A public-library bindery already agreed to start from a pretrained checkpoint. The remaining argument is whether to update a small PEFT adapter or the full weights on a modest GPU night. How should that argument be settled?
Select an answer to reveal the explanation.
Short Explanation
The checkpoint is already chosen. The remaining fight is a small PEFT adapter versus a full-weight update on a modest GPU night. Compare quality, memory, and overfit risk. PEFT is not fake training, and it does not wait on an NCCL kernel.
Full Explanation
After transfer is chosen, PEFT versus a full weight update is the next measured comparison. The readout is quality on the same val set, plus memory use and overfit risk on a modest GPU night. A small adapter is a real adaptation method, not a lesser substitute that can be dismissed without a run. Collectives and an invented ban on prompt learning are not the decision rule.