Learn
See
Try
Make

Tune a model without a big machine

What this lesson is about

Learn the cheap way of teaching a model, where you train a small extra piece instead of the whole thing.

Your cousin says a model cannot be tuned without a big computer, and you only have a free notebook. What do you tell her?

Why this matters

The Hugging Face page on cheap tuning says these methods train only a small adapter, not the whole model. That extra piece is usually a LoRA. It cuts the compute and storage cost, and lets big models be trained on ordinary hardware. The model's own numbers are left alone, which is why a free notebook is enough. One documented run, from Google's own LoRA tutorial, read on 14 Sep 2026. It tunes Gemma 3 1B on a Colab T4 at LoRA rank 4. One pass over 1,000 examples, batch size 1, inputs cut at 256 tokens. That pass took 923 seconds, about 15 minutes. The tutorial shows replies before and after, not a score, so the kept-aside score is yours to add.

Tune a small model on a free machine and report the change on examples it never saw.

The cheap method trains a small adapter and leaves the model itself alone.
Three settings are yours: the LoRA rank, the training length, and where to stop.
Measure on the examples you kept aside, with the same measure you used before.

Task

Tune a small model on a free notebook for one narrow job, and report it honestly.

Weak approach Runs training overnight, keeps the copy from the final pass, and reports that the training numbers fell steadily. The forty kept aside were never scored after the run.

Stronger approach Loads Gemma 3 1B on a free Colab T4 at LoRA rank 4, the tutorial's own setting. Runs three short passes and scores the forty kept aside after each one. Keeps the checkpoint from the pass where that score peaked, not the last one. Writes the rank, the passes, the minutes and both scores on the page.

Why the stronger approach works The weak report measures the examples the model saw. The strong one measures the examples it never saw, and names the settings that produced the number.

Try a changed situation A teammate shares a tuned model and says it works well. He sends no numbers at all. What do you ask for?

Make something yourself Tune one small model. Show the same kept-aside examples scored before and after, the settings you used, and how long the whole run took.

Lesson 7 of 8 on Open models on a small budgetnextSay plainly what it cannot doPart 2 ends in a work sample you can send

Read the primary or official source: Hugging Face PEFT