NLU-11: Week 5

Welcome to Week 5! This week we will continue studying LLMs, and discuss their scaling laws, connecting to automata and issues with security and safety. We also provide the tutorial solutions below and a new lab. Wishing you a productive week of learning!

Please note that Gradescope now accepts submissions of assignment 1 (code and text are separate).

Lab 2: Word Embeddings - in the labs page.

Tutorial solutions

The tutorial solutions are now available on the tutorials page.

Lectures

• Lecture 1: Scaling laws for LLMs [pdf]

      Required reading:

1. Scaling Laws for Neural Language Models, Kaplan et al. 2020 (Sections 1-2)
2. Training Compute-Optimal Large Language Models, Hoffmann et al. 2022 (Chinchilla scaling laws; Sections 1, 3)

• Lecture 2: LLMs as automata [pdf]

  Background material:

  ACL 2024 tutorial by Butoi et al. - You do not have to go over the tutorial in detail

   

• Lecture 3: Safety and security with LLMs [pdf]

      Required reading:

1. A Watermark for Large Language Models, Kirchenbauer et al., 2023 (sections 1-3)
2. Universal and Transferable Adversarial Attacks on Aligned Language Models, Zou et al., 2023 (sections 1-2)

     Background reading (optional):

1. Gold Doesn't Always Glitter: Spectral Removal of Linear and Nonlinear Guarded Attribute Information, Shao et al. 2023 (sections 1-3)
2. Erasure of Unaligned Attributes from Neural Representations, Shao et al. 2023 (sections 1-2, 3.1, 3.2)