Fine-tuning
Adapting a pretrained model without paying to pretrain it. Mostly a study in which parameters you are allowed to leave frozen, and what the frozen ones cost you in precision.
Five entries across two branches. One is a stub and is marked as such — read them as open questions rather than answers.
Supervised fine-tuning
sft/ ›the baseline every other method is measured against
Parameter-efficient methods
peft/ ›low-rank updates and the arithmetic of merging them
03.2O(d)03.3O(r·d)03.4O(r·d)03.5O(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.