Match LoRA rank to task type

**LoRA** fine-tunes a model by adding a small pair of trainable matrices alongside the frozen original weights. The rank is the shared inner dimension of those two matrices, and it controls how much new behavior the adapter can express. Think of rank as a capacity dial. A low rank gives the adapter a thin slice of expressive room, enough to nudge how the model phrases things. A high rank gives it far more room, enough to reshape how the model reasons about a whole new domain. The matching follows that logic. Light tone or style tuning sits at the low end. Standard instruction tuning on a few thousand examples is the comfortable default in the middle. Heavy domain adaptation, like medical or legal text with new vocabulary, needs more capacity. The catch is at the top. Past a certain point, extra rank stops helping. The parameter count balloons toward full fine-tuning, the model starts memorising your training set, and you lose the lightweight storage benefit that made LoRA appealing. So you pick the smallest rank that does the job, not the biggest one you can afford.

**LoRA rank** is the capacity knob of the adapter. It is the inner dimension shared by the two low-rank matrices that approximate the weight update. More rank means more expressive freedom, and more parameters to train and store. The mapping tracks how aggressively you reshape the model. A rank near 4 fits light style or tone tuning, where you change phrasing but not knowledge. A rank near 16 is the standard default for instruction tuning on a few thousand examples. A rank near 64 suits real domain adaptation with vocabulary shift, like a medical or legal corpus. A rank near 256 handles aggressive domain shift or multi-task adapters, edging toward full fine-tuning capacity. The relationship breaks at the top. By rank 512 and beyond you pay full fine-tuning compute without its expressive freedom. The adapter also tends to overfit on the directions the higher rank newly exposes. So the practical sweet spot sits well below the apparent ceiling. Pick the smallest rank that closes your quality gap, then scale up only if evaluation says you must.