In this two-week crash course, participants will gain a baseline understanding of the vocabulary, concepts, and methods involved in current interpretability work. Our first week overviews the basic components of generative language models like GPT-2. The second week ferries participants through a variety of approaches that researchers use to explain model behavior, ending with a day on multimodality. Throughout, our focus will be on the model as an object in its own right. Instead of using a model to analyze some external dataset, our hands-on sessions select various facets of the model itself (embeddings, outputs, etc.) and build explanations on that basis.
Syllabus¶
The course runs Monday through Thursday, 10am-1pm EST from July 20-30.
| Day | Date | Topic |
|---|---|---|
| 1 | M (7/20) | Setup and overview |
| 2 | T (7/21) | Tokenization and embeddings |
| 3 | W (7/22) | Embeddings and attention |
| 4 | R (7/23) | Sampling and generation |
| 5 | M (7/27) | Text as data |
| 6 | T (7/28) | Attribution and feature importance |
| 7 | W (7/29) | Meta-modeling |
| 8 | R (7/30) | TBD: Participants decide |
Prerequisites¶
This course is intended to be a general introduction and makes no assumptions about participant backgrounds (familiarity with programming, experience in natural language processing, etc.). It will walk participants through everything they need to know to complete each session.