Large Language Models: Training, Adaptation and Evaluation

Pre-Training Data and Objectives

Web-scale data, cleaning, deduplication and next-token prediction

Lab: build a small data pipeline

Scaling Laws and Compute

How performance grows with data, parameters and compute

Interactive: scaling-law curves

Training a Small Language Model

Training a tiny GPT end to end

Code: train a character-level GPT

Instruction Tuning

Supervised fine-tuning on instructions and conversations

Code: instruction-tune a small model

Learning from Human Feedback

RLHF, reward models, DPO and constitutional approaches

Interactive: the preference pipeline

Reasoning Models

Chain of thought, test-time compute and reinforcement learning for reasoning

Lab: compare reasoning strategies

Choosing an Approach

Prompting, retrieval, fine-tuning or training from scratch

Decision exercise

Parameter-Efficient Fine-Tuning

LoRA, QLoRA and adapters

Code: LoRA fine-tuning

W′ = W + BA

Quantisation and Distillation

Making models smaller and faster

Code: quantise and compare

Inference and Serving

KV cache, batching, speculative decoding and serving engines

Lab: serve an open model

Long Context and Memory

Context windows and strategies for long documents

Lab: long-document question answering

Evaluating LLMs

Benchmarks, contamination, LLM-as-judge and human evaluation

Code: build an evaluation suite

Multimodal LLMs

Models that read images, audio and text together

Lab: a document question-answering demo

Open and Closed Models

Licences, trade-offs, cost and model selection

Comparison exercise

Module Review and Assessment

Coding tasks and a scored quiz

Graded project