Adaptation and Fine-Tuning

Fine-tuning is the option teams reach for when prompting is not enough, and the one they

understand least. The advice around it is unusually confident and unusually unquantified.

So adaptation is attempted, the model gets better at the new task, and whether it got worse

at everything else is a question nobody measures — because measuring it means deliberately

evaluating a capability you were not trying to improve.

That is the design decision behind this book: **every experiment reports two numbers.** The

domain the model was adapted to, and the general text it started from. The second is the one

production teams do not collect, and it is where the entire cost of adaptation appears.

Nothing here is described rather than built. Low-rank adaptation is implemented from scratch —

frozen base weights, a low-rank pair on every attention projection, initialised so the adapted

model is exactly the base model before training begins. Direct preference optimisation is

implemented from scratch, with a frozen reference model. Quantised training, gradient

accumulation, loss masking and multi-task interference are all measured on the same base

model, so the comparisons differ only in the variable under test.

What gets measured:

- **The trade happens immediately, not gradually.** Probed every fifty steps, the first 8% of

training captured essentially the whole domain gain — perplexity from 19,952,720 to 8.18 —

and inflicted most of the forgetting, taking general perplexity from 256 to 5,111. There is

no stopping point at which the domain is adapted and the general ability is intact.

- **LoRA's reputation does not survive measurement — and neither did my first conclusion.**

The initial sweep showed full fine-tuning beating LoRA on both axes. That sweep tied the

scaling constant to the rank, so two variables moved together. Holding rank fixed and

varying only the scale reversed it: at a low scale LoRA retained general ability *better*

than full fine-tuning, while training 1.68% of the parameters. Both sweeps are in the book,

because the mistake is more instructive than the result.

- **Mixing 12.5% general data into each batch** cost 0.6% of the domain gain and returned

general perplexity from 6,245 to 236 — better than the base model's own 256. No method that

merely restricts the update can end above where it started.

- **Preference training costs a different order of magnitude from demonstration.** Teaching by

comparison reached perfect preference accuracy for 2.8–9.1% of general perplexity, where

supervised adaptation cost 2,339%.

- **A second adaptation erases the first.** A model at accuracy 1.000 on one task scored 0.000

after fifty steps on another — erasure rather than decline, because a discrete learned

behaviour has no partial version. Interleaving the two gave 1.000 on both, and better

general text than either order.

- **Data volume is the wrong question.** Eight examples — one per fact — reached 1.000, and

thirty-two times as much taught nothing further. A quarter of the labels could contradict

the rest with no measurable loss; at half, the failure was per-fact rather than uniform,

which spot-checking misses.

- **Quantisation costs the base model, not the adaptation.** Four-bit base weights cost 8.3%

of what the model already knew and left the adapted result unchanged, because the adapter is

trained afterwards in full precision and fits itself to the quantised base.

What you get:

- 10 chapters covering the adapt-or-retrieve decision, full fine-tuning and forgetting,

parameter-efficient methods, mitigations, instruction tuning, preference training, dataset

design, multi-task interference and evaluation, memory-constrained training, and a final

chapter of limits.

- 54 original practice questions, tagged by certification and by difficulty, where every

option is explained — not just why the right answer is right, but why each wrong answer is

wrong, because on these exams the distractors are where the teaching is.

- A reverse map from each certification to the chapters that serve it, in reading order.

- All the code — the transformer, LoRA, DPO, the quantiser — so every table can be regenerated.

Written to the published objectives of NVIDIA NCP-GENL, NVIDIA NCA-GENL, AWS Certified

Generative AI Developer – Professional, Databricks Certified Generative AI Engineer, and

NVIDIA NCP Agentic AI. Objectives were checked in August 2026; confirm the current blueprint

with the certifying body before you sit.

You need to be able to read Python. Transformers, gradients, low-rank factorisation and

perplexity are built up from the point where they are needed, and nothing in this volume

assumes you have read any other book.

Every question is original, written from published exam objectives. Nothing is reproduced

from, or based on recollection of, any live examination.