Approaches

Low-rank adaptation (LoRA)

Train low-rank adapter matrices for selected model weights.

When to consider it

  • When a supported base model needs a compact trainable adapter

Considerations

  • Adapter quality and memory savings depend on rank, targets, and training setup.

Connections

Sources

  1. PEFT methods overviewHugging Face · checked 2026-09-30

Last reviewed 2026-09-30