# Changelog

## [2025.4] - 2026-01-24
- Improved documentation and support of PowerPoint version of the lectures. 

## [2025.3] - 2025-12-15
- Added PPTX version of all the lectures. The conversion was not verified.

## [2025.2] - 2025-10-31

- Added a slide to the neural networks deck (#3) about GPU speeds up not always showing
- Several small tune-ups to the assignments following issues that came up in 2025
- Add updated leaderboard scripts

## [2025.1] - 2025

This is not a complete list of changes, but I tried to track changes as much as possible.

- Removed "beta" marker from version
- Lectures
  - Added version info to title slide for each deck
  - 01 - Introduction
    - Minor updates
    - Procedural clarifications specific to the Cornell instance of the course
  - 02 - data basics and perceptron:
    - Discuss complexity and implementation
  - 03 - neural networks
    - Discuss trade-offs of vocab size
    - Examples for vocab size
    - Sanity checks: fit an example, fit a minibatch, then train
    - Look at initial loss
    - Stopping criteria with patience — show patience pseudocode
    - Tracking stats during learning: validation loss, training loss — and how they should look like
    - Padding for batching with MLP
    - Discuss trade-offs of max length
  - 04 - word embeddings
    - Refresh word senses visuals
    - Add word2vec pseudocode
  - 05 - language models
    - Discuss implementation considerations
    - Add in-class Shannon Game
    - Discuss smoothing trade-offs
    - Improve discussion of WWII code breaking
  - 06 - tokenization
    - Revise BPE pseudoode to explicitly target vocabulary size
    - Add BPE in-class worksheet
  - 07 - neural lms and transformers
    - Add slide on dropout in transfromers
    - Visualize attention mask
    - Add slide on computing multiple attention head with a single projection
    - Small fixes to Transformer architecture to be in line with original design
  - 08 - decoding llms
    - Add explicit sampling pseudocode
  - 09 scaling up to llms
    - Slides about concrete parameter counts in Transformers
    - Examples from model families
    - Discuss stages of LLM training (details later on)
  - 10 masked lms
    - Refined Transformer architecture presentation
    - Add BERT usage best practices
  - 11 encdec pretrain
    - Improve Transformer architecture presentation
    - Moved enc-dec instruction to here
  - 13 tasks
    - Refine introduction of the task framework
    - Add BERT best practices reminder
    - Add formal notation to presentation of modeling instantiations of tasks
    - Moved enc-dec architecture to earlier lecture
  - 14 post-training llms
    - Rename aligning -> post-training
    - Add iterative post-training
    - Add targeted LLM fine-tuning
    - Minor tweaks throughout
  - 15 prompting
    - Minor tweaks
- Assignments
  - Adopt suffix notation of tensors throughout in starter code
  - Significant improvement throughout -- list below is under-specified
  - Assignment 1
    - Improve starter code
    - Add milestone
    - Clarify expected runtimes
  - Assignment 2
    - Split Assignment 3 to create a word2vec-only A2
    - Create report template and assignment PDF
    - Create starter code and unit tests
  - Assignment 3
    - Rename Spring 2024 Assignment 2 to Assignment 3
    - Starter code improvements
    - Switch to templated report
    - Small improvements in assignment PDF
    - Create unit tests and improve starter code
  - Assignment 4
    - Move the second part of Spring 2024 Assignment 3 to 4
    - Revise the assignment PDF to not include parts moved to A2
    - Revise the assignment report template
    - Improve starter code
  - Assignment 5
    - Move Spring 2024 Assignment 4 to 5
    - Improve starter code

## [2024.1beta] - 2025-02-24

- Added a changelog file.
- Marking version of initial release from Spring 2024 class.