| Dimension | Adapter-based approach | Full fine-tuning | What the buyer should request |
|---|---|---|---|
| Trainable artefact | A limited set of added/updated parameters | Broad updates to the model | Exact saved files and base revision |
| Resource use | Often chosen for lower training memory | Can require substantially greater resources | Measured peak memory, job cost and failed-run cost |
| Adaptation capacity | Depends on rank, placement, data and task | Wider parameter freedom | Controlled comparison if quality is limiting |
| Retention | Must still be tested | Must still be tested | Out-of-task and safety regression set |
| Portability | Adapter must match the base and runtime | Checkpoint still has licence/runtime dependencies | Import and inference rehearsal elsewhere |
QLoRA uses quantisation with low-rank adaptation to reduce training memory. Biderman and colleagues found standard low-rank configurations learned less of certain target domains while forgetting less outside them. These are different studies and questions; one does not invalidate the other. A procurement claim that an adapter is “as good as full fine-tuning” needs task-specific evidence.
Quantisation is also a deployment choice. Retest the actual served precision and runtime rather than assuming the training evaluation transfers unchanged; the PEFT guide explains the relationship between quantisation and adapters but does not certify a buyer’s model.