Fine-tuning / peft
Parameter-efficient methods
Full fine-tuning updates every weight and needs optimiser state for all of them. The methods here constrain the update to a low-rank or quantised form, which makes tuning cheap, adapters portable, and merging several of them into one model a question with an actual answer.
4 entries, 1 of them a stub.
Entries
03.02.1O(d)03.02.2O(r·d)03.02.3O(r·d)03.02.4O(r·d)
Adapter Mergingstubcommon
Task arithmetic, TIES and DARE over independently trained deltas.
DoRAcommon
Low-rank adaptation split into a magnitude and a direction.
LoRAstandard
Low-rank additive update to frozen weight matrices during adaptation.
QLoRAstandard
4-bit NF4 base weights with paged optimiser state and LoRA adapters.