LoRA
Also called low-rank adaptation, LoRA adapter
A cheaper way to fine-tune a model: its own weights stay frozen, and a small add-on called an adapter is trained beside them.
Fine-tuning a whole model means adjusting billions of numbers, which takes a lot of memory and time. LoRA, short for low-rank adaptation, leaves the model's own weights frozen and trains 2 small sets of numbers beside them instead. That's a small fraction of the work, with results comparable to tuning the whole model.
The trained part, the adapter, is small enough to keep several for different jobs on one base model, or to merge into the model so it runs as fast as before. With the model held at 4 bits during training, one of about 7 billion parameters can be tuned this way in as little as 5GB of graphics memory.
For example
A team trains a LoRA adapter on past support emails so a small open model sorts new ones into the right queues, then trains a second adapter for tagging invoices on the same model.
Put it to use
Where you'll meet loRA in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.