Module 4.3

Three asks, three right answers

Full fine-tune, LoRA, and prompt-only aren't ranked best-to-worst — they answer different problems. Pick a real Nimbus-bot request and see which one actually fits, and why the other two would be a mistake here.

A REQUEST COMES IN FOR NIMBUS-BOT...

The math

What each option actually costs

Prompt-only / RAG

trainable params = 0

all the "learning" happens in the prompt, per request

training cost: none · per-request cost: higher

LoRA

trainable params = 2·d·r

Module 4.1: typically 0.1%–1% of the full model

training cost: low · per-request cost: ~same as base

Full fine-tune

trainable params = 100%

every weight, every layer

training cost: high · needs a full model copy per version

Cost order of magnitude, not exact numbers — but the ordering (0 ≪ 0.1–1% ≪ 100%) holds for essentially every real model size.

The Nimbus thread

The pattern across all three scenarios

Notice what actually drove each decision: it was never "which technique is more advanced." It was how often the underlying information changes and how much labeled data exists. Facts that change daily belong in a prompt or a retrieval step (Module 5), not baked into weights. Style and tone, with even a few hundred examples, is exactly LoRA's sweet spot (Module 4.1). A genuinely new capability, with enough data to support it, is the rare case that justifies touching every weight.

So what

The question is never "which fine-tuning method is best" — it's how often does this change, and how much data proves it.