chapter two

2 How to do model adaptation

 

This chapter covers

  • How each adaptation technique trades off cost, capability, and engineering effort, with the kind of practical depth a team uses to pick
  • Hybrid stacks and how to layer them
  • Buy versus build: when a managed API is the right answer, and when self-hosting is
  • Infrastructure options, tools, and frameworks, and how to choose a base model
  • A small fine-tuning quickstart that previews the recipe chapter 5 explains in depth
  • Security implications and the operational discipline a deployed model needs

A regional investment bank wanted to build a system that could read quarterly earnings transcripts and surface insights for portfolio managers. The team had solid PyTorch chops, access to a cluster of A100 GPUs, and a clear deadline: have something ready before Q1 reporting season. They chose full fine-tuning on an 8B-parameter model, trained on three years of historical earnings transcripts. Eight weeks and $55K in compute later, the system confidently summarized calls but could not answer the one question that mattered: what was said on yesterday’s call.

2.1 The adaptation continuum at decision depth

2.2 Hybrid approaches: Most production systems stack rungs

2.3 Matching technique to use case

2.4 Buy versus build

2.4.1 Buy: Start with a managed API

2.4.2 Build: Self-host

2.5 Infrastructure options

2.6 Tools and frameworks

2.7 Selecting a base model

2.7.1 Open-weight base models for fine-tuning

2.7.2 Frontier models for reasoning tasks

2.8 A quickstart preview: The shape of a fine-tune in five steps

2.9 Security across the adaptation continuum

2.9.1 How each technique changes the attack surface

2.9.2 Mitigating the risks

2.10 Lightweight operationalization

2.11 Summary

2.12 References