Master of Science (M.Sc.) Semester: 4051 3 Credits

Course Syllabus & Policies

Comprehensive overview of Master's curriculum topics, learning outcomes, grading breakdown, prerequisites, and academic policies.

Course Description

Natural Language Processing (NLP) is at the epicenter of modern artificial intelligence. This graduate Master's course provides an in-depth mathematical and algorithmic foundation of modern NLP, focusing on large-scale language representation learning, Transformer architectures, Parameter-Efficient Fine-Tuning (PEFT/LoRA), Reinforcement Learning from Human Feedback (RLHF & DPO), Retrieval-Augmented Generation (RAG), and Multimodal LLMs.

Key Learning Outcomes

  • Master mathematical foundations of autoregressive language models, word representations, and sequence modeling.
  • Implement encoder-only, decoder-only, and encoder-decoder Transformer blocks from scratch with multi-head attention and RoPE.
  • Understand pretraining scaling laws, compute-optimal recipes, data filtering, and distributed training systems.
  • Apply parameter-efficient fine-tuning techniques (LoRA, QLoRA, prefix tuning) on open-weight LLMs.
  • Implement preference alignment pipelines using RLHF, Direct Preference Optimization (DPO), and rejection sampling.
  • Design industrial-grade Retrieval-Augmented Generation (RAG) pipelines with hybrid dense/sparse vector search.
  • Analyze multimodal language models, decoding strategies, test-time scaling, and safety/fairness considerations.

Grading Policy & Breakdown

Official preliminary grading breakdown (110% Total with 10% bonus capacity)

Total: 110%
Component Weight Description Visual Share
Assignment 1 10% Word Representations, Autoregressive LMs & Tokenization
Assignment 2 10% Transformers & Parameter-Efficient Fine-Tuning (LoRA)
Assignment 3 10% RAG Systems, Reasoning (CoT) & Model Alignment (DPO)
Assignment 4 (Course Project) 20% Course research and implementation project (Guidelines to be announced)
Midterm Exam 20% Comprehensive exam covering Sessions S1 to S17 (Week 10)
Final Exam 20% Final theoretical and applied examination on full semester topics
Quizzes (Q1, Q2, Q3) 15% Three short review quizzes (5% each) at Weeks 5, 8, and 12
In-Class Participation 5% Active discussion, paper reviews, and classroom engagement
Total Available Score 110% Includes 10% built-in bonus potential

Curriculum Structure & Module Alignment

All 11 official curriculum modules are mapped independently to course sessions, aligning with national graduate requirements and benchmarked against CMU ANLP and Stanford CS224N.

Sessions S1, S2, S4

Module 1: Introduction to Large Language Models & Distributed Semantics

History of NLP, distributed word representations (Word2Vec, GloVe), and autoregressive neural language modeling fundamentals.

Sessions S8, S9

Module 2: Pretraining Strategies, Data Scaling & Transfer Learning

Web scraping pipelines, quality filtering, deduplication, Chinchilla compute-optimal scaling laws, and transfer learning paradigms.

Sessions S5, S6, S7

Module 3: Transformer Architecture & Subword Tokenization

Recurrent sequence encoders, scaled dot-product attention, multi-head Transformer blocks, RoPE, and subword tokenization (BPE, WordPiece).

Session S19

Module 4: Security, Privacy, Hallucination & Robustness in NLP

Adversarial prompt injection, jailbreaking attacks, mechanistic interpretability, safety filters, and privacy preservation in LLMs.

Session S18

Module 5: LLM Evaluation, Benchmarks & Leaderboards

Standardized NLP benchmarks (MMLU, GSM8K, HumanEval), LLM-as-a-Judge methodologies, Chatbot Arena Elo ranking, and contamination detection.

Sessions S11, S12

Module 6: Reinforcement Learning & Human Preference Alignment

Markov decision processes, PPO policy gradients, reward modeling, RLHF (InstructGPT), and Direct Preference Optimization (DPO).

Sessions S10, S13, S14

Module 7: Advanced Architectures, Quantization & Parameter-Efficient Fine-Tuning

LoRA, QLoRA, weight quantization (AWQ, GPTQ), distributed Megatron/ZeRO training, and Mixture of Experts (MoE).

Session S16

Module 8: Retrieval-Augmented Generation (RAG) & Dense Search

Dense passage retrieval (DPR, ColBERT), vector indexing (HNSW), chunking, re-ranking, and Self-RAG frameworks.

Sessions S15, S17

Module 9: Reasoning, Chain-of-Thought & Autonomous LLM Agents

Decoding algorithms, Chain-of-Thought prompting, test-time scaling/search, ReAct prompting, tool use, and autonomous agent loops.

Sessions S20, S21

Module 10: Multimodal Models & Diffusion Generation in NLP

Vision-Language Models (CLIP, LLaVA), cross-modal attention, continuous/discrete text diffusion, and flow matching for language generation.

Session S22

Module 11: Graph Neural Networks (GNNs) for Natural Language Processing

Knowledge graph embeddings, Graph Convolutional Networks (GCN/GAT) for text classification, dependency parsing, and KG-LLM fusion.

Course Policies & FAQs

We encourage the use of modern generative AI tools (e.g. ChatGPT, Claude, Copilot) as learning and coding assistants. However, all submitted code and reports must represent your own synthesis and deep understanding. You are required to explicitly cite any significant AI assistance in your assignment submission comments. Plagiarism between student submissions is strictly prohibited and subject to IUST academic committee referral.
Each student is allocated a total of 3 slip days (grace days) across the entire semester for homework assignments, with a maximum of 2 slip days usable on any single assignment. Beyond the slip days, late submissions are penalized at 15% deduction per 24 hours. No submissions will be accepted 4 days after the original deadline.
Students are expected to have a solid background in Linear Algebra, Probability & Statistics, Machine Learning / Deep Learning, and proficiency in Python and PyTorch. Prior knowledge of basic NLP concepts is beneficial but not strictly mandatory.
The 3 quizzes (Q1 in Week 5, Q2 in Week 8, Q3 in Week 12) are designed to consolidate theoretical insights and ensure steady progress. The remaining time of each quiz session is dedicated to interactive Q&A and problem solving with the teaching assistants.