Training an AI model on a custom dataset means taking a model that already exists and continuing to adjust its internal settings — its weights — using your own examples, so it learns your specific patterns, vocabulary, or task instead of relying only on what it picked up during its original training.
You are not building a model from scratch; you are nudging an existing one toward your data. The two common versions are fine-tuning, where you feed the model matched input-and-output pairs, and continued pre-training, where you feed it a large pile of raw domain text so it absorbs your field's language before you teach it a task.
What actually changes inside the model
A model's weights are millions or billions of numbers that encode everything it learned. During ordinary use, those numbers stay frozen — the model just reads your prompt and predicts a response. Custom training unfreezes some of them and runs your examples through the model repeatedly, measuring how wrong each prediction is and nudging the weights to reduce that error.
The result is a model whose default behaviour shifts: it starts completing your kind of text, in your format, without being told every time. The data shape matters. For fine-tuning, you need input/output pairs — a question and its ideal answer, a messy record and its clean version, a support ticket and its correct reply.
For continued pre-training, you need volume of raw text: manuals, transcripts, internal documentation. A few hundred well-labelled pairs can meaningfully change a narrow task; continued pre-training generally wants far more raw text because you are teaching language patterns, not a single behaviour.
The quality of the examples usually matters more than the count. Fifty consistent, correct pairs beat five hundred sloppy ones, because the model copies the pattern it sees, mistakes included.
When custom training beats prompting or retrieval
Start by asking whether prompting alone fails. If you can describe the task in a paragraph and get good results, you do not need custom training — you need a better prompt. If the model keeps missing facts it was never told, retrieval is usually the cheaper fix: store your documents in a searchable index, fetch the relevant passages at question time, and let the model read them.
That keeps your knowledge current without retraining. Custom training earns its place when the problem is style, format, or a stable behaviour rather than missing facts. A concrete example: a clinic wants every patient summary to follow one fixed structure with specific phrasing.
Prompting gets close but drifts. Fine-tuning on, say, 300 examples of past summaries written in that exact structure teaches the model the shape directly, so the output lands correctly with a short prompt instead of a long instruction block. If instead the clinic's problem were that the model does not know its own formulary, retrieval would be the right tool — the facts change, the format does not.
Limits, costs, and failure modes
Custom training is not free or permanent. It costs money to run the training, time to prepare clean data, and ongoing effort to maintain — every time your format or policy changes, your dataset ages. You also need some technical fluency: splitting data into training and held-out test sets, watching for overfitting, and knowing when to stop.
Two failure modes are common. Overfitting happens when the model memorises your small dataset instead of learning the pattern; it aces your examples and flops on anything new. Catastrophic forgetting is the opposite risk: training hard on your narrow data can erode general abilities the model used to have, so it gets better at your task and worse at everything else.
Both are why you keep a test set the model never trained on. As for what to expect from the wider tool market, our AI tool database tracks 360 AI tools with pricing and capability snapshots recorded at verification time, and customisation features vary widely across them — some offer fine-tuning, many do not.
Pricing and availability change often, so the vendor's own page is the only reliable source before you commit. A useful rule: if your problem is knowledge, retrieve. If it is behaviour, fine-tune. If it is a one-off, just prompt.