For single-character LoRA training, rank 16 and approximately 3,000 steps are recommended, with a sample generated every 250 steps; the optimal version typically appears between steps 1,000 and 2,000 rather than at the final step.