Stop Fine-Tuning When a Prompt Already Works
Someone wanted to fine-tune a model for our tone and I talked them out of it. The prompt already did 90%, and the fine-tune would have cost a week we didn't have.
Someone on my team wanted to fine-tune a model for our tone and I talked them out of it. The prompt already did 90% and the fine-tune would have cost a week and a dataset we didn't have.
Fine-tuning earns its keep when the behavior is hard to describe, a style, a judgment, a pattern no prompt pins down. For 'sound like us,' three lines of instruction beat a training run nine times out of ten.
The trap is novelty, reaching for the impressive option because it's there. Most teams don't need a custom model; they need a better prompt and a human review step they keep skipping.
Before the fine-tune, prove the prompt failed. If it hasn't, the fancier tool is just a more expensive way to do what you can already do. Save the training for the part prompts truly can't reach.