What does fine-tuning change?

A pretrained model has already been trained once. Fine-tuning runs more training on a narrower set of examples so the model’s behavior moves toward that set.

Hu and colleagues, in the LoRA paper, describe full fine-tuning as retraining all parameters, and they use a 175-billion-parameter GPT-3 as the case where storing a separate fully fine-tuned copy becomes prohibitive.

LoRA is their alternative: freeze the base and train low-rank matrices. Both are still only as good as the examples.

When is retrieval the better change?

Fine-tuning bakes a pattern into weights. It does not keep a policy page current.

If the model is wrong because it cannot see today’s document, RAG versus fine-tuning is the decision: retrieval puts the passage in front of the model for that question, and fine-tuning changes how the model tends to respond.

Use fine-tuning for a stable format or behavior. Use retrieval for facts that move.

How do you know the run helped?

Keep a prompt-only baseline on the same held-out examples. If the prompt already produces the format, the training run may be maintenance you do not need.

Never score the model on rows it trained on. The Hugging Face training guide is one library’s documentation for running a training loop.

It is not a result for your dataset.

Common questions

What is fine-tuning?

It is further training of a pretrained model on a selected dataset, so the model’s parameters move toward that data.

Should you fine-tune to teach a model a new document?

Usually no. A document that can change belongs in retrieval.

Fine-tuning is for behavior and format that should persist without pasting the source every time.

Sources & further reading

  1. Hugging Face: Fine-tuning ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories