Low-rank adaptation
Also called: LoRA
By HubAI editorial team · Source checked: 24 October 2026What does it mean?
A fine-tuning technique that learns small additional weight updates while keeping the original model weights fixed. It changes learned behaviour; it is not simply a longer prompt.
Picture this
A fictional support team wants a particular reply format. It tests a small adapter against its unchanged base model, using separate example enquiries to see whether the format improves without losing useful behaviour.
Where you’ll come across it
Adapting a model when full fine-tuning would be impractical. An adapter belongs to a particular compatible setup, so keep the base model, training settings and evaluation notes together.
What it doesn’t mean
A smaller training update is not a guarantee of better answers or lower total running costs. Check task quality, model compatibility and data permissions before committing to an adapter.
Decide whether an adapter solves your problem
A short exercise · use fictional or non-sensitive material.
- Write down the behaviour you want to change, such as a consistent fictional support reply format.
- Compare a clearer prompt and a few examples first. If you still consider training, keep a separate set of test enquiries.
- For an adapter trial, record the exact base model, adapter and settings. Judge the result on the held-out tasks, not the training examples.
Is the problem learned behaviour, or missing information that retrieval would address more directly?
Did that make it clearer?
Likes are a private on-page choice, not a public rating. For a correction or complaint, send the pre-filled form to our editorial inbox. Nothing is sent until you submit it.Check the source
Written and explained by HubAI. The reference below lets you go deeper; it is not an endorsement of a particular product.
Technical reference: huggingface.co ↗