Skip to content

Roadmap

This book is at version 0.1 — a living, ever-evolving resource that grows alongside the field. The five chapters below have been class-tested as course assignments; additional chapters are planned as the open textbook is expanded under the OU Alternative Textbook Grant 2025–2026.

Available now

The five class-tested chapters are delivered as interactive notebooks. Although the course runs five assignments, each notebook spans several subtopics, so the chapters cover considerably more than five discrete topics. The "Inside this chapter" column lists the subsections found in each assignment notebook.

Part Chapter Inside this chapter Status
I — Language Model Foundations 1. Tokenization & Vocabulary
Supplements: Setup Tutorial, AI Usage Example
Part 1 — Analyze the Llama-3.1 vocabulary
  • Setup: Python environment, VSCode, the OSCER cluster
  • Load the Llama-3 tokenizer locally; test it on English, emojis, and combining characters
  • Explore the vocabulary with Linux command-line tools
  • Study your own open research question on the vocabulary
Part 2 — N-gram language model
  • Input files: training, test, seeds
  • Task 1 — compute sentence probability
  • Task 2 — build a language generator
Statement of AI usage
Setup Tutorial: Conda vs. Mamba vs. Venv; UV / Poetry / pip / Conda / virtualenv; Conda on Colab; intro to Jupyter; Pythonic Python
Available
II — Machine Learning Foundation and NLP Tasks 2. Text Classification Section 1 — Dataset exploration
  • Inspect the data; positive/negative class balance
  • Train/dev split; evaluation metrics
Section 2 — Logistic regression
  • Random-chance baseline classifier
  • Preprocessing and feature extraction
  • Train a binary logistic-regression classifier
Section 3 — Multinomial logistic regression
  • Exploratory data analysis on SST-5
  • Build a multiclass sentiment classifier from scratch
Available
III — Representations & Sequences 3. Embeddings & Sequence Models Part 1 — Word embedding
  • History: symbolic vs. neural vs. neuro-symbolic (WordNet → ImageNet; SHRDLU → ChatGPT)
  • The distributional hypothesis
  • Train embeddings with a neural n-gram LM: training, batch training, evaluation, n-gram probability, text generation
  • Further exploration on OSCER
Part 2 — Deep networks for sentiment
  • Deep Averaging Networks
  • PyTorch LSTM
  • Train a word embedding with Gensim
Available
IV — Transformers & LLMs 4. Transformers & LLMs
Supplement: Transformer Illustrated
Part 1 — Machine translation
  • Setup and data exploration
  • Seq2Seq model: encoder, decoder, training setup
  • Context vector at every decoding step
  • Attention-based context vector (attention decoder)
  • Transformer-based machine translation
Part 2 — Transformers & pretrained LMs
  • Transformer-based sentiment classifier
  • HuggingFace Trainer for BERT
  • Sentiment via LLM prompting — unconstrained, chat format with Pydantic & Outlines, with explanations
Transformer Illustrated: EncoderDecoder & Generator, encoder/decoder stacks, self- and multi-head attention, decoder masking, encoder–decoder attention, linear + softmax, feed-forward sub-layer, positional encoding, learnable embeddings, full model
Available
V — LLM Alignment and Agentic AI 5. Fine-Tuning with LoRA/QLoRA & Agentic NLP Part 1 — Supervised fine-tuning (SFT) with TRL (free Colab)
  • Key concepts; install dependencies; log in to Hugging Face
  • Load the dataset; configure LoRA/QLoRA
  • Train, save, and run inference
  • Push the merged model
Part 2 — Agentic NLP: an exam agent
  • Build an LLM agent with LangGraph (ReAct: act / observe / reason)
  • Design tool functions so the agent solves the quiz correctly
  • Build a second agent to grade the answers automatically
Available

Planned chapters

These topics are candidates for future chapters. The list is tentative and will evolve.

  • Evaluation of NLP & LLM systems — metrics, benchmarks, human evaluation, LLM-as-judge.
  • Post-training & reinforcement learning — RLHF, reward models, PPO, DPO, and RLAIF.
  • Self-evolving LLMs — self-improvement, self-training, and self-correction loops.
  • Responsible NLP — bias, fairness, safety, and limitations.

Suggestions welcome

If you are using this book and would like to see a particular topic covered, please reach out to the author (see the Preface).