GLOSSARY
What is LoRA fine-tuning?
LoRA fine-tuningLoRA (low-rank adaptation) fine-tunes a model by training small low-rank matrices added to its weights instead of the weights themselves. It makes adapting a System One model to your own decisions fast and cheap enough for a laptop.
For a decision model the adapters go on the encoder and the decision head is trained in full. The adapters can then be merged back, so the result is an ordinary checkpoint that loads like the original.
Variants change how the adapters learn. DoRA also trains a magnitude for every output row, rsLoRA scales the update by α/√r so larger ranks keep learning, and LoRA+ trains one of the two matrices faster than the other.
System One Studio does all of this for Laya on your own machine (MLX on a Mac, PyTorch on Windows and Linux), measures the base and the tuned model on the same held-out split, and publishes the result to the registry in one step.
Read more
Next term: System One model · Every System One model