LoRA

LoRA stands for low-rank adaptation and is a method for parameter-efficient adaptation of pretrained neural networks. Rather than changing all weights, it typically freezes the base weights and trains smaller low-rank matrices for selected layers. The resulting adapters require substantially fewer trainable parameters and less storage than full fine-tuning.

The method was introduced for large language models and is now also used with image and multimodal models. A LoRA file is not a complete model: it requires a compatible base, suitable integration, and often a chosen activation weight. Small file size guarantees neither quality nor safe generalisation. Training data, consent, personality and copyright rights, and the licences of both base and adapter require review.

  • Parameter-efficient adaptation method
  • Base weights typically remain unchanged
  • Trains small low-rank matrices
  • Adapter requires a compatible base model
  • Still assess data rights and licences

Frequently asked questions

A LoRA is a compact set of trained adaptation parameters that modifies the behaviour of a compatible base model.

No. A LoRA adapter contains only adaptation parameters and cannot operate without a compatible architecture and base weights.
Zum Glossar