Table of Contents
🌏 中文版
Berkeley CS288 Spring 2026 is a graduate-level Advanced Natural Language Processing course. It does not begin with an LLM API. It builds from n-grams, word representations, classification, and sequence models before reaching Transformers, pre-training, post-training, RAG, inference-time compute, reasoning, and agents.
This series is a guide to the materials, not a reconstruction of the classroom. The official site says that recordings require a Berkeley login. Anonymous readers get 18 slide units, three assignments, two starter repositories, and project specifications. The articles stay within that evidence.
The prerequisites are operational
The course assumes machine-learning experience and proficiency with PyTorch, NumPy, and neural networks; it provides no introductory tutorials. For undergraduate and master's students, CS182, CS188, CS189, or EECS183/283A is strongly encouraged.
Try this tonight: implement a PyTorch classifier with a forward pass, loss, backward pass, optimizer step, and a defensible train/validation split. If that is not yet routine, repair the foundation first.
Six articles cover all 18 slide units
| order | scope | public slides |
|---|---|---|
| 1 | Counts to classification | 01–04: introduction, n-gram LM, word representation, text classification |
| 2 | Sequences to Transformers | 05–07: sequence models, seq2seq, Transformers |
| 3 | Making trained models useful | 08–12: pre-training, advanced pre-training, post-training, generation, evaluation |
| 4 | External knowledge and architecture | 13–14: retrieval/RAG and advanced architectures, plus the social-impact materials gap |
| 5 | Reasoning and agents | 15–18: embodied perception, inference-time compute, agent reasoning, embodied agents |
The schedule also lists impact, speech, continual learning, and guest lectures. Where no anonymous materials accompany a session, this series records the gap instead of inventing its contents.
The assignments form one capability chain
Assignment 1 covers an n-gram LM, a neural n-gram LM, a perceptron, and an NBOW MLP. It forces a comparison between designed features and learned representations before large models enter the picture.
Assignment 2 turns Transformer vocabulary back into components: BPE, RoPE, causal attention, blocks, optimizers, warmup, cosine decay, and gradient clipping. A bonus mini-study connects pre-training, fine-tuning, and prompting.
Assignment 3 has no starter code. Students collect Berkeley EECS pages, create QA validation data, build a retrieval corpus, compare sparse and dense retrieval, and use ablations to separate retrieval failures from generation failures.
A3 is not the enrolled experience
The materials and starter code are sufficient for meaningful work, hence A3. Course Info says recordings are available to enrolled students and Cal-affiliated auditors; Ed, Gradescope, and staff feedback also belong to the enrolled environment.
Official assessment for Assignments 2 and 3 additionally relies on hidden tests, Gradescope, and a course-provided OpenRouter wrapper. Off-campus readers do not receive those pieces. The final project adds in-course team matching and checkpoint feedback. The course home only acknowledges VESSL AI and Google Cloud project compute credits; it does not promise public access to them.
A2's Transformer and A3's RAG can both incur compute costs. Start with small data, small models, and CPU baselines before renting a GPU.
References
Loading...