little over which is going
to go to the dataset, create brands, and create
completions accordingly. So I'm going to take all
group together over here. Let's take the rest
of the code as well. Go execute it. So as you can see, for loop is going to go fetch
data from the CSV file. It's also going to
be adding this data to the lists that
I created above. And these two lists are to be added to the original CSV file. And this all is to be saved in a new CSV file called
prepared data CSV. So I'm going to run the code. And once I run the
code and it's stolen, I can tell that they're
most VC is the file called prepared data CSV fired at it till I
enjoy over here. So let's go to the files. I'm just going to close
and open to refresh. It's cool down today into it. And as you can see, here's my foot once
I click on it. So as you can see, this is the name of the movie. It's earnings that
general the director, whatever data I included
previously are here. On the left side I have something
called generated brown. This was added by the Python
code, generated completion. And these two columns are the
only ones that I'm going to need in order to use to
fine tune or bimodal. So that's it tweaks cuz it
decode my data's ready to go. Thank you for watching.
11. Building a Well-Structured Project Directory: Hello, welcome to
another lesson. In our lesson today
we're going to learn how to build a good
project directory, how to structure and
manage data within it. This is an essential skill for any d designs or
fine-tuning project because it ensures that
we can get back to the data and Rick and
understand it in a good way. So there are many benefits for building a good
project structure. One of those benefits is that if you were working with a team, this will allow each
team member to come back to the data and
complete and work on it. This will remove them
as consumptions. It will also remove overlapping, and this actually saves our time and improves our professions. Another thing is that a good structure for a project allows me
to come back to it, to reuse it again, e.g. I. Can take these data files, adjust them, and include
them in your projects. I'm gonna be sharing some of the types that you
may be used in order to create a good
project directory. First of all, remember always to keep your data in folders. Inside those folders, you
can include subfolders, but then moved to use
descriptive names. Remember to use
understandable names of which whenever you want to
look for a specific file, you can come back to its
search, to its binding. Another thing is that you should always keep a read
me file in the read me file and you can keep data about how did
you do the project, how to use it? What datasets to do, use Python codes to
use the urn, so on. So as you can see here, I prepared here an example for a projects directory or how would the products
structure look like? So first of all, you can
include different for folders. Some of them maybe about codes, the script of the codes. Let's imagine that you can use different codes for
preparing data. You have to keep a copy
if you use a code for data analysis and the code that used for
training the model. I keep particles over here. You can also keep notebooks. In the notebooks, you
can keep the files of data that you use,
the original ones. So this is a CSV file, and if you use the next one, you can include it as well. I also here have other fonts, more detail, but these
are the processed ones. So this is called prepared data, and this is the prepared
data from the text. And it is actually a recommended tuning them
with numbers, e.g. if you prepared the data
more than one state, you have to include
prepare data one, data two, and so on. As well as you can
keep the outputs of the process that you
did or the project e.g. like keeping copies from the modals are keeping
a read me files. In the read me file,
you can include whatever data that you
need, breaks out poll, what are the suffixes that you used in the data before you point to and what was
the model that used? What was the modal's ID? Which API key did
you use in order to build and fine
tune this model? You can also write red birds about these low
dose and keep them. This will actually make
it easier for you or any team member or any person to refer
back to the project, see what happened,
take this data, work on it to some other point, tuning products and so on. So that was a brief
example of how to create a good
project directory, how to set up an organized
on top within it. And as we saw, we shared a small
example to that. This project can actually
be uploaded to your GitHub. It can be kept in
specific files on your drive or on your computer. But this actually will be the
key for succeeding in using this project and learning from it and doing some other
projects in the future.
12. How to Choose a Pre-Trained Model for Fine-Tuning: Hello and welcome back. In our lesson today, we will learn how to choose
a pre-trained model that actually meets the requirements
of our own project. We know that GPT-3 is a set
of large and powerful models, but we want to use one of them all in order to create
our own fine-tune. In this lesson, we
will preview with the specifications
of those models. And we will get to
know which one of them would meet our
requirements in order to get to know
those models and think about which one of
them should we choose? We should go and preview
them from their source, which is the Open AI's website. So I'm gonna go do it openly. I specifically to
the playground. Why? Because I want to go drew
their documentation section. Here it says at the top it will go to the
documentation section. Going, scrolling
down on the left, we can go to Modules section. As you guys can see,
this is an overview of every single GPT model
that was done by OpenAI. They reached a typically for
now there has to be 3.5. There's this Dolly than
one that can generate images and there are
some other modals. The thing is we're going
to be depending on GPT-3. Gpt-3 is the basic
set of TBT models, the ones that we can actually
use for fine-tuning. I'm gonna go down, scrolling down to
GPT-3 specifically, we can prove you GPT-3. Now, we want to see the
basic GPT-3 models. Those are the da Vinci, the Curie they are
now and the Babbage. And these are the
only models that people can actually
use for flight. No other models can be used fine-tuning except these three. Now, the thing is,
the latest model of them is also
generated over here, but we will, when we
run a two fine tuning, we only depend on
these four basic ones. And you need to know which one of them can
take more tokens, which one of them can watch the coast of each one of them. What is their capabilities
and abilities? And this is what you
need to consider renal whenever you
choose between them. So the thing is, Da Vinci
is the biggest one, it's the most capable on. It's amazing, but it has a higher Coast going to
puree by badge and ADA. Curie is good in
sentiment analysis. Also, if you were working
on a project that has to classify data into
positive and negative q0 wouldn't be a good
suggestion for that. Ada is the fastest one, it is the fastest one. So if your data
means that you need to work in a fast
responsive model button, not very impressive responses, let's say very direct
responses to false. You can depend on ADA and so on. So these are the
basic four models. Consider them, you have to think about what data types are using, what are the requirements
of your project? Thinking about the speed, thinking about the
resources that you have. How much do I want to pay, or what is my ability to be to do this fine
tuning process. And this is when
you get to be able to know which one
of them to choose. And we will use the name
of the model in order to do the fine-tuning process
using this specific order.
13. Introduction to Fine Tuning Process in Python: To have an overview
over the steps of the fine-tuning process from
the beginning till the end. I prepared this for you
to do the fine tuning. First of all, we have
to download Python. How do we download Python? We go to the Internet. We can search for Python, download heroin or Veritas. You just have to make
sure you download the latest version of Python to be able to
excuse your Python codes. Here we go. So on my laptop, I'm gonna go search for Python. And as you guys can see, I have 3.11 version, which is the latest version. So this is the first step
and I'm ready for it. The second thing we
have to do is to open our C and D. What is the C? Becky? C and D is the Windows
command processor. When you go to the
search bar and your Windows ten or 11 laptop, when you go with C and D, you can find somebody
called command prompt. What is this? This is an interface where you can run commands over here. These commands can
operate or work on your programs or your
operating systems. And it also can help link your
laptop with the Internet, asking, sending for requests,
receiving responses. And so this is where the whole
fine-tuning process is can happen on our CMD or our
commons processor. Did. Other thing that we will
be doing today is we have to locate our
fun on the CMT. The whole processes
is going to happen in a file that we
prepared previously. In the next class.
I'm going to share this slide with you in a link in the description
of the lesson. Then after we locate our file, we have to download some
needed libraries and packages. Those libraries such
as the opening on library pandas and
some other libraries that we will need
in order to work on our data and do the
fine tuning process. After that, we're
gonna be running Python codes and we're gonna
be running so coal mines in order to go to US
Open AI's tools and to prepare our
data for fine tuning. And then we're going to run the Python code to finish
the fine-tuning process. And after we do that, it will provide us
with a model ID. We'll be using this ID
in order to know what happened to the fine-tuning
process to follow our model, the one that we find, to find turn until make sure it was creating change and
everything is okay. So this is a brief explanation, a very short explanation
to the steps. Again, to do the fine tuning, we need to download some
libraries done on Python, run some Python code, do some commands to
communicate with OpenAI, to do with the fine tuning. And finally, to check if our model was fine
tuned at Seoul. So this is a fast overview. In the upcoming lesson, we're gonna be executing
some of these steps by e.g. first of all, preparing our
data in the best way ever. And then fine tuning it and then testing our fine tune
into all those things you
14. Fine-tuning a Pre-trained Model: A Three-Stage Process: Hello and welcome back. In our lesson today
we're going to start executing the
fine-tuning process. And I can split that
into three main stages. The first stage is to
change the format of the data we prepared
to adjacent format. We spoke about that
in previous lessons. We said that any font
that you want to get to a pre-trained model should
have the adjacent format. And we'll see how to change
that using a Python code. The second stage is to take this data and give it to
a pre-trained model to do the fine tuning on
the third stage is to test our
fine-tune the model. As a review, today, we're going to work
on the following. We will first, I'll
make sure that we downloaded Python
and we're going to work on our CMD as we explained in the
previous lesson, we will open our
file and we will start work on our Python code. The first Python code
is to help us to change the CSV data file
to achieve some needle. Once we've finished this, we are sure that
our data is ready to be given to all
the pre-trained model to do the cartoony. So let's go and start
executing our tests. First of all, we're gonna go to our CMD theorists voice EMT. Remember we have
to open our file. So I'm gonna go with CD and opening the
font that I called. Test. This fine over here is included in the description
below the nestling. And it includes the two Python code
there I'm going to use. The first one is to change
the format of the data to adjacent from the
second one is to communicate with
open EI and asked for doing the fine tuning
using our API key is. And here is the explanation to the whole steps that we did. Here. I included here my API key
because I'm going to need it in order to link my
pre-trained model on my phone to
unimodal wave from my API key to be able to use
it for further applications. And this is the data that
we're using good the dataflow. Let's preview our data. As you guys can see, I have the left side, over here, the bronze, and on the right, the second column
is the confessions. In the prompt sides, I have a question and after it, I had spaced and after this
face I included a suffix. This will help them understand. Then this is where
they prompt finishes. On the right side, I
have the completions, that completions as well. It starts with a blank space or a 1D space over here
before the completion. And this space makes the
model understand that this is where the collision starts
and why they innovate. I added a suffix. This suffix will actually
help me or help them understand that this is
where the completion to the first prompt ends unsold. So my data's ready,
I downloaded it, make sure you have a clean tea, a good amount of detail. Make sure everything is set. Make sure you have no
extra information in any other cells or any
other sheets and so on. Once you're ready, once your
data's ready, You are good. Took up. Going back to our C and D, Let's start by running
our first Python code. Before we run the code, let me preview the code. This is it. I'm going to open it as a notes. I'll as text file and
see what does it do. First of all is import
some needed libraries. And then it creates a workbook. In this workbook
it's going to load the data from wind test file. Remember, if you named your data file in
a different name. As you guys can see,
I named the zest. So if you have a different name, you have to write the
name your file over here. Good. Then it starts going to
the sheets inside life, why we only have one sheet. Then it's gonna go and
using this for loop, it's going to start creating the prompts and Nicole
deletions in this format to save them in a JSON format file that is called data
test dot json. Then new fault that's
going to be created by this code is named dates test. And inside this file
you will find our data, everything we included
in this Excel file inside a fine But with
adjacent per month. That's how I prepare
my data to give it to a pre-trained model
to fine tune it. Good. So after I make sure my code is ready and
everything is changed, the name of the
files and everything needs go to our C and
D to run this code. So we're going to
take the name of this file over here,
excel to JSON. So let's just rename,
copy the name. We're going to place the name of the file and click dot pi
sin this to be executed. Once it was executed, as you guys can see over here, look back to our file and we have an extra
fun here it is. It's called data test. Let's preview our new file. So Afghans can see
here is the prom, next to it is the completion. And all of this is prepared
in a decent format. Good, so my data is
ready now for font. Now our next step is
to take this data and send it to be fine tuned
or to fine tune a model. So I'm gonna go to the
next set in my C and D. And that is to run the Python code that helps
me to find two are my model, which is over here. Let's preview our code. I'm gonna go back to open
this code as a text file. Let's have a look. It's also imports,
they needed libraries. It switches API key and after
it finishes my API key, it starts communicating
with OpenAI in order to create a new trained model
or a new fine-tuned model. And it's going to be
asking for requests list of the previous pre-trained
models and other data. What I care about
in this code is the last part over here, because this is where I
get to decide the name of the file that I want to create or the name of the
model that I wanted to create. And what is the name of the band model that you
want me to depend on? So looking at this
data over here, first of all, what am I
going to load my data? The one that I'm going
to use for fine tuning. You have to take it
from a file called digitalis Jason. This is my f1. And what is the name of
the new pre-trained model? First of all, I can name
my model as I want. I called it Mary
disorder and model. And which is the
pre-trained model that I'm gonna depend on, that is to differentiate. So whatever was the name
of this file over here, you have to include
it over here. Remember it's the teeth and L5, you have to choose the name of your model and you also have to choose
the pre-trained model. Now, when I make sure
that my code is ready, I can start on this cuticle. The thing is, I also wanted to share something with you in order to be able to
communicate with OpenAI and receive requests, you have to add your API key to the environment
variables of a laptop. What does that mean? If you go down to the
search bar and go through the
environment variables. One site openness
setting box below, we overhear you have this
top environment variables. Once you click on it, you have to add
the open API key. In the fire variables. The variables allow my
system to communicate using this API key with OpenAI
earned include my EBIT. So if you didn't
have it over here, you're just going to
go with the following. Open, name it, OpenAI from
the score API underscore G. And it has to be written with
capital letters this way. And then you can go
to the playground of touchy P t and get of OpenAI and gets your API key and
include it and save it. So if this was saved
over here correctly, this will make your
sustained able to communicate with OpenAI and Syndrome quest for
fine tuning and receive responses after the
foreign to me plus is good. So to execute this code, the fine tuning code, I'm going to go back to my Cmd. I'm going to go with
typing the name pointer. Okay, let's again copy its name. Go to our CMD and jot high, because this is a Python code. What was, it sends
requests to API. It has to respond
back with one bit. It to this response
came back from OpenAI. It shows me what happened and it actually gives me a
very important thing, which is the ID of my fine-tuned
to model the new one. So this is the whole history of one-to-many models
that I did before. And most of these cases where succeeded,
succeeded, earned. I'm going to show them to
you in my next lesson when I show you how do we test
our pre-trained model. Going back to the end of this
response that I received, I want to see what's
happened over here. As you guys can see, look, this is the ID of the model that I'm trying
to point you in today. So I'm going to need
this ID in order to shake the status of this model, the one that we're
fine tuning now, Da Vinci is the
pre-trained model we're using and we're gonna do with
the fine tuning process. And the name of the new model is narrow disorder and so on. So these are the important
information we need. And as you guys
can see over here, it says it is still pending. Pending means that it's still doing the fine tuning
model because it wasn't, It is still not fine-tuned. It's doing it. If you guys look over here
at the previous time, I tried to find two
unimodal I failed. There was a problem
with something. If you go back above, you will see the
history of every fun to model and each one
of them with its ID. So this actually can help me remember what did I
do previously if I need an ID or for previously a pre-trained
model and so on. So in order to check what happened with mine
and tuning process, I can use an Open AI command that included for you
inside this file. This can be the final stage. This command is
prepared by OpenAI. And this command
actually twice to follow the status of
the fine-tuned model, but you have to include the
ID of the modal over here. So go back to your CMD, copy the ID, the idea of
the fine-tuned model. Here. I'm going to copy it on double gonna
go back over here, paste the ID after
the letter. I. Take this response or this
command, the way it has. Go back to your sandy, paste it and fun. Okay, good. It says that open EI is not recognized as an internal
or external commands. So good that we have
this issue over here. In order to resolve
this problem, I'm gonna go with installing
the opening library. So I go with pip,
install Open AI. Once I do this,
I'll get opening. I installed. Good. It's ready to be used. Okay, let's try again. Paste our command with the
idea of our model and send it to open it and see what's happened
to our fine-tuned model. And this is ours. As you guys can see, open EBI foreign to follow. After it followed, it says that our model was created good. So it did create our
fine-tuned model. And this is the idea
of the new fine-tune. So that's it. This is the whole process
for fine-tuning a new model. And in the next class
we're going to be previewing how to
test our point tool.
15. Testing Your Fine-Tuned Model on OpenAI Playground: Hello and welcome back. In our lesson today,
we're going to be testing our fine tune model. Where do we test that? We're fine tune global
retested on the playground. It's the first place of which
we brought our API keys. And it's the place
where developers or people can actually test if their account and open AI is working and they can
test our APIs key is on. So what are we going to
go to a new web browser? Go to Open AI playground. And once we open it, okay, There you go. I'm going to click on
plaguing old on the top. I love. This is the polygonal
and this is where I get to decimal and
fine tune the model. As we said variously, on the right side over here
is the place where you get to choose the one that
you want to deal with, the model that you wanted
to respond back to you. Once you click on it, you can preview GPT is models. You have these basic models and down you can find
your fine tune. For our model, the one we created
together and visa sense, it was married disorder mobile. So here it is. Now once I choose my model, I can also change the temperature to make sure
the model is not actually using some extra
information or trying to be creative by adding
some more information, referring back to the
dimensional model, I also can decide if I
wanted my answers to be shorter than a specific
limit of tokenization. And I also need to use
this top sequence, the subjects that I added to my data before I did
the fine tuning. So the first thing is I'm going to write my prompt over here. Since this model
is about disorder, I'm going to go with writing
what is mental disorder? And I remember that
this suffix or the separator that I
added at the end of each prompt was
backslash n backslash. So I'm going to
add it over here. I also have to add
a stop sequence, the suffix that I added to
my completions in my data, which was triple hashtags. Once you write the stop sequence
character that you used, you click top. So you added it. Now you can submit your prompt and see what
response do you get. As you see, I get a specific
answer and a short one, even though I didn't actually
change the limit over here. Let's imagine I forgot to add these suffixes over
here and over here. Let's see what's
going to happen. So just remove Y
suffix over here. And I'll remove
this top sequence. And I will put e.g. then tokenization maximal
length and they'll submit, I'll actually get a
very short answer. And if I add more frees up 50, I'm going to submit it again. And as you can see, I get a longer answer and
it's told by this number. So this top sequence is actually
something that helps me. Here, as you can see,
it stopped here. And it's still added some
extra tokens over here in order to use all of
the 50 tokenization. So the thing is, those suffixes are
actually here because they help us make our
modal understand that you have to stop where every
completion you read when you were trained ended with
these treble hashtags? This is how I get a
specific fixed answer to my question over here. That's it. This is the way how
we test our model. We actually come
to the right side. We pick which model
are we testing. Do we write our approach
here and then defining the suffixes of the
completions on process. And that's all.
Mariam Omar, Skilled in AI, Chatbots, Robotics
