Hello Model
← Training basics

Lesson 2 of 8

Prompting vs fine-tuning vs training from scratch

In plain words: Start with the cheapest option that could work, and only move up when you have the data and a reason.
  • Prompting / zero-shot: use a pretrained model or LLM as-is. No training data needed; great for prototypes.
  • fine-tuning / transfer learning: adapt a pretrained model with hundreds to thousands of your own examples. The usual sweet spot.
  • Training from scratch: only for large, unusual datasets (e.g. tabular data with gradient-boosted trees, or very specialised domains).
  • For tabular data, 'from scratch' with LightGBM/XGBoost is cheap and normal; for text, images and audio, fine-tune instead.
Tip If prompting already reaches your target metric, ship it and collect data for later.