Elambharathi Thangavel

Using Large Language Models (LLMs)

5 min read

LLM

Advanced type of artificial intelligence model specializing in processing, understanding, and generating human language (They consist of a vast number of parameters - Parameters are aspects of model learned from training data enabling recognition, prediction and production of language)

Transformer Architecture : Most of the current LLMs are based on a machine learning approach called a Transformer Architecture, which allows for efficient processing of sequences of text and the ability to pay attention to different parts of the input when generating an output

Model creation process

  1. Data collection
  2. Pre-processing
  3. Model design
  4. Model training
  5. Computing Power
  6. Additional Training
  7. Deploy Model

Using LLMs with PartyRock

PartyRock → Service provided by AWS that allows us to experiment with Amazon Bedrock

  • When you give any prompt to PartyRock, it’ll build an app in the background in Bedrock using foundational models and present it to you

LLMs for Code Generation

CodeWhisperer is a service in AWS. Given an text prompt, it’ll generate code for us (Works natively with AWS Cloud9 and VS Code as well)

  • Users simply input a comment describing the desired function.
  • Amazon CodeWhisperer then uses autocomplete to suggest relevant code.

Natural Language Models

Branch of AI that combines computational linguistics (rule based modeling of human language) with statistical, machine learning, and deep learning models

Key Processes in NLP

  1. Tokenization - Breaking down text into smaller units (tokens), such as words or phrases, for easier analysis
  2. Parsing - Analyzing the grammatical structure of sentences to understand how words relate to each other.
  3. Semantic Analysis - Understanding the meaning behind words by considering context, synonyms, and ambiguities.
  4. Contextual Understanding - Utilizing the context of surrounding sentences to enhance interpretation, including understanding implied meanings and intentions.
  5. Statistical Inference - Using probabilities to predict subsequent words or appropriate responses in a conversation
  6. Machine Learning Integration - Continuously learning from new inputs to improve language prediction and understanding

Evolution of LLMs

  1. Rule-Based Systems - Early models based on explicit grammatical and syntactic rules, limited and inflexible.
  2. Statistical Models - Shifted to probability-based predictions, enabling more complex language processing.
  3. Machine Learning Models - Began understanding context and patterns, but struggled with long-range dependencies.
  4. Neural Networks (RNNs and LSTMs) - Improved handling of sequences and memory, marking significant progress.
  5. Transformer-Based Architectures - Introduced attention mechanisms, allowing models to weigh the importance of each part of the sentence, thus capturing intricate relationships and context.

How Large Language Models Work

  1. Transformer-Based Architectures - These are types of neural network architectures that have become fundamental in state-of-the-art natural language processing (NLP) models. They’re particularly adept at handling long sequences of data and learning complex patterns.
  2. Attention Mechanism - A core concept in Transformer architectures. The attention mechanism, particularly self-attention, allows the model to weigh the importance of each word in a sentence in relation to every other word.
  3. Context Capture in Text - Transformers are notable for their ability to capture context across long stretches of text. This is a significant advancement over rule-based, statistical, and traditional machine learning approaches.
  4. Tokenization - The process of breaking down a sentence into tokens, which can be individual words or parts of words.
  5. Embeddings - The numerical representations of words or tokens, typically in the form of vectors. Embeddings convert words into a format that can be processed by neural networks and other algorithms. They capture and quantify aspects of word meanings, their use in different contexts, and their syntactic roles.
  6. Self-Attention in Transformers - This technique is used to calculate attention scores for each token, determining how much focus to put on other tokens in the sentence. It leads to a context-aware representation of each word.

Encoders and Decoders

In Transformer models, these are distinct components. The encoder processes the input text, converting it into numerical values known as embeddings. The decoder then uses these embeddings to generate output text.

Decoding Strategies

  • Greedy Decoding - Picks the most likely next word at each step. Efficient but can lead to suboptimal sequences.
  • Beam Search - Tracks a number of possible sequences (beam width) to find a better sequence of words. It balances between the best local and overall choices.
  • Top-K Sampling - Randomly picks the next word from the top K most likely candidates, introducing randomness and diversity.
  • Top-P (Nucleus) Sampling: Chooses words from a set whose cumulative probability exceeds a threshold, focusing on high-probability options.

Building LLMs with Code

Tokenization - The sentence is broken into tokens

# Tokenization using Hugging Face's Transformers 
 
from transformers import AutoTokenizer 
 
tokenizer = AutoTokenizer.from_pretrained("gpt-2") 
tokens = tokenizer.tokenize("A young girl named Alice sits bored by a riverbank...")

Embedding - Tokens are transformed into numerical vectors

# Embedding and Processing with a Transformer Model
 
from transformers import AutoModel 
 
model = AutoModel.from_pretrained("gpt-2") 
inputs = tokenizer("A young girl named Alice sits bored by a riverbank...", return_tensors="pt") 
outputs = model(**inputs) 
last_hidden_states = outputs.last_hidden_state
  1. Self-Attention - The model calculates attention scores for each token, understanding their importance in context.
  2. Layered Processing - Multiple neural network layers process these embeddings, enhancing understanding at each step.
  3. Encoders and Decoders - The encoder processes input text into embeddings, and the decoder generates output text, using strategies like Greedy Decoding or Beam Search.

Graph View