I have a recurring classification task where the labels are fairly specific to our business. A detailed prompt gets acceptable results, but examples and edge cases are making it longer than the actual input.

For people who have tried both, when did OpenAI fine-tuning become worthwhile compared with keeping the task prompt-based? I’m especially curious whether it helped with consistency on borderline examples, and how much time you had to spend preparing and maintaining training data. A few real-world experiences would help me decide whether to test it or keep refining the prompt.