Techniques

Fine-Tuning

Continuing training of a pre-trained model on domain-specific data to specialize its behavior.

Fine-tuning takes a general-purpose foundation model and updates its weights on a curated dataset of examples. The result is a model that follows a specific tone, format, or classification schema more reliably than prompting alone.

Modern fine-tuning is dominated by parameter-efficient methods like LoRA and QLoRA, which train only 0.1–1% of parameters. A useful fine-tune can be produced in hours on a single GPU for a few hundred dollars.

Fine-tuning is not the right tool for teaching a model new facts — that is what RAG is for. Fine-tuning teaches behaviors: 'always respond in this JSON schema', 'classify this ticket into these 12 categories', 'write in our brand voice'.

Key points

  • Needs 100–10,000 high-quality labeled examples
  • LoRA / QLoRA reduce cost 10–100x vs full fine-tuning
  • Fine-tuned models can run on smaller base models, cutting inference cost
  • Iteration cycle is slower than prompt engineering — plan evaluation harness first

Common use cases

Custom classifiers (support tickets, sentiment, intent)
Structured extraction with strict schemas
Brand-voice content generation
Domain-specific code assistants

Frequently asked

Related terms

Applying Fine-Tuning in your business?

We help enterprises design, ship, and govern AI systems end-to-end.