00:00Are you sick and tired of ChatGPT not giving you the right responses?
00:04Well, maybe it's time to consider training your own AI.
00:07In this video, I'm going to show you how to fine-tune your own AI model
00:10that's tailored towards your needs and preferences.
00:13Whether you want an assistant that understands your specific workflows,
00:16speaks in your tone, or has a certain deep expertise in a particular niche.
00:20To do this, I'll be using a platform called FPT AI Factory.
00:24They're an NVIDIA preferred cloud partner that helps provide users
00:27with end-to-end solutions to train AI models.
00:30You can visit my link in the description below for $100 in free credits,
00:34which is going to be more than enough to follow along with today's demo.
00:37So yeah, make sure you take advantage of this opportunity.
00:40Okay, let's get started.
00:41First things first, you might be wondering,
00:43Lucy, why should we build our own AI assistant when ChatGPT exists?
00:48Well, the main advantage is customization.
00:50ChatGPT gives us generic answers, and so if we want to do a specific task,
00:54we have to explain the context every time.
00:56For example, saying something like,
00:58act as a career coach and give me advice.
01:01However, even with that prompt,
01:02ChatGPT sometimes provides some pretty standard responses.
01:06What I wanted was to build an AI assistant
01:08that can act as an interview coach for cloud learners,
01:11which, by the way, I'm going to show you how to build in this video.
01:14Okay, now that you understand the benefits,
01:16here are the steps you'll need to take to fine-tune your own AI model.
01:19First, start with a pre-trained model
01:21that has already learned from billions of web sources.
01:24Next, prepare your custom training data.
01:26This is where you'll teach your model your specific knowledge.
01:29After that, fine-tune the model on your data.
01:32And finally, deploy and test your model.
01:34But before we get right into the steps, here's the problem.
01:37To fine-tune a model, you need some serious compute power.
01:40We're talking $2 to $5 an hour, sometimes even more.
01:43Because let's be real, GPUs are expensive.
01:46Even if you're just experimenting, this can add up fast.
01:49And that's been a barrier for people who want to try this.
01:52And so, since FPT AI Factory is giving away $100 in free credits
01:56for users to try their platform,
01:58I thought it'd be the perfect opportunity
02:00to show you how to fine-tune your own AI model.
02:02Again, you can redeem this in the description below.
02:05Alright, let's head over to my computer now and start building.
02:08Alright, so the first thing we need to do
02:10is go on to FPT AI Factory and create an account.
02:14Once you log in, you'll be directed to this dashboard
02:16where you can redeem $100 of free credits.
02:19What we're going to do now is launch a new AI notebook
02:22to start the process of creating our own AI assistant.
02:26So, let's click on AI notebook over here.
02:29For the GPU, let's select GPU H200.
02:32And here it is.
02:34We've now got a Jupyter notebook environment ready to use.
02:37What's great about this is that pretty much everything we need
02:40is already installed.
02:42Python, PyTorch, CUDA drivers, and so on.
02:45Because of this, we can start working inside the notebook straight away.
02:48Now, we do need to install a few extra tools to help us train the AI
02:52because we need some specific add-ons like transformers and datasets.
02:56These are the tools that will actually teach the AI
02:59how to read our notes and learn from them.
03:01So, let's install the required packages and click run.
03:04Nice, okay.
03:05Everything is downloaded successfully.
03:07By the way, you'll be able to access these instructions
03:10in a GitHub file that are linked down below.
03:12Okay, now that our tools are ready,
03:14we need to give the AI something to learn from.
03:16This is the most important part.
03:18So, in my case, I've prepared a list of questions and answers
03:21based on my own experience to help learners get hired in Cloud and AWS.
03:25This training material includes things like how to answer star interview questions
03:29and what interviewers look for in candidates.
03:32Depending on the type of AI assistant you'd like to build,
03:35you can add in whatever relevant material you'd like.
03:37I'd recommend formatting your text in a question-and-answer format
03:41because this will be useful later where we form our prompt response pairs
03:45that the AI can learn from.
03:46Alright, once you've got your text pasted in the notebook,
03:49all you have to do is click run or you can just hit shift enter.
03:53This is going to format our notes into a special file called train.jsonl.
03:58Okay, we're now ready to train the model.
04:00Let's paste in our script and we're going to train our AI model
04:03based on the material we provided earlier.
04:06The model we're using is called Mistral
04:08and you can think of Mistral as a very smart student
04:11who has read the whole internet
04:12but haven't read our specific interview notes yet.
04:15We're also using something called LoRa
04:16which lets us add in our new notes
04:18as a small expert layer on top of the AI.
04:21So instead of training the whole thing,
04:23which is slow and expensive,
04:25we can just teach it our specific style.
04:27This saves us a lot of time
04:28and keeps the AI focused on exactly what we wanted to learn.
04:32Cool, so let's wait for it to run
04:33and this training should only take you a few minutes.
04:36Alright, here we go.
04:37It says training complete
04:39and so our new AI assistant is now saved
04:42in this AWS playbook model.
04:44What we have to do now is load the model
04:46to see if it actually works.
04:48So let's copy in the loading script from the instructions
04:51and paste it into a new cell.
04:53What's happening here is that we're loading
04:54the base Mistral model again
04:56but this time with the training that we've done.
04:58This script also sets up a special ask AI function
05:01that cleans up the answers.
05:03It tells the model to stay focused
05:04so that it doesn't make up random things.
05:06Okay, once that's done,
05:08we'll get a confirmation message
05:09saying that the model has loaded successfully.
05:12Great, so now that I have the model loaded,
05:14we're now ready to ask it questions.
05:16Let's try it out.
05:17In a new cell,
05:18I'm going to use the print command
05:20with our ask AI tool.
05:21So let me type in
05:23how do I write an effective cloud portfolio?
05:28Nice.
05:28Okay, so it's explaining how to choose your projects,
05:31how to show visual evidence
05:33and where to publish your portfolio.
05:35It even reminds you to treat it as a cloud diary,
05:38which is what I would say.
05:39Let's try one more question.
05:41So I'll type in
05:42what are the top cloud computing careers?
05:48And here we go.
05:49It's breaking down the different roles
05:51like solutions architect,
05:52cloud architect,
05:53and explaining what they do.
05:55It's not just giving a generic answer from the internet.
05:57It's actually using my specific style.
06:00And honestly,
06:01this same approach works for everything.
06:03Whether you're building customer support bots,
06:05technical documentation assistance,
06:07or just personalized AI tools.
06:09Whatever expertise you have,
06:11you can turn it into an AI that sounds like you.
06:13So that was a walkthrough
06:14of how to fine tune your own AI model.
06:17But let me tell you a bit more
06:18about the platform that we used,
06:20FPT AI Factory.
06:21It's a complete end-to-end AI platform
06:23that actually goes way beyond
06:25helping you fine tune models.
06:26They've got FPT AI Agents,
06:28which is a comprehensive platform
06:30for multilingual AI development,
06:31and a whole bunch of other AI
06:33and infrastructure offerings.
06:34What stood out to me was their performance.
06:36The GPUs that they have are seriously fast.
06:39The flexibility is nice too,
06:41whether you want containers,
06:42notebooks,
06:43or serverless endpoints.
06:44It's all built on NVIDIA's latest infrastructure,
06:46which is a gold standard
06:47for AI computers these days.
06:49So yeah,
06:49if you're building anything with AI,
06:51make sure you check them out
06:52because their platform handles everything
06:54from experimenting with ideas
06:56to serving models to real users.
06:58Okay,
06:58this brings us to the end of this video.
07:00Let me know in the comments below,
07:01would you train your own AI model
07:03or would you prefer to just use ChatGPT?
07:05I'd love to start a discussion.
07:06Thanks for watching
07:07and I'll see you in the next video.
07:09Bye for now.
