In plain English
Fine-tuning continues a model's training on your examples, so it behaves as they show. LoRA freezes the model and trains a small adapter, which is faster and cheaper. Use it for style, format and narrow tasks. Do not use it to teach facts that change.
- Fine-tuning
- Extra training on your own examples.
- LoRA
- Training a small adapter on top of a frozen model.
- Adapter
- A small add-on file that changes how a model behaves.
Picture it
Two ways to fine-tune
Full fine-tune
Change every number. Heavy and costly
LoRA adapter
Freeze the model, train a small add-on
Quantized LoRA
Add-on on a compressed model, even lighter
Try it yourself
Fine-tune, or fix another way?
Decide the best first move for each problem.
The model must always write in our brand voice
It forgets to include the headings we asked for
It does not know our current price list
A high-volume tagging task where a small tuned model could replace a big one
It answers too long
It cannot answer questions about last week's policy change
0 of 6 right