LLMs (Large Language Models) are large neural networks models that are built up of interconnected layers of nodes.

LLMs parameters are connections between the input tokens of the model (or the nodes of a layer) and other nodes of another layer. They are numerical values that represent the model's understanding of a domain (global or specific). Each parameter has a weight; a numerical value given to a connection.

Types of LLMs:

• Representation models: encoder-only models that are used for specific tasks.
  Representation models are sequence-to-value models; they take a text (sequence of tokens) as input and generate a value.
  Example: "The weather today is great!" ==> "1"
  
  
• Generative models: decoder-only models that generate text.
  Generative models are sequence-to-sequence models; they take a text (sequence of tokens) as input and generate text (sequence of tokens). They are not trained on specific tasks. When given a text, the generative models need to understand the context and they need to be given clear instructions on the expected output.
  Example: "What's 1 + 1?" ==> "The answer is 2"
  

The transformer architecture is an encoder-decoder architecture. It has 12 encoders and 12 decoders stacked together. The encoder-decoder models are sequence-to-sequence models. They are pre-trained using masked language modeling where sets of tokens are masked:
INPUT (masked sequence): LLMs can be used for [MASK], a form of [MASK]. OUTPUT (predicted masks): [MASK] = text generation [MASK] = generative ai
T5 (Text To Text Transfer Transformer) architecture is an encoder-decoder architecture.

Applications of LLMs:

• text generation
  
• text classification
  
• text clustering
  
• semantic search
  
• ...
  

They are used for specific tasks: for example, text classification

Example: predict masked token
INPUT: [CLS] LLMs are large [MASK] models OUTPUT (predicted value for [MASK]): language
Examples of representation models:

• Google BERT large model (open source) consists of 340 million parameters.
  BERT: Bidirectional Encoder Representations from Transformers
  

The model takes an input (a.k.a. the user prompt, user query) and returns an output that's expected to follow the user prompt.

Generative models are also called completion models (auto-complete the user prompt).

Example: predict next token
INPUT: "LLM is a" OUTPUT: "Large Language Model"
Example: auto completion
INPUT: "What's LLMs?" OUTPUT: "LLMs are ..."
Examples of representation models:

• Open AI GPT-4 model (proprietary) consists of 1.75 trillion parameters
  GPT: Generative Pre-trained Transformer
  
  
• Google Gemini model (proprietary) consists of 27 billion parameters
  
  
• Meta AI LLaMA 4 model (open source) consists of 2 trillion parameters
  

Creating a language model takes place in two steps: Training, Fine-tuning

Training:

• The model is trained on a lot of data allowing it to learn the language grammar and understand the semantic, context, and patterns of text.
  
• The training allow the model to predict the next token (doesn't target specific tasks).
  
• The trained models (a.k.a. the pre-trained models) are called foundation models or base models.
  
• The training takes a lot of computation (GPUs, VRAMs).
  
• The training takes a lot of training time.
  
• The training is very costly.
  
• The training requires a lot of data (un-supervised).
  

Fine-tuning:

• Fine-tuning uses the pre-trained model to train it on a specific task (for example: text classification task).
  
• It produces fine-tuned models.
  
• It takes less of computation (GPUs, VRAMs).
  
• It takes less of training time.
  
• It requires less data (supervised).
  

Training techniques: supervised and un-supervised

• Supervised training techniques: uses labeled data (supervised text classification).
  
• Un-supervised training techniques: no prior labeling (text clustering).
  

Generative Models can be trained to answer questions. They can be fine-tuned to create models that respond to instructions.

Training representation models uses a technique called masked language modeling. It masks tokens of the input and instructs the model to predict it.