What is Fine-tuning?

Fine-tuning continues training an existing language model on your own labelled examples so it adopts a specific tone, format, or task more reliably. It changes the model's weights, unlike RAG, which leaves the model untouched and supplies knowledge at query time.

Read in RomanianReglaj fin (fine-tuning)

Why it matters

Fine-tuning is how you get consistent behaviour (a fixed output shape, a house style, a narrow classification) rather than fresh knowledge. Used well it makes a smaller, cheaper model perform a specific job as well as a much larger one.

Fine-tuning vs RAG

They solve different problems. RAG injects up-to-date facts the model can cite; fine-tuning teaches a durable skill or format. If your answers are wrong because the model lacks current information, reach for RAG. If they're wrong because the model won't follow your structure or voice, fine-tuning is the lever. Many production systems use both.

Where teams get it wrong

Teams fine-tune to "add knowledge," then watch the model confidently invent facts. Fine-tuning bakes in patterns, not a reliable, updatable fact store, that's what RAG is for. It also needs real labelled data; a few dozen examples rarely move the needle.

Related terms

Used in

RAG vs fine-tuning: which one does your AI feature need?Teams reach for fine-tuning when they usually mean grounding. Here is the honest difference, and why most products want retrieval first.

← Back to glossary