Train a tiny add-on instead of the whole model; QLoRA also loads the base in 4-bit.
- FreezeKeep the base weights fixed.
- Add adaptersInsert two small matrices beside each layer.
- TrainUpdate only the adapters on your examples.
- 4-bit base (QLoRA)Load the frozen base in 4-bit.
- Merge or swapFold the adapter in, or swap per task.