Materials and Resources

Textbooks, one paper per week, derivations, toolchain

Core Textbooks

Code Source Used for
JM3 Jurafsky & Martin, Speech and Language Processing, 3rd ed. online draft, 2026-08-19 release (https://web.stanford.edu/~jurafsky/slp3/) The teaching baseline for the whole course. This release is reorganized into three volumes and 26 chapters, with Volume I devoted to LLMs
RAS Raschka, Build a Large Language Model (From Scratch), Manning, 2024 (a Traditional Chinese edition exists) The primary text for building components by hand. The labs in W3, W4 and W8 take their skeletons from it
RAS-R Raschka, Build a Reasoning Model (From Scratch), Manning Hand-built material for W9 and W10 — the DPO loss, verifiers, budget forcing
XZ Xiao & Zhu, Foundations of Large Language Models, arXiv:2501.09223 (2025-01, last revised 2025-06) An arXiv monograph, not a published book. Treat it as free lecture notes rather than as a textbook
LLMBook 趙鑫、李軍毅、周昆、唐天一、文繼榮, 大語言模型, 高等教育出版社, 2024 (13 chapters) The Chinese companion. Its site carries lecture PDFs, code and the original slides on request
BB Bishop & Bishop, Deep Learning: Foundations and Concepts, Springer, 2024 Prerequisite mathematics and deep-learning background; free online edition
UDL Prince, Understanding Deep Learning, MIT Press; GitHub v5.0.3, 2026-02-09, ch 1–21, CC BY-NC-ND The best catch-up text for a class with no prerequisites: free PDF, notebooks, slides and an answer booklet, and still maintained
HOLLM Alammar & Grootendorst, Hands-On Large Language Models, O’Reilly, 2024 (12 chapters) A hands-on companion. The visualizations in ch 2 (tokens and embeddings) and ch 8 (RAG) are unusually good. Its notebooks are written for Colab, so the backend has to be swapped on the demo machine
ZZM Zong, Zhao & Ma, Natural Language Processing and Large Language Models: Theory, Hands-on Codes, and Case Studies, Springer Nature Singapore / Tsinghua University Press, 2026 (ISBN 978-981-92-0681-0; open access, 400 pages) The English companion for W1–W3. Ch 5 (n-grams, smoothing, perplexity, FNN- and LSTM-based LMs), ch 3 (word2vec, ELMo) and ch 4 (seq2seq through Transformer) line up almost section by section with W1 and W3; ch 2 is the shortest catch-up chapter for anyone arriving without the prerequisites; ch 7 is the only treatment of Chinese word segmentation on this list. Read it, do not run it — see the note below
LLMSurvey Zhao et al., A Survey of Large Language Models, arXiv:2303.18223 (last revised 2026-03-18, 144 pages, 1081 references, marked ongoing) A reference, not a textbook

Three things to know about JM3

Checked directly against the Stanford page.

  1. It is still a draft with no print edition (the site’s own answer to “when will the book be finished?” is “don’t ask”). Download and keep the 2026-08-19 release, because chapter numbers move between releases and this course cites that one.
  2. The structure of the 2026-08-19 release — the largest improvement since the 2024 version, since pretraining and post-training finally have separate chapters:
    • Vol I — Large Language Models: 1 Introduction (rewritten for this release) / 2 Words and Tokens / 3 N-gram LM / 4 Logistic Regression & Text Classification / 5 Embeddings / 6 Neural Networks / 7 Transformers and Pretraining / 8 Post-training
    • Vol II — Advanced LLM Topics and Tools: 9 Masked LM / 10 Interpretability (incomplete) / 11 Information Retrieval and RAG / 12 Agents “[not written yet]” / 13 MT / 14 RNNs & LSTMs / 15–17 speech
    • Vol III — Annotating Linguistic Structure: 18–26 (POS/NER, parsing, IE, SRL, coreference, discourse, conversation)
  3. W13 has no textbook — ch 12 Agents is unwritten, so that week runs entirely on papers and lecture notes. Ch 10 Interpretability is also incomplete; use it with care in W14.

Two limits on ZZM

  1. The code is on a different stack. Chapters 7–15 are written entirely in PaddlePaddle / PaddleNLP against Baidu’s ERNIE models and AI Studio: the book mentions paddle 474 times and paddlenlp 123 times, and torch and pytorch exactly zero times. This course runs on Hugging Face transformers with Qwen3, so ZZM is reading, not lab material.
  2. Coverage stops around 2023. Ch 6 compresses instruction tuning and RLHF into a section each, and the substance of W5–W14 — RoPE and long context, MoE and state-space models, scaling laws, DPO and GRPO, reasoning and test-time compute, inference efficiency, RAG, agents, evaluation and interpretability — is simply not in the book.

The authors are candid about the application chapters themselves: ch 1.2 says they demonstrate “the implementation process of the task through a specific example, rather than getting a practical system with the best performance.” Read the application chapters in that spirit.

Deliberately not assigned

Goldberg, Neural Network Methods for NLP (2017) and Paaß & Giesselbach, Foundation Models for NLP (2023) have both dated badly, and appear only as historical context in W1. Raaijmakers, Large Language Models (MIT Press, 2025) is conceptual and social rather than technical — not deep enough for a graduate course. Burkov, The Hundred-Page Language Models Book (2025) is too thin to carry the course, but works as pre-reading before week 1.

One Paper per Week

Week Topic If you read only one
W1 From n-grams to seq2seq Bahdanau et al., Neural Machine Translation by Jointly Learning to Align and Translate, ICLR 2015
W2 Tokenization Limisiewicz et al., Compute Optimal Tokenization, arXiv:2605.01188
W3 Attention Vaswani et al., Attention Is All You Need, NeurIPS 2017 — with Jain & Wallace, Attention is not Explanation
W4 The full block Wortsman et al., Small-scale proxies for large-scale Transformer training instabilities, arXiv:2309.14322
W5 Position and long context Liu et al., Lost in the Middle, TACL 2023
W6 Architecture routes Du et al., Kimi Linear, arXiv:2510.26692 — with Dao & Gu, Mamba-2
W7 Pre-training and scaling Hoffmann et al., Chinchilla, arXiv:2203.15556 — with Besiroglu et al.’s replication
W8 SFT and PEFT Shuttleworth et al., LoRA vs Full Fine-tuning: An Illusion of Equivalence, arXiv:2410.21228
W9 Alignment and RL Rafailov et al., DPO, NeurIPS 2023 — with the GRPO section of Shao et al., DeepSeekMath
W10 Reasoning and test-time compute Yue et al., Does RL Really Incentivize Reasoning Capacity Beyond the Base Model?, arXiv:2504.13837
W11 Inference efficiency Leviathan et al., Fast Inference from Transformers via Speculative Decoding, ICML 2023
W12 RAG Li et al., RAG or Long-Context LLMs?, EMNLP 2024 industry track
W13 Agentic systems Cemri et al., Why Do Multi-Agent LLM Systems Fail?, arXiv:2503.13657
W14 Evaluation and interpretability Korznikov et al., Sanity Checks for Sparse Autoencoders, arXiv:2602.14111 — with Miller, Adding Error Bars to Evals

Why these. The table leans deliberately towards critical work and negative results rather than the paper that first proposed each method. The originals are largely in the textbook — JM3 Vol I covers the spine of W1–W9 — while “here is where this method breaks” has no textbook and is worth far more to a graduate student. If you read one paper for a week, read the critical one.

Derivations and Hand-Built Components

One board derivation and one component per week. Everything is demonstrated in class. The last column says what each component needs to run on, not how long it takes: the demo machine has no CUDA, and nothing here has been timed on it yet. Timings will be added once they are measured rather than estimated.

Week At the board Built by hand Runs on
W1 Cross-entropy as compression; the power-law decay of unseen mass under Zipf–Heaps; the BPTT product and its singular-value bound — (demos only) CPU
W2 Vocabulary size inside the scaling law BPE training, encoding, decoding CPU — tokenizers are hardware-independent
W3 The variance of q^\top k1/\sqrt{d_k}; the KV-cache formula Scaled dot-product attention + KV cache CPU — the gradient check needs float64, which Metal does not have
W4 Parameter accounting → N \approx 12 n_{\text{layer}}d^2; C \approx 6ND A full GPT block, and a tiny model trained MPS or CPU. SDPA on MPS refuses attention dropout
W5 RoPE’s relative property in complex form; the phase argument for extrapolation RoPE and its extrapolation variants MPS; the long-context demo uses a quantized mid-size model
W6 The linear-attention associativity identity → recurrent form; the SSD skeleton Both algorithms for linear attention MPS. SSM and hybrid models run on MLX’s own Metal kernels
W7 Chinchilla’s first-order condition (one of the course’s three peaks) A scaling law fitted at small scale MPS — four or five tiny transformers
W8 LoRA’s parameter and memory accounting; the SVD behind intruder dimensions LoRA SFT with mlx_lm.lora MLX. Pointing it at a quantized model is QLoRA
W9 DPO in four steps (one of the three peaks); GRPO’s advantage normalization The DPO loss, with four tests CPU
W10 The two crossing pass@k curves; the binomial bound behind self-consistency Budget forcing MLX, a mid-size model with a thinking mode
W11 Roofline and the batch threshold for decode; the three-line losslessness proof for speculative decoding A roofline measured MLX (mlx_lm.benchmark, --draft-model)
W12 InfoNCE, and how the negatives define “similar”; RAG’s three-term decomposition A minimal RAG, and three deliberate failures CPU
W13 The compound error rate p^n, and what a verifier does to it An agent broken by prompt injection A small local model is enough
W14 The binomial standard error and paired bootstrap; the sparsity–reconstruction Pareto An evaluation workshop (no new model) CPU — TransformerLens avoids MPS by default

Toolchain

Students need no compute for this course. It is lectures throughout, and every demo is run live in class on an Apple-silicon laptop with 64 GB of unified memory and no CUDA — which is what decides most of this table.

Purpose First choice Notes
Inference for in-class demos MLX (mlx-lm), or llama.cpp / Ollama (GGUF) Native to Apple silicon. Unified memory is the real advantage here: models far larger than 16 GB load comfortably
Building components by hand Plain torch (MPS backend, or CPU) The hand-built labs in W3, W4, W6 and W9 are small enough that MPS — or even CPU — finishes them in minutes
Tokenizer experiments tokenizers, sentencepiece, tiktoken Pure CPU, hardware-independent. The W2 fertility table
Fine-tuning demos mlx_lm.lora W8. QLoRA is not a flag: point --model at an already-quantized model and that is QLoRA. Adapters are merged back with mlx_lm.fuse
Evaluation mlx_lm.evaluate W14. The harness itself has no MLX backend — the integration lives on the MLX side and registers itself into lm-eval. The workshop looks at per-item results, not just aggregate scores
RAG sentence-transformers + faiss-cpu, or plain numpy W12. Brute-force search on CPU is enough; no vector database needed
Interpretability transformer-lens (runs on CPU in practice) Activation patching in W14. The library declines to auto-select MPS unless TRANSFORMERLENS_ALLOW_MPS=1, and currently marks no torch version as MPS-safe. CPU is fine at these model sizes

What cannot be demonstrated on this hardware

Not available Why What stands in for it
bitsandbytes 4/8-bit, QLoRA CUDA only mlx_lm.lora against a quantized model. But “you cannot do 4-bit in transformers on a Mac” stopped being true in 2026: transformers now ships a Metal quantization backend built on MLX’s kernels
vLLM’s PagedAttention and continuous batching Not usable on Apple silicon W11 teaches the argument and cites the SOSP 2023 numbers. PagedAttention’s memory argument was always clearest on a whiteboard
FP8 / NVFP4 measurements Apple silicon has no such tensor formats W7 and W11 cite the figures from the Quartet and NVFP4-pretraining reports
FlashAttention, custom Triton kernels CUDA only The IO argument in W3 is a board derivation, not an implementation exercise
auto-gptq, autoawq CUDA mlx-lm’s own AWQ / GPTQ / DWQ / dynamic-quant tools for the “what quantization changes besides perplexity” comparison
float64 — finite-difference gradient checks, ill-conditioned linalg Metal has no double type. Not slow; absent Pin those exercises to the CPU
Attention dropout through scaled_dot_product_attention MPS’s SDPA refuses dropout_p > 0.0, and has no dedicated backward kernel The hand-built version writes its own dropout — which is one more argument for building it by hand
Multi-process accelerate launch, DDP gloo and nccl do not work with mps Single device only. In-class demos never needed it
CPU offload via device_map="auto" MPS requires the whole model to fit in unified memory This is where 64 GB becomes a hard edge: too big is too big, with no automatic fallback

Models used in the demos

The demos are not limited to models that fit a 16 GB card, so the sizes are chosen for what each point needs rather than for what is cheap to run:

Used for Suggested size Why it is worth scaling up
Hand-built components (W3, W4, W6, W9) Qwen3-0.6B / 1.7B Small is the right choice here — the run finishes while the class is still watching
Long-context behaviour (W5) A quantized mid-size model The lost-in-the-middle curve is far more convincing on a real model than on 0.6B
MoE routing (W6) A genuine open-weight MoE No need to simulate with a tiny model plus four experts
Reasoning (W10) A mid-size model with a thinking mode Budget forcing and overthinking only show up on a real reasoning model
The 2026 flagships Read the paper’s table DeepSeek-V4, Kimi K3 and the rest are still out of reach — that has not changed

Four Things This Course Admits It Does Not Know

These gaps are load-bearing rather than embarrassing — each is a project topic, and saying so is the honest graduate-level position.

  • W2 — there is no verifiable published work on Traditional Chinese token efficiency. The fertility comparison has to be demonstrated live.
  • W12 — chunking and query rewriting have no citable representative papers; both are engineering folklore more than literature, so the course substitutes an ablation lab. RAG evaluation has fragmented, and nothing has become the accepted successor to RAGAS.
  • W13 — no textbook (JM3 ch 12 is unwritten), and no canonical design paper for the Model Context Protocol; its academic literature is almost entirely about security.
  • W14 — item response theory for benchmarks is still scattered; the text-side benchmarks for Taiwanese Hokkien and Hakka are effectively absent; and model welfare has no citable academic literature.
回到頂端