ThirdBrAIn.tech

QWEN-3 EASIEST WAY TO FINE-TUNE WITH REASONING 🙌

2 min read

QWEN-3 EASIEST WAY TO FINE-TUNE WITH REASONING 🙌

AI Summary

In this video, the presenter demonstrates how to fine-tune a Quinn 3 model on a custom dataset. Key points include:

  • Structure of data is crucial due to the hybrid thinking mode of Quinn models, enabling or disabling reasoning with hyperparameters.
  • The video uses Unslaught for fine-tuning and discusses the concept of catastrophic forgetting during continual fine-tuning.
  • The LORA (Low-Rank Adaption) method is introduced to avoid changing the original weights of the model, allowing for smaller adjustments through adapter weights.
  • The importance of dataset preparation for fine-tuning is emphasized, including combining reasoning traces with non-reasoning datasets to preserve reasoning capabilities.
  • Examples are provided showing how to convert datasets into a suitable format for the model.
  • Hyperparameters are discussed, including batch size, learning rate, and temperature settings for optimal inference performance.
  • Finally, the presenter shows how to run inference in both thinking and non-thinking modes, highlighting the model’s adaptability.

Description

Learn how to fine‑tune Qwen‑3‑14B on your own data—with LoRA adapters, Unsloth’s 4‑bit quantization, and just 12 GB of VRAM—while preserving its chain‑of‑thought reasoning. I’ll walk you through dataset prep, the key hyper‑parameters that prevent catastrophic forgetting, and the exact Colab notebook to get you running in minutes. Build a lightweight, reasoning‑ready Qwen‑3 model tailored to your project today!

LINKS:
https://qwenlm.github.io/blog/qwen3/
https://docs.unsloth.ai/basics/unsloth-dynamic-2.0-ggufs
https://huggingface.co/datasets/unsloth/OpenMathReasoning-mini
https://docs.unsloth.ai/basics/qwen3-how-to-run-and-fine-tune
https://huggingface.co/datasets/mlabonne/FineTome-100k
https://docs.unsloth.ai/get-started/fine-tuning-guide
https://arxiv.org/html/2308.08747v5
https://heidloff.net/article/efficient-fine-tuning-lora/

NOTEBOOK: https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Qwen3_(14B)-Reasoning-Conversational.ipynb
Fine-tuning Playlist: https://www.youtube.com/playlist?list=PLVEEucA9MYhPjLFhcIoNxw8FkN28-ixAn

Website: https://engineerprompt.ai/

RAG Beyond Basics Course:
https://prompt-s-site.thinkific.com/courses/rag

Let’s Connect:
🦾 Discord: https://discord.com/invite/t4eYQRUcXB
☕ Buy me a Coffee: https://ko-fi.com/promptengineering
|🔴 Patreon: https://www.patreon.com/PromptEngineering
💼Consulting: https://calendly.com/engineerprompt/consulting-call
📧 Business Contact: engineerprompt@gmail.com
Become Member: http://tinyurl.com/y5h28s6h

💻 Pre-configured localGPT VM: https://bit.ly/localGPT (use Code: PromptEngineering for 50% off).

Signup for Newsletter, localgpt:
https://tally.so/r/3y9bb0

Fine-Tuning Qwen-3 Models: Step-by-Step Guide

00:00 Introduction to Fine-Tuning Qwen-3
01:24 Understanding Catastrophic Forgetting and LoRa Adapters
03:06 Installing and Using unsloth for Fine-Tuning
04:19 Code Walkthrough: Preparing Your Dataset
07:14 Combining Reasoning and Non-Reasoning Datasets
09:48 Prompt Templates and Fine-Tuning
16:13 Inference and Hyperparameter Settings
18:11 Saving and Loading LoRa Adapters

Graph View