How I Fine-tuned Qwen3-4B on a Single GPU in 5 Steps

Note: AI Weekender has moved. New posts are published at ai-weekender.com, and this Substack is now an archive.

To keep receiving weekly issues, please subscribe at ai-weekender.com instead of here.

For this week’s technical deep dive, I fine-tuned a small LLM (Qwen3-4B-Instruct-2507) on Banking77. This is a public dataset of 13k real customer inquiries to a banking help desk, each tagged with one of 77 support intents, such as card delivery, lost card etc.

On a held-out eval set of 100 unseen messages, the base model got 0% right. After QLoRA training, the LLM correctly classified 83%!

That’s a real ops win for a banking startup. At 83% auto-routing, humans only handle the ~17% the model misses, so you can put ops on the edge cases that need judgment, not mindless label-picking.

I ran the Jupyter notebook on Google Colab’s free tier when a T4 was available (took ~62 min to train). Note that GPU type and session limits aren’t guaranteed on free Colab.

Paid subscribers get the full five-step walkthrough with code and also a link to my Colab notebook. You can copy the notebook and try it out on your own when compute is available.

Note: AI Weekender has moved. New posts are published at ai-weekender.com, and this Substack is now an archive.

To keep receiving weekly issues, please subscribe at ai-weekender.com instead of here.