---
title: "Discussion on \"Fine-Tune an LLM on Your Mac with MLX: QLoRA in 16 Seconds (Real Run, No Cloud)\" | Hashnode"
description: "Discussion on \"Fine-Tune an LLM on Your Mac with MLX: QLoRA in 16 Seconds (Real Run, No Cloud)\". A real MLX QLoRA run on a 16 GB Mac: Qwen2.5-3B, 16-second training, 2.5 GB memory, and why the best-loss checkpoint gave the worst answer."
url: "https://hashnode.com/posts/fine-tune-an-llm-on-your-mac-with-mlx-qlora-in-16-seconds-real-run-no-cloud/6acb2294061765bf8144aadf"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Fine-Tune an LLM on Your Mac with MLX: QLoRA in 16 Seconds (Real Run, No Cloud)](https://rainytech.hashnode.dev/fine-tune-an-llm-on-your-mac-with-mlx-qlora-in-16-seconds-real-run-no-cloud?utm_source=hashnode&utm_medium=feed)

A real MLX QLoRA run on a 16 GB Mac: Qwen2.5-3B, 16-second training, 2.5 GB memory, and why the best-loss checkpoint gave the worst answer.

Oct 11 · 8 min read · rainytech.hashnode.dev  
Updated 2d ago  
Read the article

-   [#ai](https://hashnode.com/tag/ai)
-   [#macos](https://hashnode.com/tag/macos)
-   [#machine-learning](https://hashnode.com/tag/machine-learning)
-   [#python](https://hashnode.com/tag/python)

## Written by

[RainyTech](https://hashnode.com/@rainytechlab)

@rainytechlab · Phnom Penh, Cambodia

  
[View profile](https://hashnode.com/@rainytechlab)

AI, coding and tech explained clearly. Real tutorials with real numbers. New YouTube videos every week.

RainyTech is a YouTube channel by Rainy Reaksmey, a software engineer and mobile developer from Cambodia. Real tutorials, real numbers: local AI on Mac, Python, React and the newest devices. Watch: youtube.com/@RainyTechLab Website: rainytech.dev

## Responses 3

Markdown supported · ⌘↵ to post

[Ahmet Özel](https://hashnode.com/@ahmetozel)

[Oct 11](https://hashnode.com/posts/fine-tune-an-llm-on-your-mac-with-mlx-qlora-in-16-seconds-real-run-no-cloud/6acb2294061765bf8144aadf/comment/6acb5a734e3e0972859ce1ec)

The step-20 answer dropping authorization is a useful completeness failure, but the displayed comparison question also appears in your training JSONL example. That probe shows how a trained example behaves; it should stay separate from the two genuinely held-out test prompts.

I would choose the checkpoint using a small validation rubric that checks both required concepts, then report untouched test generations only after that choice. Repeatedly reading test answers to pick step 10 would make those prompts part of selection. Keeping these roles explicit would strengthen the smoke test without overstating what twenty examples establish.

[RainyTech](https://hashnode.com/@rainytechlab)

[Oct 13](https://hashnode.com/posts/fine-tune-an-llm-on-your-mac-with-mlx-qlora-in-16-seconds-real-run-no-cloud/6acb2294061765bf8144aadf/comment/6acb5a734e3e0972859ce1ec#reply-6ace1e2381c276fc5b82f918)

Good points, Ahmet. The sample line in the post is the same authentication/authorization question I used for the before/after answers, so as written, that comparison shows behaviour on a trained example, not generalisation. To be precise on selection: I picked step 10 from that one answer, not from the two held-out test prompts. That's still a weak basis, though. A small validation rubric to pick the checkpoint, then untouched held-out answers, is the cleaner setup. Thanks for spelling it out.

[Ahmet Özel](https://hashnode.com/@ahmetozel)

[Oct 15](https://hashnode.com/posts/fine-tune-an-llm-on-your-mac-with-mlx-qlora-in-16-seconds-real-run-no-cloud/6acb2294061765bf8144aadf/comment/6acb5a734e3e0972859ce1ec#reply-6ad13c6f23aa965ffb78d4f2)

You are right to separate those: I overstated the selection concern by leaving open whether the two held-out prompts influenced step 10. Your clarification says they did not. The trained-example comparison still measures retention of that example rather than transfer.

For the next run, the validation rubric can include unseen phrasings that require both authentication and authorization, with checkpoint selection fixed before opening the final test answers. Reporting the trained probe separately would keep it useful as a smoke test without making it carry the generalization claim.
