Skip to main contentSkip to main navigationSkip to footer content

Foundations of Large Language Models: Tools, Techniques, and Applications



 

 

Reserve your spot | Seats are limited
Register now Learn more
About this course You may also be interested in
ChatGPT and the LLM Revolution
Four weeksFive hours per week

There was an error: {{ status.errorMessage }}
OnlineWith live instruction and Q&A sessions
$1,450.00 + applicable taxesView available early bird and alumni discounts

Technologies like OpenAI's ChatGPT and Google's Bard are changing the way we work and generative AI will become increasingly popular. As data professionals and developers, it is important to know how these tools work under the hood and how you can leverage large language models (LLMs) for your work.

LLMs have revolutionized the field of natural language processing (NLP) and are increasingly being used to solve a wide range of NLP problems in various industries. Understanding LLMs can help developers and data scientists, like you, to:

  • Build better NLP models: LLMs are state-of-the-art models for many NLP tasks, and understanding how they work can help developers and data scientists to build better models and achieve better performance on their NLP tasks.
  • Develop custom NLP applications: LLMs can be fine-tuned to specific NLP tasks, making them highly adaptable to different domains and use cases. Developers and data scientists who understand LLMs can leverage this flexibility to develop custom NLP applications for their specific needs.
  • Optimize model performance: Understanding LLMs can help developers and data scientists to optimize model performance by selecting the appropriate architecture, prompt engineering, fine-tuning strategies, and downstream tasks for their specific use case.

With the most recent release of OpenAI's GPT-4 language model, it is being used by Morgan Stanley wealth management to organize its vast knowledge base, Be My Eyes to transform visual accessibility, Stripe to streamline user experinece and combat fraud, and the Government of Iceland to preserve its language. 

This course will provide you with a comprehensive understanding of the latest techniques, tools, and applications of LLMs so you can build applications or processes and further improve your effectiveness and efficiency when working with large language models.

 

  • Software engineers and developers
  • Data analysts and data scientists
  • Machine learning and artificial intelligence engineers
  • Programmers and developers looking to apply NLP, machine learning, and prompt engineering to their stack
  • How to apply and integrate various LLMs to an organization’s existing data infrastructure and systems. This includes understanding open-source alternatives to ChatGPT, Sydney, and Bard.
  • Understanding the evolution of transformer architectures and the historical context behind ChatGPT, including different types of LLMs and their lifecycles (pre-training, fine-tuning, and inference).
  • Familiarity with various machine learning paradigms, including unsupervised, supervised, self-supervised, and in-context learning.
  • Knowledge of different downstream tasks that LLMs can be applied to, such as prediction, extraction, sequence labeling, sequence transformation, and generation.
  • Ability to perform prompt engineering and effective fine-tuning, including prompt construction, effective completions, and understanding the tradeoffs between zero-shot, k-shot, domain/knowledge transfer, in-context learning, and supervised fine-tuning.
  • Understanding LLMs as components in larger architectures, including their use in embeddings for dense retrieval, recommendations, clustering, synthetic data generation, negative mining, and managing model size through knowledge distillation, pruning, and quantization.
  • Weekly live sessions via Zoom every Tuesday from 12 - 1:30 p.m. ET (also available as recordings). Sessions include live instruction with Q&A.
  • Hands-on assignments and weekly quizzes to independently practice and hone skills with programming notebooks.
  • Curated readings designed to expand your knowledge about LLMs.
  • A complete system that uses both ChatGPT and a search engine to answer questions.
  • Attendance requirement: It is highly recommended that participants attend live sessions, but it is not required. Live sessions will be recorded and available for later viewing.

Prerequisites

  • Proficient in reading and writing code in Python. (understanding of different data types and the basics of object-oriented programming).
  • Intermediate to advanced experience with data and machine learning Python libraries such as Numpy, Scikit-Learn, and Pandas.
  • Comfortable working with web applications and big data.
  • Experience working with API endpoints and the ability to write a simple Python code to send requests and parse responses from an endpoint.

Course requirements

 

Receive a certificate from the University of Waterloo

Upon successful completion of this program, you will receive a professional education certificate from the University of Waterloo.

Module 1: Basics of LLMs
  • Understand the evolution of transformer architectures and the historical context behind ChatGPT: Decoder-only models (e.g., GPT-4), encoder-only models (e.g., BERT), encoder-decoder models (e.g., T5).
  • Understand the use cases for working with different machine learning paradigms (supervised, self-supervised, in-context learning).
  • Explain the lifecycle of LLMs: pre-training, fine-tuning, and inference.
  • Know and work with different types of downstream tasks: Text classification, text similarity, search, question-answering, summarization, translation, and named entity recognition.

Module 2: Prompt engineering and fine-tuning
  • Construct prompts and effective completions with OpenAI APIs.
  • Understand the use cases for working with zero-shot and few-shot in-context learning approaches.
  • Fine-tune and evaluate models in with the Hugging Face Transformers library.

Module 3: Picking the right tool for the task
  • Understand the tradeoffs between zero-shot, k-shot, domain/knowledge transfer, in-context learning, and supervised fine-tuning.
  • Choose the right architecture for the downstream task.
  • Understand how to perform knowledge transfer with out-of-distribution datasets.
  • Know and work with custom models with further pre-training.
  • Know how to adapt case studies to new problems involving multilingual retrieval, QA, summarization, sequence labeling, machine translation, and other tasks.

Module 4: LLMs as components in larger architectures
  • Understand how to use embeddings for dense retrieval, recommendations, and clustering.
  • Explain synthetic data generation and negative mining pipelines.
  • Know how to reduce model size via knowledge distillation and quantization.
Professional headshot of Brian Zimmerman

Brian Zimmerman

Course Instructor, WatSPEED

Brian is a current PhD student at the David R. Cheriton School of Computer Science at the University of Waterloo. He is doing research in controllable text generation and techniques to more efficiently train large language models. Brian also spent several years as a software engineer based out of San Diego, California, working on creating an intelligent search interface for government contracts.

Headshot of Rodrigo Nogueira

Rodrigo Nogueira, PhD

Course Author, WatSPEED

Rodrigo Nogueira was a postdoctoral fellow at the University of Waterloo in Canada under the guidance of professor Jimmy Lin. He holds a Ph.D. in Computer Science from New York University, where he was advised by the renowned Professor Kyunghyun Cho.

He is also a computer scientist who serves as an adjunct professor at UNICAMP in Brazil and consults for multiple companies in the search engine industry. He was a pioneer in the use of transformers for search, including monoT5, doc2query, and InPars, and is a co-author of the book "Pre-trained Transformers for Text Ranking."

Throughout his career, Rodrigo has won several IR competitions and made contributions to the IR and NLP fields through the creation and open sourcing of multilingual models, such as BERTimbau and PTT5, and datasets, such as mMARCO.

Headshot of Dr. Jimmy Lin

Dr. Jimmy Lin

Cheriton Chair, Cheriton School of Computer Science | Co-director, Waterloo.AI | Course Author, WatSPEED

Dr. Jimmy Lin is a professor and holds the David R. Cheriton Chair in the David R. Cheriton School of Computer Science at the University of Waterloo. He also serves as the co-director of the Waterloo AI Institute, which has the mission to promote cross-disciplinary research at the frontiers of artificial intelligence and its applications across the entire campus.

Dr. Lin’s area of research lies at the intersection between natural language processing and information retrieval. In addition to being one of the most cited artificial intelligence scholars in the world, he has been frequently and deeply engaged with both the private and public sectors throughout his career, including an extended sabbatical at Twitter and visiting positions at the US National Library of Medicine (NLM), part of the National Institutes of Health (NIH).

He presently serves as the chief technology officer of Primal, a Waterloo-based AI company focused on creating meaning that computers can understand. Dr. Lin holds a Ph.D. in Electrical Engineering and Computer Science from MIT and is a fellow of the ACM.

Questions? Let's chat!

Office hours: Monday to Friday, 8:30 a.m. - 4:30 p.m. ET

  +1 (519) 888-4773

  watspeed@uwaterloo.ca

Hear from our learners

As an aging software developer, the Foundations of Large Language Models course was able to provide me with an understanding of how LLMs actually work and provide a new edge in the market. The elegance of applied mathematics truly continues to change mankind for the better, and I am thankful for the clarity into these amazing systems.

Samuel H. | MindModel AI Inc.
Foundations of Large Language Models Course

 

Register now