Artificial Intelligence

LLM Introduction

Overview

This course provides a practical, end-to-end understanding of Large Language Models (LLMs), from how they are trained to how they are adapted, deployed, and evaluated in real-world scenarios.

Participants begin with the foundations of modern LLM architectures and pre-training methods, then explore the post-training techniques that shape model behavior, including supervised fine-tuning, preference tuning, LoRA and RLHF/DPO.

The course also covers safety and alignment considerations, evaluation strategies, and performance measurement.

Through guided hands-on labs, learners will build and deploy small LLMs, fine-tune models for instruction following, and evaluate outputs across downstream tasks.

Duration

1 Day

Who Should Take This Course

Audience

Developers, Data/DevOps/AIMLOps Engineers, Technical Managers and Analysts.

Prerequisites

Participants should have basic programming skills in Python and some experience with data analysis/science.

Why You Should Take This Course

Upon completing this course, participants will be able to:

  • Understand the technical foundations of Large Language Models and how they are trained
  • Build pre-training pipelines, including dataset sourcing and data cleaning
  • Use various tokenization methods and Transformer-based architectures
  • Perform pre-training on GPT-style models
  • Use post-training techniques such as Supervised Fine-Tuning, PEFT/LoRA and DPO/RLHF to shape model behavior
  • Implement safety and guardrail mechanisms, including refusal behavior, tone control, and persona/system prompt design
  • Evaluate LLMs using traditional performance metrics, task-specific benchmarks, and human preference scoring frameworks

Course Outline

LLM Introduction

1. LLM Overview

  • LLM technical foundations
  • LLM capabilities and limitations
  • Popular LLMs (GPT family, Llama family, Claude, etc.)
  • Use cases and applications
  • LAB: Building an LLM Endpoint

2. Pre-Training

  • Data collection (crawling, Common Crawl)
  • Data cleaning (RefinedWeb, Dolma, FineWeb)
  • Tokenization (Byte Pair Encoding (BPE), SentencePiece and Unigram language models)
  • Architecture (neural networks, Transformers, GPT family, Llama family)
  • Text generation (greedy and beam search, top-k, top-p)
  • LAB: Pretraining a Tiny GPT LLM

3. Post-Training

  • SFT (Supervised Fine-Tuning) and instruction following
  • PEFT (Preference-Enhanced Fine-Tuning) and LoRA (Low-Rank Adaptation)
  • RLHF (Reinforcement Learning from Human Feedback)/DPO (Direct Preference Optimization)
  • Safety & Guardrail tuning
  • System prompts and persona shaping
  • LAB: Fine-Tuning an LLM for Instruction Following

4. Evaluation

  • Traditional metrics
  • Task-specific benchmarks
  • Safety and Alignment Evaluation
  • Human evaluation and leaderboards
  • LAB: Evaluating LLM Performance on Downstream Tasks
Search UMBC Training Centers