Unsloth Turns Qwen3.5-4B Into a 10-Intent Classifier on a Single T4

A large language model does not always need to converse. In many products, it is expected to deliver a more practical result: identify a customer's intent, choose the next tool, or decide whether a message should proceed. A conventional LLM may generate an explanation when the system needs only one short answer—for example, “balance inquiry” with a confidence score.

Machine Learning Mastery showed how to build such a classifier on top of the open Qwen3.5-4B model. The setup combines Unsloth, 4-bit LoRA, and the CLINC150 dataset. The demonstration runs in Google Colab on a Tesla T4 with 14.6 GB of VRAM and limits the task to ten intents. It is not a production-ready model, but a clear starting point for a developer who wants to replace a prompt chain with a dedicated selection mechanism.

The Model Chooses Instead of Composing

Consider the message: “Show the balance in my savings account.” A generative model might write instructions, ask which bank the user means, or begin reasoning. A classifier must do something else: select the balance intent from a predefined set and return a confidence score.

In Unsloth, FastDecisionModel handles this task. A specialized head is added to the base language model to score the permitted options. Instead of generating tokens sequentially, the system compares fixed decisions. It supports selecting one option, answering yes or no, and scoring on an ordered scale.

This format works for request routing, moderation, sentiment detection, and agentic workflows. For example, a small model can sit in front of the main LLM and decide whether to route a request to search, a knowledge base, a payment service, or a human. The output is constrained by a schema, making it easier to validate and pass to the next component.

How the Training Setup Works

The author loads Qwen3.5-4B in 4-bit mode and attaches rank-16 LoRA adapters. Most of the original parameters remain frozen, while training updates the compact adapters and the decision head. Gradient checkpointing further reduces memory consumption.

The data must also be restructured. Each record contains three parts: the user's text, a description of the available decisions, and the correct option. In the CLINC150 example, 10 of the 150 intents are selected at random, including date, time, directions, ingredient substitution, and jump-starting a car. Only examples associated with those intents are then retained.

The option list is what makes the task controllable: the developer defines in advance which actions the system is allowed to take. But that also introduces a cost. The more options there are, the longer the input and the more demanding the training process. Ten classes on a T4 do not yet prove that the same configuration will handle all 150 categories—or real customer messages—without changes.

Where the Demonstration Ends

The available excerpt provides no final accuracy, confusion matrix, or comparison with a conventional classifier. It is therefore impossible to claim that Qwen3.5-4B already outperforms cheaper models. The quality of recognizing requests outside the supported set is not shown either, although CLINC150 includes separate examples of such queries.

For production, model confidence alone is not enough. It must be calibrated on a held-out dataset, a rejection threshold must be chosen, and errors between similar intents must be examined. If two actions differ in only one detail—for example, “freeze card” and “close card”—the cost of a wrong decision may be higher than the latency introduced by asking a follow-up question.

The practical conclusion is straightforward: FastDecisionModel is interesting as a way to turn an open LLM into a tightly constrained router. It makes sense to begin with a small set of clearly distinguishable actions, a mandatory “I don't know” class, and an error log. The solution can then be compared with a compact BERT classifier, rules, and calls to a larger model. What matters here is not the architectural flourish, but the share of correct decisions at an acceptable latency and cost.