Applied GenAI Curriculum for AI PMs  ·  Depth Layer  ·  Architecture Thread

Unit 19 — Finetuning: What It Actually Involves

The finetuning landscape at a conceptual level — enough to interrogate a proposal's data requirements and maintenance burden, and make the finetune / don't-finetune call as a business decision.

Tags
architecturedata-cost
Layer
DEPTH — when someone seriously proposes finetuning
Objective
Make the finetune / don't-finetune call a business decision — informed but not implemented by you.
Depends on
Unit 05 (the four-way decision; "RAG for facts, finetuning for form"). Pull in when finetuning is seriously proposed.

Why this is a decision unit, not a how-to

Unit 05 gave you the reflex to reach for prompting and RAG before finetuning, and Unit 11 covered the data it demands. This unit is for the moment a team seriously proposes changing the model's weights anyway. You won't run the training. Your job is to interrogate the proposal — what data, how much, who maintains it — and make the call as a business decision. The good news: deciding whether to finetune is the hard part; the mechanics are comparatively straightforward.

1. What finetuning actually is

Finetuning is transfer learning: take a base model that has some of what you need and refine it for your task. It works because the model already learned a huge amount during pre-training, so finetuning transfers that knowledge — the way knowing piano speeds up learning another instrument.

Two ideas that reframe the whole thing Sample efficiency: training a legal-QA model from scratch might need millions of examples; finetuning a good base model might need a few hundred. Unlocking, not teaching: the most useful lens is that finetuning often just surfaces capabilities the model already has but that are hard to reach through prompting alone. If the base model genuinely can't do the task at all, finetuning is a much steeper climb than if it can almost do it.

Any training after pre-training counts as finetuning, so it takes several forms. The ones an application team uses:

Supervised finetuningTrain on labeled (instruction, response) pairs so the model learns to produce the outputs you want. The workhorse.
Preference finetuningTrain on comparisons (instruction, better response, worse response) to align outputs with human preference.
Continued pre-trainingSelf-supervised finetuning on cheap, unlabeled domain text (raw legal docs, a big Vietnamese corpus) before the expensive labeled step — a cost-saving warm-up.

2. The three choices to actually run it

Once the decision is made, execution comes down to picking three things:

Base model

Same selection criteria as Unit 07 (size, license, benchmarks). Start with the strongest you can afford to prove feasibility, then work down.

Method

Lightweight adapters (e.g. LoRA/PEFT) vs. full finetuning — a data-and-serving trade (below).

Framework

A finetuning API (upload data, get a model) vs. a self-run framework — a control-vs-convenience trade (below).

Method: lightweight adapters vs. full finetuning

You don't need the math, just the trade. Adapter methods (LoRA and other PEFT techniques) train a small add-on rather than the whole model: cheaper, work with small datasets (a few hundred examples), run on modest hardware, and — importantly — let you serve many finetuned variants on one shared base model. Full finetuning updates everything: best ceiling on performance, but wants thousands+ of examples and means serving a separate full model each time.

Adapter (LoRA / PEFT)Full finetuning
Data neededSmall — hundreds can workLarge — thousands and up
PerformanceGood, usually below fullHighest ceiling
ServingMany variants on one base modelA full model per variant
WhenStart here; small dataLater, with lots of data

Practical rule: start with an adapter method, move to full finetuning later if needed. With only a few hundred examples, full finetuning often won't even beat LoRA.

Framework: API vs. run-it-yourself

Finetuning APIUpload data, pick a base model, get a finetuned model back. Quick and easy — but limited to supported base models and the knobs the API exposes.
Self-run frameworkMore flexibility and control (adapter or full), but you provision the compute — a mid-tier GPU suffices for adapters; full finetuning at scale needs distributed training.

3. Two development paths worth recognizing

When engineers describe how they'll approach it, it usually maps to one of these:

Progression path

  1. Test the code on the cheapest, fastest model.
  2. Test the data on a mid model (loss should drop with more data).
  3. Push performance with the best model.
  4. Run all models to map the price/performance frontier and pick.

Distillation path

  1. Finetune the strongest model you can afford on a small dataset.
  2. Use it to generate more training data.
  3. Train a cheaper model on that generated data.

Recognizing which path a proposal is on tells you what it's optimizing for — the progression path for finding the best cost/quality model, the distillation path for getting a small, cheap production model out of a strong one.

4. What to interrogate in a finetuning proposal

This is the unit's real job. When finetuning is on the table, these are the questions that turn it from an engineering enthusiasm into a business decision:

The proposal interrogation Data: Do we have the (instruction, response) pairs, at the quality and volume this method needs — or are we about to fund a big annotation project (Unit 11)? Base capability: can the model almost do this already (a good finetune) or not at all (a long shot)? Alternatives exhausted: did systematic prompting and RAG genuinely fail first (Unit 05)? Maintenance: who owns retraining and monitoring as base models improve — because a better base model can outrun our finetuned one, stranding the investment? Serving: one model or many variants (which points at adapter vs. full)?

What "good" looks like after this unit

You can now:

Deliberately out of scope here: the memory/quantization and PEFT internals — those are the engineers' implementation depth, not yours. That closes the architecture thread. The remaining DEPTH units cover data (20), monitoring (21), and safety (22).