Definition
Fine-tuning starts from a model that already learned general patterns (language, vision, speech) and continues training on task-specific data — Somali news classification, medical Q&A, or your company's support tickets.
You update some or all model weights so behavior shifts toward your domain without training from random initialization, which would need far more data and compute.
In simple terms
A general contractor knows how to build houses. Fine-tuning is like hiring them to specialize in your local building codes and materials — same core skills, adjusted for your exact job.
Where you see it
- OpenAI and others fine-tune base models for instruction-following chat assistants.
- Whisper is fine-tuned on Somali audio for better ASR.
- Companies fine-tune LLMs on internal docs for domain-specific Q&A.
How it works
1.Pick a base model
Choose a checkpoint trained on broad data (GPT, Llama, BERT, etc.).
2.Prepare task data
Collect labeled or paired examples for your target behavior.
3.Train with lower learning rate
Small weight updates preserve general knowledge while learning the task.
4.Evaluate and deploy
Test on held-out data, then serve the adapted checkpoint.
Why it matters
- Fine-tuning is how most teams customize AI without training billion-parameter models from scratch.
- It is the standard path for low-resource languages and specialized domains.
Often confused
Fine-tuning always beats prompting.
Prompting and RAG are often faster and cheaper; fine-tune when you need consistent style, format, or domain depth.