◆ LoRA Trade-off Playground

rank × learning rate × iterations — feel the capacity/speed/overfit trade-off

Try this 123

rank r

learning rate

iterations 50 / 200

real lab stopped at 50 · beyond is extrapolated

train loss—

val loss—

train − val gap—

adapter size —

overfit signature healthy

train loss val loss iter 50

Training trace — what mlx_lm.lora would print ⓘ