LLM Customization and Fine-Tuning: Adaptation, distillation, and alignment

Get the eBook free when you register your print book at Manning.

LLM Customization and Fine-Tuning is a hands-on playbook for turning a general-purpose open-weights model into a focused, cost-efficient system that’s tailored to your business. You’ll explore the complete adaptation spectrum, from prompting and RAG, through LoRA and QLoRA, to fully supervised fine-tuning, knowledge distillation, and preference alignment with DPO. One running example, a fictitious enterprise and its IT help desk, carries through every chapter, so you see the same problem solved at each step and can compare the techniques head to head. You’ll soon be swapping generic LLMs for ones that know your business, respect your budget, run on your infrastructure, and stay reliable in production!

Everything you learn is fully reproducible on cost-effective and easy-to-access hardware. You’ll train LoRA and QLoRA adapters on a modest consumer card and perform full fine-tuning on a single 24 GB card such as an A30 or RTX 4090. The published numbers in the book match what you’ll see on your own machine, within run-to-run variance. All with no requirement for a datacenter. The book is also honest about the tradeoffs: it shows you where a technique doesn't win, and it publishes all the code, the trained models, and the training and evaluation logs, including the runs that didn't work, so you can verify every result yourself.

As you go, you’ll build a decision framework that weighs cost, latency, privacy, and ROI to choose the right technique for each problem. You’ll construct a training-data pipeline that curates real data, generates teacher-model outputs, and tracks lineage. You’ll distill a smaller, cheaper student from a stronger teacher, and align a model with DPO while running safety regressions at every step. And because most fine-tuned models fail not at launch but months later, you’ll implement the operational layer that other LLM books skip entirely: a model and data registry, TF-IDF drift detection with canary prompts, rollback procedures, a red-team safety monitor, and an outcome-based retraining cadence.

The methods you learn in this book scale unchanged from Qwen3-4B on your workstation to frontier models on a cluster. You’ll build your own intuition around training AI, own your own pipeline, and make build-versus-buy decisions with real numbers.

What's inside

• A framework for choosing between prompting, RAG, LoRA/QLoRA, SFT, distillation, and DPO
• End-to-end LoRA and QLoRA fine-tuning on a single GPU
• Building a training-data pipeline with quality gates and lineage tracking
• Distilling smaller student models and aligning them with DPO
• Production ops: drift detection, canary prompts, rollback, and safety monitoring

About the reader

For ML engineers, data scientists, and MLOps practitioners who need to adapt open-weights LLMs for specific enterprise use cases and run them reliably in production.

About the author

Amit Bahree is CTO in the office of the CEO at G42 Americas, where he leads engineering and AI platform strategy. Previously, he led Applied AI Engineering in Core AI at Microsoft, covering Azure OpenAI, Azure AI Services, and agentic systems across training, inference, and customization at cloud scale. He holds multiple patents, pursued applied software-engineering research at the University of Oxford, and is the author of several books on AI and software, including Generative AI in Action and Practical Weak Supervision.

Weehyong Tok is Partner, Director of Product Management at Microsoft, with over 18 years of experience driving product strategy for Microsoft’s data and AI platforms. He previously served as Head of AI Labs, leading global AI strategy and innovation. He holds a PhD in Computer Science from the National University of Singapore, specializing in large-scale database and data streaming systems and AI in query processing, and he is the author of more than 10 books on AI and data, including Practical Weak Supervision and Practical Automated Machine Learning.