𝗜 𝗙𝗶𝗻𝗲 𝗧𝘂𝗻𝗲𝗱 𝗔𝗻 𝗟𝗟𝗠 𝗔𝗻𝗱 𝗧𝗵𝗲𝗻 𝗦𝗮𝗶𝗱 𝗡𝗼

𝗜 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗲𝗱 𝗔𝗻 𝗟𝗟𝗠 𝗔𝗻𝗱 𝗧𝗵𝗲𝗻 𝗦𝗮𝗶𝗱 𝗡𝗼

A client wanted an AI for a landscaping business. They wanted a specific brand voice. I built a fine-tuned model. Then I told them not to use it.

I tested two versions.

Version A: Base model with a knowledge base (RAG). Version B: Base model with supervised fine-tuning.

Version B sounded better. It was warmer. It took ownership of mistakes.

But small data sets create risks. The model made up fake warranties. It lied to sound professional. This is a liability for a business.

The Azure bill ended the debate.

One day of cost:

• Training: 0.58 euros.
• Tokens: near zero.
• Hosting: 54.47 euros.

Hosting costs 1,630 euros per month. You pay this before a user sends one message. The base model has no standing fee. You pay only for what you use.

My advice for this client:

• Keep the base model.
• Use RAG for facts.
• Put voice examples in the prompt.

Fine-tuning earns its place when:

• High volume makes tokens too expensive.
• You have a large, clean dataset.
• Prompting does not work.

This client did not meet those needs.

The most valuable part of my work was saying no. Many people follow tutorials to build a model. Few people measure the cost to see if it makes sense.

Source: https://github.com/xaphor/landscaping-llm-brandvoice-eval Optional learning community: https://t.me/GyaanSetuAi