LLM Fine-Tuning

How it works

A pre-trained foundation model already has broad general knowledge from its initial training. Fine-tuning continues training that same model on a smaller, curated dataset specific to a task or domain — computing the difference between the model's predictions and the correct labeled answers, then updating the model's parameters to reduce that gap. Because it starts from an already-capable pre-trained model rather than training from scratch, fine-tuning requires far less data and compute than initial model training, while still meaningfully changing how the model responds.

Fine-tuning vs. prompt engineering

Prompt engineering shapes a model's output through instructions given at query time — it doesn't alter the model's underlying parameters at all. Fine-tuning changes the model itself by updating its parameters through additional training. Prompt engineering is faster and cheaper to iterate on, and is often sufficient for adjusting output format or tone. Fine-tuning is the stronger tool when the required behavior change is deep or consistent enough that no amount of prompting reliably produces it — a specialized domain vocabulary, a rigid structured output format under all conditions, or a reasoning pattern the base model doesn't naturally perform well.

Two main approaches

Instruction tuning trains a model to better follow instructions and perform a broader range of tasks competently, generally improving its usefulness across many use cases. Alignment tuning shapes the model's responses to match specific values, tone, or safety requirements — used to make a model's behavior consistent with an organization's specific standards, beyond raw task capability.

When it applies

Fine-tuning earns its cost when a use case needs a consistent, specific output style or format across every response, a specialized domain vocabulary or reasoning pattern the base model doesn't handle well out of the box, or behavior that needs to hold reliably without depending on a carefully engineered prompt every time. See RAG vs fine-tuning for the full decision framework distinguishing this from retrieval-based approaches.

When it doesn't apply

Fine-tuning is the wrong tool for teaching a model new, frequently-changing facts — retrieval-augmented generation handles that far more cheaply and reliably, since updating a RAG system means re-indexing a document while updating a fine-tuned model requires another training pass. Fine-tuning is also unnecessary when prompt engineering alone reliably produces the needed output — reaching for fine-tuning by default, before confirming prompting genuinely can't solve the problem, adds training cost and ongoing model-version maintenance burden without a corresponding need.

Fine-tuning vs. RAG at a glance

What Code Ninety does

Code Ninety reserves fine-tuning for cases where a client genuinely needs a specific output style or specialized behavior that retrieval alone can't produce, defaulting to RAG for fact-grounded, auditable use cases specifically because of the source-attribution advantage RAG carries for fintech and healthcare clients. Code Ninety is an AI agency in Pakistan building enterprise RAG and agentic systems. See the LLM integration case study for how this was applied in a regulated fintech environment.

Related terms

• → Retrieval-augmented generation
• → Vector database
• → Model Context Protocol