PKU Class 2026 Spring: Introduction to Foundation Models

Instructor: Kun Yuan (kunyuan@pku.edu.cn)

Teaching assistants:

• Daibo Li (2501210088@stu.pku.edu.cn)
  
• Yilong Song (2301213059@pku.edu.cn)
  
• Zhoutong Wu (2501111519@stu.pku.edu.cn)
  

Classroom: 6:30pm - 8:30pm Tuesday, 1:00pm - 3:00pm Thursday, 三教506

Office hour: 4pm - 5pm Wednesday, 静园六院220

References

Stanford CS224n: Natural Language Processing with Deep Learning

Projects

Course Projects

Lectrures

Lecture 1: Introduction to LLM

• Introduction to deep learning [Slides1]
  
• Introduction to large language model [Slides2]
• Reading:
  
  • Andrej Karpathy, State of GPT
    
  • Andrej Karpathy, The busy person’s intro to LLM
    

Lecture 2: Machine Learning Basics

• Preliminary [Notes]
  
• Linear regression; Logistic regression; Multi-classification; Neural network [Slides]
• Reading:
  
  • Stanford CS231n, Linear classification
    
  • Stanford CS231n, Neural netowrk part I

Lecture 3: Language Models

• Word embedding; Language models; Recurrent neural networks [Slides]
  
• Seq2Seq; Attention; Transformer [Slides]
• Reading:
  
  • Stanford CS224N: Week 4
    
  • Ilya Sutskever et. al., Sequence to Sequence Learning with Neural Networks
    
  • Ashish Vaswani et. al., Attention Is All You Need
    

Lecture 4: Parameters, Computations, and Memories in Language Models

• Parameters, Computations, and Memory Costs in Transformer [Slides01] [Slides02]

Lecture 5: Popular LLM Models

• Teacher forcing; Pretrain and Finetuning; BERT; GPTs [Slides]
• DeepSeek [Slides]
• Reading:
  
  • J. Devlin et.al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
  • A. Radford et.al., Improving Language Understanding by Generative Pre-Training
  • A. Radford et.al., Language models are unsupervised multitask learners
  • T. B. Brown et.al., Language Models are Few-Shot Learners
  • DeepSeek, DeepSeek-V3 Technical Report

Lecture 6: Gradient descent

• Convex set; Convex functions; Convex problems; Gradient descent [Slides] [Notes]
• Forward-backward propagation [Slides] [Notes]

Lecture 7: Accelerated Gradient Descent

• Momentum gradient descent; Nesterov acceleration; Anderson acceleration [Slides] [Notes]
  

Lecture 8: Stochastic Gradient Descent

• Stochastic gradient descent [Slides] [Notes]
  

Lecture 9: Momentum and Adaptive SGD

• Momentum SGD [Slides] [Notes]
  
• AdaGrad; RMSProp; Adam [Slides]

Lecture 10: Flash Attention

• Memory access cost; Kernal fusion; Flash attention [Slides]

Lecture 11: Mixed-Precision

• FP32; FP16; Mixed-precision training [Slides]

Lecture 12: 3D Parallelism

• Data Parallelism; Pipeline Parallelism; Tensor Parallelism [Slides]

Lecture 13: Agents (by TA)

• From LLMs to Agents [Slides]
• Introduction to Claude Code [Slides]
• Building Agents with Claude Code [Slides]

Lecture 14: Scaling Law

• Scaling law [Slides]
• Reading:
  
  • Jared Kaplan, et. al. Scaling Laws for Neural Language Models
  • Jordan Hoffmann, et. al. Training Compute-Optimal Large Language Models

Lecture 15: Principals in Prompt Engineering

• Pricipals in Prompt Engineering [Slides]
• Chain of Thoughts [Slides]
• Reading:
  
  • OpenAI, Prompt Engineering
  • Jason Wei, et. al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
  • Takeshi Kojima, et. al. Large Language Models are Zero-Shot Reasoners
  • Zhuosheng Zhang, et. al. Automatic Chain of Thought Prompting in Large Language Models
  • Shunyu Yao, et. al. Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Lecture 16: Data Preparation

• Data source; Deduplication; Quality filtering; Sensitive information reduction; Data composition; Data curriculum [Slides]
• Reading:
  
  • Penedo, Guilherme, et al., The Refined Web Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only
  • Soldaini, Luca, et al., Dolma: An Open Corpus of Three Trillion Tokens for Language Model Pretraining Research
  • Kandpal, Nikhil, Eric Wallace, and Colin Raffel., Deduplicating Training Data Mitigates Privacy Risks in Language Models
  • Xie, Sang Michael, et al., Data Selection for Language Models via Importance Resampling
  • Chen, Mayee, et al., Skill-it! A data-driven skills framework for understanding and training language models

Lecture 17: Parameter-Efficient Fine-Tuning

• Low-Rank adaptation (LoRA); LoRA+; DoRA; LISA; BAdam [Slides]
• Reading:
  
  • E. Hu, et. al., LoRA: Low-Rank Adaptation of Large Language Models
  • S. Hayou, et. al., LoRA+: Efficient Low Rank Adaptation of Large Models
  • S. Liu, et. al., DoRA: Weight-Decomposed Low-Rank Adaptation
  • R. Pan, et.al., LISA: Layerwise Importance Sampling for Memory-Efficient Large Language Model Fine-Tuning
  • Q. Luo, et.al., BAdam: A memory efficient full parameter optimization method for large language models

Lecture 18: Inference

• KV Cache; MLA; H2O; Streaming LLM; Quest; Speculative decoding; Page Attention [Slides]
• Reading:
  
  • J. Ainslie et al., GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
  • DeepSeek Team, DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
  • Z. Zhang et al., H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models
  • G. Xiao et al., Efficient Streaming Language Models with Attention Sinks
  • J. Tang et al., Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference
  • M. Stern et al., Blockwise Parallel Decoding for Deep Autoregressive Models
  • W. Kwon et al., Efficient Memory Management for Large Language Model Serving with Paged Attention

Lecture 19: Alignment

• SFT and RLHF [Slides]
• Proximal policy optimization (PPO) (Slides are adapted from Hung-yi Lee’s lecture at Bilibili)
• Reasoning Large Language Models; DeepSeek-R1 (Slides are adapted from Hung-yi Lee’s lecture at Bilibili)
• Reading:
  
  • Andrej Karpathy, State of GPT
  • L. Ouyang, et. al., Training language models to follow instructions with human feedback
  • Jan Leike, Aligning language models with humans
  • Devon Wood-Thomas, AI Alignment and LLMs
  • DeepSeek Team, DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

Lecture 20: Guest Lecture

• Leheng Chen: Harness Engineering [Slides]