Extra training that shapes an already-trained model for a specific job or style.
Fine-tuning takes a model that has finished pretraining and trains it a bit more on a smaller set of examples, so it fits a particular task or tone. It doesn't start over; it adjusts what the model already knows. Teaching a model to answer questions helpfully, using model answers, is one kind of fine-tuning.
Often the original numbers stay frozen and only a small add-on set is trained. LoRA, published in 2021, works this way, which makes fine-tuning even big models fairly cheap. Fine-tuning is good at changing style and format; for adding new facts, RAG, which looks things up and answers with the source open, is often a better fit.
Think of it like this
Say a friend who's already great at soccer is asked to play goalkeeper. They don't relearn soccer from scratch; they add a few weeks of goalkeeper practice. Fine-tuning likewise adds one skill on top of what's already there.
For example
From the same base model, fine-tuning can produce a coding helper, a math tutor and a friendly chat buddy as separate versions.
Records about this
Source-checked moments from AI history.
The 137B model outperformed the 175B GPT-3 zero-shot on 20 of 25 tasks
Freezes the weights and trains only low-rank matrices, cutting trainable parameters to 1/10,000 for GPT-3 175B, with no increase in inference latency