How to Train Your own Decision Model like Jev using Unsloth

In this article, you will learn how to turn an open-weight language model into a decision model that classifies user intent, by fine-tuning Qwen3.5-4B on the CLINC150 dataset with Unsloth. What decision models are, and how Unsloth’s FastDecisionModel scores a fixed set of choices instead of generating text. How to convert CLINC150 into Unsloth’s decision format and fine-tune Qwen3.5-4B with 4-bit LoRA in Google Colab. How to evaluate, calibrate, test, and save the resulting intent classifier. What if you could turn a regular language model into a decision model that chooses an answer instead of generating a long response? Models like Jev, Laya, and Clef are designed to make structured decisions. Instead of generating text token by token, they select from predefined options and return probability estimates. Unsloth now supports training decision models from open-weight language models such as Qwen and Gemma. Its official decision-model training guide explains the method and provides ready-to-run notebooks. In this tutorial, I’m using Google Colab, Unsloth, and Qwen3.5-4B to build a custom intent classifier using the CLINC150 dataset. We will: By the end, you’ll have your own decision model that can classify user requests and return structured predictions with confidence scores. So, let’s get…