Pick a LoRA rank for a domain task and justify why r=512 isn't a free upgrade

**LoRA** fine-tunes a model by freezing the original weights and training two small matrices whose product is added back. The rank, written r, sets how big those matrices are. Higher rank means more trainable parameters and more capacity. For 30k legal-clause examples on a 7B model, start at r=32 or r=64. Legal text is a real domain shift, so it needs more capacity than a light style tweak. But 30k examples is a modest dataset, so you do not want a huge adapter that just memorises it. The person asking for r=512 thinks bigger is safer. It is not. At r=512 you train so many parameters that you lose the cheapness and easy swapping that made LoRA attractive. You also give the model enough room to memorise quirks, which hurts on new clauses. There is a subtler trap too. LoRA scales its update by alpha divided by r. If you raise r without raising alpha, each update step shrinks. So a bigger rank can actually learn less per step and look worse. The safe move is a small sweep. Try r values like 16, 32, 64, and 128, evaluate on held-out clauses, and keep the smallest rank that hits your quality bar.

For 30k legal-clause examples on a 7B base, I would start at **r=32** and test r=64. Legal language is a genuine domain shift, so it needs more capacity than a style tweak at r=4-8. But 30k examples is modest, so r=64 is plenty without inviting overfitting. The r=512 argument fails on three fronts. First, parameter count. At r=512 the trainable params approach a real fraction of full fine-tuning, so you lose the storage and adapter-swap wins that motivate LoRA. Second, overfitting. More rank means more capacity to memorise 30k idiosyncrasies, and held-out benchmarks regress. Third, alpha scaling. The effective update multiplier is alpha over r, so raising r alone shrinks every step. The disciplined move is a rank sweep. Train r in {16, 32, 64, 128} on a held-out clause set, keep alpha tied to r, and pick the smallest rank that hits target. That leaves headroom to grow rank only if a larger future dataset justifies it.