Fine-Tuning
Last updated:
What fine-tuning is
Large language models start with general pre-training on huge amounts of text. Fine-tuning is a second, much smaller training run on examples you choose, usually pairs of inputs and the outputs you want. The result is a new version of the model whose internal parameters, called weights, have shifted toward those examples.
Fine-tuning is good at teaching patterns: a consistent tone, a fixed output format, a classification scheme or domain-specific phrasing. It is less suited to teaching facts that change, such as prices, stock or opening hours, because every change would need another training run.
How fine-tuning works
A typical fine-tuning project follows four steps:
- Collect examples: a set of high-quality input and output pairs that show the behavior you want. Quality and consistency matter more than volume.
- Train: the model provider or your own infrastructure adjusts the model’s weights on those examples, often with methods such as LoRA that update only a small part of the model.
- Evaluate: compare the tuned model with the original on test questions it has not seen, to check that it improved without losing general ability.
- Deploy and maintain: the tuned model replaces the base model in your application and has to be retrained when the base model is retired or the desired behavior changes.
Example
A law firm wants its internal assistant to summarize client calls in a strict five-part format. Instructions alone get the format right most of the time, but not always. The firm fine-tunes a model on a few hundred past summaries written by its staff, and the tuned model follows the format more consistently. For anything factual, such as the current fee schedule, the firm still relies on retrieval, because those details change during the year.
Why it matters for business chatbots
For most website chatbots, the main job is answering accurately from the company’s current content. Retrieval-augmented generation (RAG) usually handles that better than fine-tuning: updates take effect once content is re-indexed, answers can be traced to sources, and there is no training data set to build. Fine-tuning is worth considering when a business needs a specific behavior that clear instructions and examples cannot achieve, and has the data and budget to maintain a custom model.
| Fine-tuning | RAG | |
|---|---|---|
| What changes | The model’s weights | The text given to the model |
| Best for | Style, format, narrow tasks | Facts, policies, product information |
| Updating knowledge | Train the model again | Edit and re-index the content |
| Traceability | Hard to tell where an answer came from | Answers can point to source passages |
Fine-tuning and intoCHAT
intoCHAT does not fine-tune models. When you “train” an intoCHAT agent, it processes your website pages, files, text snippets and Q&A pairs into searchable chunks with embeddings. The underlying model, one OpenAI model chosen and managed by intoCHAT, stays the same. That is why changes take effect after you retrain a source, without waiting for a training run. To shape tone and behavior, you write the agent’s instructions and add Q&A pairs for answers that must be exact.