aiaimodel.cards
Capabilities23

Fine-tuning

Training a model further on your own examples

Fine-tuning

Further training an existing model on your own examples so it follows a format, style or narrow task more reliably.

Why it matters. Can make a smaller, cheaper model good enough for a narrow task, but it takes data and evaluation work and is a poor way to add facts.

How it works

Fine-tuning takes a trained model and continues training it on a dataset you supply, usually example inputs paired with the outputs you want. The result is a new model version that behaves more like your examples.

For hosted models, makers offer fine-tuning on some models only, with their own prices for training and for using the tuned model, which can be higher per token than the base model. With open weights you can fine-tune yourself, often with parameter-efficient methods such as LoRA that train a small set of added weights.

Try clear prompts, in-prompt examples and retrieval first. Fine-tuning fits consistent format, tone or narrow classification; facts that change are better supplied through retrieval.

Worked example

Illustrative: 2,000 training examples of 500 tokens each is 1M tokens per pass; 3 passes (epochs) is 3M training tokens. At $25 per 1M training tokens, the run costs $75, before evaluation and any higher rate for using the tuned model.

Common mistakes

  • Fine-tuning to add knowledge that should come from retrieval, then getting outdated or invented facts.
  • Training on a few hundred inconsistent examples and getting inconsistent behavior back.
  • Skipping a held-out test set, so there is no evidence the tuned model beats a good prompt.

See it on a real card: Feature support on Claude Opus 5.5.

Related terms