Fine-tuning
Fine-tuning is additional training that teaches a general model your specific patterns: you show it hundreds or thousands of examples of your tasks, your tone and your formats, and it adjusts its internal weights until that behaviour becomes its default.
No jargon, no math
In plain words, what is it?
Fine-tuning starts from a capable base model and specialises it on your examples. You build a dataset of input-output pairs: customer questions and correct replies, raw instrument readings and finished reports, orders and standard responses. The model works through them for many rounds: at each pass, the error on your pairs is propagated backwards and the weights shift a little in the direction that reduces it. Afterwards the model has not learned new rules; it has moved its statistical habits until your format is the most likely one. The cost is real (rented GPUs, days of data preparation) but orders of magnitude below training from scratch. Where it breaks: too few examples or noisy labels lead to overfitting, where the model memorises your examples and stumbles on anything new. And remember: fine-tuning teaches style and format, not fresh facts; for facts that change weekly you still need retrieval.
- 01
The base model. You start from a general model that is already capable.
- 02
Your examples. You build input-output pairs from your real tasks.
- 03
Training. The model works through the pairs; the error nudges its weights towards your patterns.
- 04
Evaluation. A held-out set of examples the model has never seen tests whether the specialisation holds.
- 05
Release. Pass the evaluation and the model ships; fail it and you train again.
Terms that live nearby.
Training
Training is the process that turns an untrained model into a working AI: you feed it thousands or millions of examples, compare its guesses with the correct answers, and nudge its internal settings until the guesses get good.
ReadOverfitting
Overfitting is when a model memorises its training data instead of learning the underlying pattern: near-perfect on examples it has seen, unreliable on anything new.
ReadDataset
A dataset is the organised collection of examples an AI learns from or is judged on: thousands or millions of rows of text, images or sensor readings, cleaned, labelled and split into training, validation and test parts.
Read§ 04
Want this running in your stack?
We design and ship AI systems with permissions, logs and human approval. It starts with a 30-minute scoping call.