chapter one

1 Why model adaptation?

 

This chapter covers

  • Questions every team asks before adapting an LLM
  • Scenarios where adaptation pays off and the ROI math behind each
  • Enterprise pain points: accuracy gaps, API cost at scale, regulatory constraints, vendor lock-in
  • When NOT to adapt: the cases where prompting alone is the right answer
  • A decision framework for choosing your first technique
  • The base model, dataset, and tools we use for hands-on examples

A demo on Thursday nailed every test case. Leadership pushed for users by Friday. By the following Wednesday, the support queue had caught the first fabrication: a paragraph the model had invented, citing a reimbursement clause that does not exist and recommending escalation to a manager who had left two years ago. The phrasing was plausible. That was the problem. A human reviewer had to know the policy by heart to flag it. By the time the queue caught up with the fabrication, four other employees had already acted on similar invented guidance elsewhere in the assistant.

This is what happens when a general-purpose large language model (LLM) meets an organization for the first time. The fluency and general capability are real. So is the gap: the model does not know your terminology, workflows, or current policy, and has no way to flag when it is guessing. This book is about closing that gap. It is also about knowing when not to try.

1.1 The adaptation problem

1.2 Scenarios where adaptation pays off

1.2.1 Scenario 1: Industry-specific models

1.2.2 Scenario 2: Regional and cultural adaptation

1.2.3 Scenario 3: Enterprise tone and brand

1.3 The numbers: ROI in real terms

1.4 Enterprise pain points

1.5 When NOT to adapt

1.6 How model adaptation works

1.6.1 The adaptation lifecycle

1.6.2 The continuum of techniques

1.6.3 What the gap actually looks like

1.7 Choosing a starting technique

1.8 Considerations: Safety and adaptation risk

1.9 What you’ll need to follow along

1.9.1 The base model: Qwen3-4B-Instruct-2507

1.9.2 The starter dataset: IT support Q&A from Stack Exchange and Dolly

1.9.3 Tooling and hardware

1.9.4 The code repository

1.10 How this book teaches model adaptation

1.11 Summary

1.12 References