Instructor: Prof. Travis Breaux
Associate Professor
Software and Societal Systems Department
https://www.cs.cmu.edu/~breaux
Syllabus
Introduction. Advances in large language models have created opportunities to adapt natural language processing tasks to new domains without the computational labor of fine-tuning models on tens of thousands of training examples. Instead, users can write prompts with instructions, demonstrations and other context to facilitate in-context learning using a model with frozen parameters.
This course begins with a brief history of language modeling, after which the course covers prompt engineering strategies and techniques. Topics covered include prompt-tuning, chain-of-thought prompting, multi-hop reasoning, multi-stage prompting, among others. Students will learn about prompt engineering benchmark datasets and task goals, evaluation metrics, self-consistency, and calibration to evalaute the efficacy of prompt designs. Finally, the course will cover alignment and the ethics of large language models, while surveying a cross-section of domain-specific applications across law, medicine and computer science.
Learning Objectives. After completing this course, you will be able to:
- Intuit instructional design for shaping answers
- Select and order demonstrations
- Decompose complex tasks into multi-stage prompts
- Design experiments to evalaute prompt designs
Assessments. Students learn more by applying and explaining ideas to others, thus, the course requires the following activities:
- Homework assignments, or individual work to help students focus on important points in the readings and to exercise particular skills
- Project, or group work to allow students to construct larger, more complex systems by combining multiple prompting strategies and techniques
- Quizzes to check your learning and reinforce key concepts
- Final Exam, to demonstrate your cumulative knowledge on practical examples. The final exam will be a take-home, open book, open notes exam that is due one week after the final day of class.
- Class participation, to enrich the discussion with your insight, relevant experience, critical questions, and analysis of the material. The quality of contribution is more important than the quantity.
| Assessment | Final Grade % |
|---|---|
| Individual Assignments | 20% |
| Group assignments | 30% |
| Quizzes | 15% |
| Final exam | 25% |
| Class participation | 10% |
Required Readings. The concepts and ideas taught in this class are largely drawn from emerging research, some of which is only available in pre-publication format. Students will be directed to read papers drawn from over 50 foundational papers in NeurIPS, ACL and JMR, among other sources.
Lectures will cover the readings in-depth with examples drawn from the papers and from other works cited by the readings. Reading ahead allows students to engage in richer discussions during the course and thus this is highly encouraged to get the most out of class time.
See the Course Schedule...