LLM Full Course 2026 | LLM Tutorial For Beginners | Introduction to LLM | LLM Training | Simplilearn

[music] Tragedy crossed 100 million users in just 60 days. It's faster than Instagram and any social media platform. And the engineers who have built these systems, they're some of the most in demand people right now. [snorts] But here's what nobody tells you. The core idea behind all of this, and that is LLM. So today in this video, we're going to go from what is an LLM to all the way building, training, fine-tuning, and deploying one. I will understand how they work in the back end. We'll also cover the core concepts behind the transformers and attention mechanism. We'll walk through hands-on code for training and fine-tuning the models and explore real world applications like chatbots, code generation, and document analysis and dive deep into advanced optimization techniques. Also, a quick information if you interested in LLMs and want to take your skills to the next level, you can explore this professional certificate course in Jai and machine learning. It's offered by simply learn in collaboration with ENIs academy IT Kpur. This 11month live online program basically covers everything from the machine learning and deep learning to generative AI prompt engineering and LLM with master classes delivered by IIT Kpur faculty. You'll get hands-on experience through 15 plus real world projects using tools like charg hugging phase and lang chain. You'll also get a certificate issued by IIT Kpur. Whether you're looking to break into AI or upskill in your current role, this program would include career support through Simple Learn by assist. So if this sounds like the right next step for you, the link is in the description below. Before we start, let me ask you a quick question. What is a primary task an LLM is trained to perform at its core? Your options are A searching the internet for relevant information, B translating text between languages, C predicting the next word in a sequence, and D classifying images into categories. Leave your answers in the comments below. So now let's get started. So now let's understand what are large language models. So do you know when you type something under chart GPT what's actually happening in the background? Most people think of these models as a very smart search engine. But that's not quite right. Large language models are statistical prediction machines. At their core, they're doing one thing. Given a sequence of word, they predict what word comes next. Now that sounds almost too simple, but here's one thing. When you train a model on essentially the entire internet, billions of books, articles, code, and scientific papers, something remarkable happens. The model doesn't just learn words. It learns concepts, relationships, and reasoning patterns. An LLM is a neural network typically with billions of parameters that is trained to understand and generate human language. GPT2 had 1.5 billion parameters. GPT3 jumped to 175 billion. And the models being trained today, there are hundreds of billions, potentially trillions of them. Each parameter is a tiny knob. And training is the process of adjusting all of them until the model gets really good at predicting the language. So now we understood about this. Let's understand what's the evolution of natural language process. So this is all started in the early days of natural language processing. We had rule-based systems. So the programmers would manually write the rules such as if the sentence contains X then respond with whites like that. Then came statistical methods. It has bag of word models, engrams and navy based classifiers. The real revolution came with deep learning. In 2013, word to vector showed that you would represent words as vectors. Then in 2017, one paper that changed everything was a paper named attention is all you need. Google researchers introduced a transformer architecture. It is the foundation of every major large language model today. So now we understood the evolution. Let's move into the core concepts. So the first is we need to understand about neural networks. If you never heard of it, don't worry. I'm going to teach you. So a neural network is inspired loosely by the brain. You have layers of nodes called neurons. Data flows in through the input layer. It gets transformed through the hidden layers and produces the output. Each connection between the neurons has a weight. Training means adjusting weights so that the network's output gets closer and closer to the correct answer. The process of adjusting these weights is called back propagation. This model makes a prediction. It compares it to the right answer and calculates the error and propagates that error backwards through the network to update the weights. That's how neural network works. If we do this process n number of times on massive data sets, we get a model that actually learns. So now you got a slight idea of how neural network works. Let's move into But how does that work? Here is where things get genuinely interesting. Before transformers, the main architecture for language was RNN that is recurrent neural networks. They process text sequentially word by word. But one drawback for RNN was that they struggled to remember things from early in the sentence by the time they reached the end. Transformer erased that out entirely. So as you've seen in the sentence, the cat which was chased by the dog and hid under the table was hungry. So when the RNN reads the sentence, it forgets what the sentence is about by the time it reaches the end of the sentence. So the sentence is the cat which was chased by the dog and hid under the table was hungry. So by the time it reads the word hungry, it forgets what the cat was, was it a singular or plural. So it replaces the word were to was and singular to plural like that. So that was the drawback of RNN. So as I've explained it to you, it'll predict were instead of was and it has forgotten that cat was the subject because there are many words in between. It only focuses on the closer noun or the table because the word is at the end of the sentence. And the transformers erase that out entirely. Instead of crossing sequentially, they look at all the words simultaneously. And they use something called the attention mechanism to figure out which words should pay attention to which other words. Let me give you an example. The trophy didn't fit in the suitcase because it was too big. Now as human we can easily understand what does it refers to. The trophy or the suitcase. You instantly know it's the trophy. The attention mechanism loves a model to make that connection. It learns to draw relationships between distant words, capturing context in a way that RNN simply couldn't. Mathematically, attention computes three things for each word. A query, a key, and a value. Think of it like a search engine. The query is what you're looking for, and the keys are what's available, and the values are what you can actually retrieve. So, basically, training a large model from the scratch takes weeks on thousands of GPUs and costs millions of dollars. So instead, what if you could take a model already trained on a massive general data set and fine-tune it for a specific task with a fraction of compute. That is what transfer learning is all about and it's why you can take an open-source model like Llama and within a day you can turn it into a specialized assistant. So now that you understood transformers, let's understand what happens inside the transformer and let's build a mental model of what actually happens when the text goes into the large language model. So the first step would be sending the input text and then we have tokenization. Here our text doesn't go into the model as words. It goes in as tokens. A token is basically a rough three to four characters on average. The word unhappiness might become three tokens. Unhappy and nest. Now you might be wondering why tokens instead of words because this lets the model handle any word by breaking them into recognizable pieces. The next step is embedding. Each token gets converted into a highdimensional vector typically hundreds or thousands of numbers. And next we have transformer layers. The embedding pass through the multiple transformer blocks. And modern models have anywhere from 12 to 96 of these layers. Each block has two main components. A multi-head retention layer and a feed forward network. Multi head attention means the model runs attention multiple times in parallel. Each head learns to attend to different types of relationships. One head might focus on syntax and another on long range dependencies. Since the transformers look at all the tokens simultaneously, they don't know the order of words. For example, let's say dog bites man and man bites dog would look identical without some way to encode the position. So the original transformer added a positional encoding vector to each token embedding. It's a mathematical signature. Basically, it's based on position in the sequence. It used sine and cosine waves at different frequencies. So every position gets a unique pattern the model can interpret. Newer models like GPT use learned positional embeddings instead. And the very latest architecture use rotary position embedding. It handles very long sequences much better. So the last step is the output. Here the next token will be predicted. Now as we've seen a journey of a prom through LLM. Let's understand about big three architectures. The first one is GPT decoder only. GBD is what's called a decoder only transformer. It's autogressive. It generates text one token at a time from left to right. Each new token based on everything before it. It's perfect for text generation, chatbots or creative writing. Next, we have bird that is it is a different approach. It's an encoder optimized for understanding the text. It looks at a full sentence birectionally. It can see context from both left and right. It is suitable for classification, question answering and information retrieval. Google uses BER to significantly improve the search quality. And the next we have T5 encoder decoder. It combines both an encoder that reads and understand input, a decoder that generates the output. It is used for transformation tasks, translation, summarization, and question answering. So now let's understand about the workflow and what happens inside it. Pre-training is the initial phase. The model learns from raw text using a self-supervised objective typically next token prediction. So this is how the training loop works. First it loads a batch of token sequences. Next comes forward pass. It gets a predictions and then calculates the loss. How wrong were the predictions and then it back propagates it and compute the gradients and it next optimizes the step updates the weights and repeats this process n number of times. So now we've got an idea of how LLM works. You might be wondering how it works in the real world. Let me tell you. So now we've seen how the LLM got trained. Here are a few scenarios of real world applications and how LLM works in real world. The first one is chatbots. Chatbots use AI to stimulate humanlike conversations and assist users in real time. Context management helps them remember previous interactions for better response. Tools like Lchain and Lama index helps connect data sources and improve the reasoning. The next one would be code generation. GitHub copilot autocomplete and testing. AI tools can generate code suggest completions and help developers write programs faster. They also assist in identifying errors, writing test cases and debugging issues. This boosts productivity and it reduces manual effort. Next one is document analysis. AI can quickly process large documents and extract key information within seconds. It helps in summarizing, identifying important clauses and analyzing data efficiently. This saves time compared to manual reading. The next we have API integration. APIs allow developers to integrate AI capabilities into applications without building models from scratch. Hosted services like OpenAI and Anthropic provide readyto-use AI models. So this makes the development faster and scalable. Next we have self-hosting. It means running AI models on your own infrastructure instead of using the cloud services. Tools like Olama and BLM enable better control privacy and customization. It's useful for organization with sensitive data and ondevice AI. It runs directly on the mobile or local devices without needing the internet. Frameworks like CoreML and ML enable faster processing and better privacy. It's commonly used in applications for features like face detection and voice recognition. So now we understood how LLM works. Let's just build our first LLM project and understand how it works. We'll first install the necessary data sets from the transformers. Also change the runtime to T4 GPU when you run the code because if you're uh using the CPU it would take around 4 to 5 hours for you. So just run it in T4 JPO if you want uh like if you want the output in 10 to 15 minutes and then rerun it. So let me tell you how it works. First we'll install the data sets from the transformers and then uh we load the GP2 tokenizer which converts the raw text into the numbers the model can understand here. Then we will build a data set class. It will define a custom class that reads a text file, tokenizes it and chops it into 128 token chunks. So here each chunk becomes one training example basically. So as you can see it's getting trained here. Next we'll be building a model. We created a smaller version of GP2 with six layers and 256 dimensions. It is uh lightweight enough to train on free hardware, but I think it's still a real working transformer. Next year, what happens is we'll be downloading the data set that is wiki text data set. It will clean the Wikipedia articles and save it as a text file that our data set class will read from. And next comes setting up the training arguments. Here we define how training should run and like the batch size, learning rate, automatic GPU detection, everything we'll mention it here. And next comes the training. Here it will train and predict the next step. So it's getting trained. So what happens here is that the trainer runs the full training loop. It'll watch the loss number drop with every 50 steps and then that's how your model is learning the language patterns. That's how that's what is happening here. So as you can see it has completed training. Now we can test it. So I've written a code in such a way that it'll import pipeline from transformers and it should generate a sentence with 100 tokens max tokens. Basically here the sentence should be history of artificial intelligence and after this the next sentence should end after 100 tokens. That's how the code is written. So it has given us the correct output history of artificial intelligence. It started it with that and then in December 2015 it has started and ended after 100 tokens. That's how you know it has the training is done successfully. So now we can train with more examples. So now if you observe I have written a code in such a way that it should give me the answer for the prompt that I have given it that is what is machine learning explain the water cycle the benefits of exercise it should read from the Wikipedia text and provide me the correct answer but that's not happening here what it has given us is that some random sentences from the Wikipedia text what is machine learning and the answer is and the time of gone ment some random answer it is not at all right so this is the time to fine-tune the code so for finetuning what we do is we install all the necessary data sets from the transformers and that is PFT and bits and byes that is how this fine tuning works the next is we load this Facebook's opt 125 million model uh it is basically a real world uh model which has 1.5 million parameters. So it's enough for us to train it on Google Collab and then we'll have this Laura configuration. This is mainly very important for fine-tuning the code. This is what makes the difference. So what it does is it freezes all the original model weights and it adds tiny trainable matrices only for the attention layers. So what happened in this sentence that is print trainable parameter is that only 0.5 0.5% of the model is actually being trained. So that's what happens in Laura. It's it gives us the same results but with less number of uh parameters being trained. So it is given an error that the data test data set text field got unexpected key. Okay, it is not identifying the data set text field. So here we have not added data set training parameter. Here I've given that the data set text field is text. So the error has gone now and it's getting trained. So here we will load the dolly data set. It's Dolly data bricks 15,000 parameters. It uh it is a real instruction following data set with 15,000 human written examples. So now here we will structure each prompt in a clean instruction how it should produce the output such as the context should be good and it should be a good response. This teaches the model not just the language but it will tell how to follow the instructions and give us the proper output and here we've provided is uh the training arguments how much the learning rate should be and logging steps all of these. So while it's fine tunes let me tell you an example of how fine-tuning actually works and how important it is with a real example. So let's just say the base model is a well read person who knows a lot about everything but we want them to specifically get good at following instructions. So instead of telling them everything, we'll just give them a short introduction or like a training session that is about fine-tuning and Laura makes that training session even better like instead of updating every single thing the person knows, we just add a small notes on top with specific information and only this information gets updated during the training and not the entire topic. So that's why only 0.5 of the parameters are trainable. So the data set that we are using that is dolly 15,000 is basically 15,000 examples of here's an instruction here's a right response all these prompts so that the model reads all of these and learns the pattern of being helpful for the result. So our main goal is to get proper answers for the questions that we ask. So here when we asked it about machine learning yeah machine learning water cycle it has given us some random sentences which doesn't even make sense. So instead of giving us these answers we're trading it to give us the perfect answers for the perfect questions. So it has also given us the output such as how many training samples per second it has done and training steps per second. It has given what all happened here. Let's just test the code now. So yeah, so now as you can see the answers are somewhere near to what we were expecting like what is machine learning? It has given us the exact answer. It is a concept of computing mathematical model that is able to generate information from data using different methods. So it has given us a correct answer. The next is water cycle. In simple terms it is not given because it has not been fine-tuned enough to give us the all the answers. So we have to fine-tune more and more. So the more and more you fine-tune a model the exact answer it will provide because then it'll understand what to you know generate for each answers each question. So you might get an perfect answer if you train it for more. Here we having what are benefits of excess? It is given as the exact answer. It improves your mood. It'll you'll feel more confident, focused. It gives you a break. All this is a correct answer instead of just giving random sentences. That's what fine-tuning does for the code. Let's just test it with uh different examples as well. So here I've written a code to generate the capital of France basic questions to understand if it has been fine- tuned or not. And I've given different types of uh prompts such as the first one is capital of France and the next is how internet works that is explanation sentence. Next we have tips for learning programming that would be a list because I'm asking for tips and then write one sentence story here it is story that should be creative and what is the difference between GPT? GPT and BRT are the LLMs. So I've given different uh models and different topics. So let's just see. So yeah, it has given us the answers. The couple of Francis Paris and Internet is a free open source web application. It has given us the correct answer. Three tips for learning programming. Now here it is not given us. You can learn programming by listening to the programming. If you're not familiar, you can find a good way to learn. See is not it just gave us a random number. So as I'm telling you, you just have to fine-tune it more and more. Just try to add more details into the prompt and train it more and more to get us to get a perfect answer. Some like 60% of it has been trained because it is giving us the answers. A sentence story about a robot. There are three types of robots. Uh there are two types of robots. It has given us. Uh the other robot is a human. So it has read what is there in the Wikipedia and it's just giving us the answer. Uh difference between GPD and bird. It is a program designed to help scientists underlying physics of matter. Now here also it is given the exact same answer. It's programmed to help scientists understand underlying physics of matter. It is given the same answer. So we need to finetune it more. So I've written a code. I fine-tuned it more and more but it is not giving us the exact answer also here now it's changed France capital of France is just telling the population of France instead of capital and the other is explain how the internet works internet is a network here it is given us the correct answer it is also given what is a server and client website everything and three tips for learning programming build a programming language, read books, play games. See, it's not a correct answer. Play games would not help you to learn programming. So that's how it works. Basically, you have to fine-tune it more. And send in story about robot. Robot is made up of robots with goal of solving a problem. It is correct only. The robot is controlled by a command object. It can be used to manipulate the object. Robot is used to look up at the object and respond a few commands. So it is also given the correct answer as well. GPD is a software tool used to map the physical world. It is uh here it is given the previous answer only same answer. So we have to fine-tune it more and more like give it more instructions more prompts basically help it to learn how it should uh identify few terms like instead of capital it has identified it as population and then gave us the to get the answer. So that's about building our first LLM. Let's also name it and save it. Uh yeah, so that's about building our LLM and I hope you understood how the LLM works in the back end and how it's generating the answers. So basically it will take a data set that we have given that is wiki data set. It'll clean it and it will read the data set basically and it will generate the answers for the prompts that we've given it. So it's basically used in many real world applications like I've told you about document analysis, chat bots or hosting APIs etc. So yeah this is how LLM works So now we've seen the demo. Let's quickly understand how we actually plug this into a real product. The first option would be hosted API. Here the open AI anthropic we can send a prompt and get a response. You will be having zero infrastructure and you can pay per token. So that's beneficial. The next one is open source and self-host. So deploy llama or mistrill using Lama or WLM and you will have more control and better privacy and lower marginal cost at scale as well. Next we have hybrid. So hosted API for userfacing features, self-hosted for batch processing or sensitive data. For web we have fast API or node.js backend and server send events for streaming react front end that basically render in stream tokens and gives you a typewriter effect. Now let's just quickly understand about advanced techniques and optimization. So the more the bigger models are the better it would work but they're slower and it's more expensive as well. So now let's understand about uh how we can reduce it. The first one is quantization. We can reduce the weight precision from FP32 to FP16 int8 to int4. A model which needs 28GB at FP32 might run in 7GB and at 4bit quantization and with small quantity loss as well. So the tools you need are GBDQ bits and byes. Hugging phase has quantized versions of major models which are ready to download. Next we'll see about the cost. Uh here in production latency matters the most. Users will notice when responses feel slow. Batching is when group multiple requests together better GPU utilization. It would be lower cost per request. And then we have KV catching during generation cash value pairs to avoid recmp computation across the terms and basically most interference frameworks do this automatically. Then we have speculative decoding. It uses a small fast model to generate candidate tokens and then the large model verifies them in parallel. It works in two to 3x speeds with no quality loss. Next, we have prompt caching. It caches the processing of long system prompts. So, if your system prompt is 10,000 tokens and you run 10,000 requests a day, this saves enormous compute. So, let me tell you about some cloud platforms. So, the first one is AWS SageMaker. It's best for training and hosting and it's very important for managed model access. The next one would be Google cloud. It's best for TPU access and Vortex AI for managed machine learning. It is strong for training very large models. Next we have Azure which has deep open AAI partnership. It's best for enterprise compliance with GPD4 API and Lambda Labs. It's cheapest and it is pure GPU cloud for training and it has lower cost as well. and hugging face. It's basically simplest part to hosting opensource models. Now let me tell you about ethical considerations and being responsible. So that's about the cloud platforms. Now let's understand the problems that are actually caused in real world. These are the problems caused in real world. So understanding them would be better for you for working with LLM. So LLMs inherit biases from the training data documented examples models generating gendered compilations for professions performing words for certain names. It reduces uneven quality across languages. So you might be wondering how to address it. Basically diverse training data bias audits using tools like fair learn or IBM AI red teaming before deployment model cards documenting capabilities and limitations. When you deploy an LLM, you are basically responsible for what it does. So basically LLMs confidently generate false information as well. They don't have a concept of I don't know unless trained for it. So never use an LLM as the sole authorative source for consequential data or consequential decisions in that matter. Be very careful while you're giving or training the data. Malicious users can craft inputs that override your system prompt as well. So be careful. Know exactly where user data goes. Many API providers use your data to improve their models by default. So check the terms and conditions. The same capabilities useful for writing assistance also enables generating misinformation at scale. So think carefully about misusing vectors before deploying it. So the EU AI act classifies AI systems by risk level and it imposes requirements accordingly. High-risk applications in healthcare and critical infrastructure face strict rules. GDPR and data protection laws restrict how personal data can be used in training. So the bottom line is it's not responsible just for ethics. It's also increasingly legal compliance. So who understands both technical and ethical dimensions will be the one who is trusted to deploy the project. So since you've built your project, how do you get it in front of users? So first option would be docker containerize your model server environment weights dependencies into a reproducible image. If it runs on your machine, it'll run on your prod as well. So the standard interface server for production LLM deployment is VVLM. It implements page retention or memory management technique that dramatically increases the throughput. So it's significantly more concurrent requests on the same hardware. The next one would be Kubernetes horizontal pod autoscaling. It means that your system automatically adds GPU instances during traffic spies and it scales down during the traffic. So a deployed model is a living system. You need to know when it's misbehaving. So the metrics to track are latency, P50, P95 and P99 response times and throughput requests per second and error rates should be failure timeouts and safety filter triggers. and always know what you're spending on. Models can degrade as usage pattern changes. So, benchmark tests run regularly and catch the regressions. So, as we've come to an end, let me tell you three case studies. The first one would be customer support automation. So, software as service company deployed a fine-tuned LLM on the documentation and 60% of Taiwan tickets automatically got resolved and average resolution time dropped from 4 hours to 3 minutes. So, Cesat scores actually improved because users got instant answers to simple questions and humans in this case would take a lot of time and had complex issues. Second case study is legal documentation review. A law firm's LLM pipeline flags unusual clauses, summarizes key terms and identify different issues. It associates reviews in 20 minutes what used to take 2 days for a human. The firm didn't eliminate associate roles. It just took 4x more clients. So here if you understand the pattern the LM are working faster and better than humans. They're not replacing them. The organizations winning with AI are the ones that figure out the right human AI workflow and they'll not try to eliminate human judgment entirely but they work faster than that. And with that we come to end of the session. We cover the basics of LLM from understanding foundational concepts of large language models to building training and deploying one. If you have any questions or would like to discuss any concept further, please leave them in the comment section below. And if you like this video, like and subscribe for more videos like this.