LoRA-conditioned model skips the first line of its output when the caption is short. Any idea why? [D]

I'm doing an MSc project on BrickNet (kulits/BrickNet on GitHub, CVPR 2026). It writes LEGO models as plain text and fine-tunes Qwen3 (I'm using 0.6B) with LoRA: a PT adapter that learns the text format, and an SFT adapt