What it means
Fine tuning continues training a pretrained model on a new dataset so it specializes for a target task. Unlike pure feature extraction, fine tuning actually updates the model weights.
How to do it carefully
Because the model already knows a lot, fine tuning usually uses a small learning rate to avoid washing away useful pretrained knowledge.
- Start from pretrained weights, not random ones
- Use a low learning rate so updates are gentle
- Train for fewer steps than full pretraining
A common danger is catastrophic forgetting, where aggressive updates erase general skills. Another is overfitting when the target dataset is tiny.
Instruction tuning
For language models a popular variant is instruction tuning, where the model is fine tuned on many examples of instructions paired with good responses. This teaches the model to follow directions rather than just predict the next token. It often comes before alignment steps that use human feedback.
Key idea
Fine tuning updates a pretrained model on task data with a small learning rate, specializing it while avoiding catastrophic forgetting.