What is Fine-Tuning?

Fine-Tuning is the process of continuing the training of a pre-trained Large Language Model on a curated dataset so it adopts a domain, style or output format more reliably than a base model would with prompting alone. In commerce it is one of three knobs available to teams, alongside Prompt Engineering and Retrieval-Augmented Generation, and the one with the highest cost and operational weight.

Definition

Fine-Tuning updates a subset of the model's weights based on input-output pairs supplied by the team. Parameter-efficient methods like LoRA train small adapter matrices instead of the full model, which slashes compute and storage cost and lets several adapters coexist on the same base. Instruction tuning focuses on following structured tasks, while preference tuning, such as DPO or RLHF, aligns model behaviour with chosen versus rejected outputs from human raters. Datasets typically range from a few hundred to tens of thousands of high-quality examples, and the model is evaluated on a held-out set covering edge cases, Hallucination, and adherence to the target schema before deployment.

Why it matters

Fine-Tuning is the right tool when prompt-only approaches blow past the context window, when a task is so repetitive that paying for long instructions every call becomes expensive, or when a specific output shape is non-negotiable. It can drop tokens per request by an order of magnitude and improve latency, which matters for high-volume surfaces like Storefront API calls and chat. It is the wrong tool when the underlying knowledge changes frequently, where Retrieval-Augmented Generation is cheaper and fresher. Fine-tuned models also need version control and re-evaluation when the base model updates, which adds to operational overhead in a Composable Commerce environment.

Use cases

A merchant fine-tunes a small model on brand-voice samples for high-volume product description generation, then serves it behind a CMS tool. A classification model is tuned on labelled support tickets to route conversations before a larger agent picks them up. A schema-extraction model is tuned to turn unstructured supplier feeds into validated JSON for the PIM.

Related

Explore Agentic Frontend Management Platform · Composable Digital Experience Platform.