Hello,
I followed the tutorial and used Qwen2.5VL 7B as the base model. I conducted pre-training on the SVGXCore 250K dataset, followed by supervised fine-tuning (SFT) on the SVGX1M dataset, with a learning rate of 3e-4. During testing, I observed that the generated images consistently appear concentrated in the top-left corner of the output.

Is this behavior expected, or could it indicate an issue with the training process or configuration? Have I potentially overlooked any critical steps or parameters during pre-training or fine-tuning? Any guidance would be greatly appreciated. Thank you!