Advanced Topics in Computer Science: Advances in AI

Advanced Topics in Computer Science/Business Computing (CS3001/3606)

Advances in AI: Large Language Models (LLMs) for Question Answering (QA) & Text Generation

David Bell

News

Task

ONE

Task 1

Quizzes & Discussion

Introduction

Overview of Topics

• Language Models

• Large Language Models

• Fine-Tuning Large Language Models

• Prompt Engineering

• Retrieval Augmented Generation

• LLMs in Fintech & Healthcare

• Evaluation

Concepts

• Language Models: Statistical models predicting the next word in a sentence based on prior words.

• Large Language Models (LLMs): Advanced language models with millions/billions of parameters capable of understanding and generating human-like text.

• Embeddings: Continuous vector representations of words or phrases that capture semantic meaning.

• Fine-tuning: Adaptation of pre-trained models to specific tasks by continued training on smaller, task-specific datasets.

• Prompt Engineering: Designing inputs to effectively guide the response generation of LLMs.

• Retrieval Augmented Generation (RAG): Technique that enhances LLM outputs by integrating external factual sources.

• Zero-/Single-/n-shot Prompting: Approaches to prompt responses without examples (zero-shot), with one example (single-shot), and with multiple examples (n-shot).

• Chaining: A technique that connects multiple prompts/results for complex tasks.

• Transformers: Architecture for LLMs leveraging self-attention to process input data.

• Attention: Focus mechanism within Transformers that enhances the processing of input sequence parts.

• Quantization: Reducing a model's precision to lower resource usage while maintaining acceptable performance.

• LLM Evaluation: Methods for assessing the effectiveness of LLM outputs, e.g., ROUGE, BLEU.

• Vector Databases: Specialized databases optimized for storing and querying vector representations.

• Distillation: A process of creating smaller models that retain the knowledge of a larger model.

Learning Outcomes

• Explanation of terms: Large Language Model (LLM), Prompt Engineering, Fine-Tuning

• Design LLM systems using advanced methods and RAG

• Understand LLM concepts and make design choices

• Assess LLM system outputs

GenAI and LLMs

Example Prompt

• Task: Create a report on fine-tuning LLMs

• Response from ChatGPT:

  1. Introduction: LLMs revolutionize natural language processing. Fine-tuning enhances their performance on specific tasks.

  2. Overview of Large Language Models:

  • Definition and Characteristics: Capable of performing multiple NLU tasks such as translation and summarization.

  • Pre-training Phase: Uses transfer learning with transformer architectures and self-attention mechanisms.

  • Notable LLMs: GPT series, BERT, XLNet, T5.

  1. Fine-Tuning LLMs: Overview of methodologies and techniques from literature.

What is a Language Model?

• Definition: Probability model denoted as P(Text | Preceding Text)

• Concept: Models the likelihood of a word following one or more preceding words based on statistical patterns in training data.

Conditional Probability Formula

• Formula: $P(W1,W2,W3,W4, ext{…},Wn) = P(W1) imes P(W2 | W1) imes P(W3 | W1,W2) imes … imes P(Wn | W1,W2, ext{…},W(n-1))$

• Exploration: Captures the sequence's implicit order through conditional probability.

Large Language Models

• Examples of LLMs:

  • OpenAI's GPT

  • DeepMind's Chinchilla

  • Bloomberg GPT

  • Google's Med-Gemini

  • Meta's LLAMA

  • T5/Gemma

  • Mistral

Characteristics of LLMs

• Utilize transformer-based architectures.

• Efficiently manage long-range dependencies by processing sequences of tokens.

Typical LLM Projects in Computer Science

• Text-oriented tasks:

  • Email writing

  • Code generation

  • Document analysis

• Training and Fine-tuning:

  • Utilization of datasets for specialized tasks

• Replacing humans with chatbots: Performing repetitive text generation tasks.

Size of LLMs

• Illustrations of LLM capabilities with performance metrics including human expert levels, measured via various benchmarks such as MMLU.

• Key models and their sizes include:

  • Gopher

  • U-PALM

  • GPT-4 Classic

  • ChatGPT (GPT-3.5 turbo)

  • Llama (65B parameters)

  • Chinchilla and others.

Architectural Designs

• Fine-tuning and Prompt Engineering:

  • Techniques and frameworks for optimizing LLMs for specific applications.

Advanced Prompt Engineering Techniques

• Retrieval Augmented Generation (RAG):

  • Combines prompt engineering with external knowledge sources for more accurate LLM outputs.

  • Components:

    • Prompt embedding

    • Query Data

    • Improved Prompt for the LLM

Example Data (from GPT-4)

• Generate question-answer pairs related to financial risks.

Fine-Tuning vs Quantization

• Fine-Tuning: Modifying a pre-trained LLM for specific tasks using domain-specific data. Typical cost for training models like GPT-4 is approximately $100 million.

• Quantization: Reduces resource requirements by lowering precision levels (e.g., from float32 to float16 or int8).

Distillation

• Process involving training smaller models to emulate larger models' behavior, thereby retaining knowledge while improving efficiency.

• Key roles of Knowledge Distillation (KD) in LLMs:

  1. Enhances capabilities.

  2. Provides model compression.

  3. Encourages self-improvement.

LLMs in Business

Applications in Fintech

• Bloomberg GPT: A large language model tailored for finance with superior performance metrics compared to open models.

• LLM-based Trade Document Analysis: Leveraging LLMs to ensure accuracy and reliability in trade documents.

Applications in Healthcare

• Utilization of LLMs to manage and summarize scientific knowledge and translate complex medical content.

Evaluation Metrics

ROUGE

• An evaluation metric measuring summary quality by overlap with reference texts, utilizing F1 score for N-gram precision and recall.

• Recall and Precision Goals:

  • Recall = accurate n-grams / reference total

  • Precision = quality n-grams / output total

Application Example

• Example: Model Output = 'the house on the hill', Reference = 'walk up the hill to the house'.

  • 1-gram overlap evaluation.

Other Evaluation Metrics

• BLEU, Perplexity, Human evaluation, etc.

Future Lectures

Expectations

• Preparation for exam structure focusing on LLMs, their applications, and integration methods.

• Discussion questions regarding RAG, output comparisons, and potential LLM applications in specific domains such as healthcare and finance.

Bibliography

• Includes multiple referenced works discussing advancements in and applications of LLMs.