Comparison
Description-based AI classification vs. a fine-tuned classifier
Short answer
Use description-based classification when you have little labeled data, categories that change or need to ship quickly: you write label descriptions instead of training. Fine-tune your own classifier when you have thousands of stable, labeled examples, very high volume and a team to maintain it. Many teams start with descriptions and collect data for fine-tuning later.
Before large language models, building a text classifier meant collecting and labeling data, training a model such as a BERT variant, then hosting and monitoring it. Now you can describe your categories and classify immediately. Both approaches have a place, and the right one depends on your data, your volume and how often your categories change.
Get Early Access to ClassifierHub: 2× credits in your first paid month.
Compared
- Labeled data: description-based needs none to start (only an evaluation set). Fine-tuning typically needs hundreds to thousands of examples per class.
- Time to first result: minutes vs. days or weeks.
- Changing categories: edit a description vs. relabel and retrain.
- Accuracy: fine-tuned models can win on narrow, stable tasks with lots of data. Description-based models win when data is scarce or labels are subjective and evolving.
- Operations: an API call vs. hosting, versioning and monitoring your own model.
- Cost at scale: self-hosted small classifiers can be very cheap per call at very high volume, once you've paid the engineering cost.
A practical path
Start with descriptions. Log every input, prediction and probability, and have people correct a sample. After a few months you have a labeled dataset from your real traffic. If volume and stability justify it, train a dedicated model on that data and keep the description-based decision as a fallback for low-confidence cases or new categories.
Where decision models sit
Decision models like Jev are description-based, but unlike chat LLMs they return calibrated probabilities over your labels in a fixed shape and bill no output tokens. That removes the main operational drawbacks of classifying with a chat model while keeping the no-training advantage.
Frequently asked questions
Related guides
Last updated . ClassifierHub is an independent product built on top of the Jev decision model, accessed through OpenRouter. It is not affiliated with or endorsed by TypeSafe or OpenRouter.