What configuration was used to compare the training performance of oft and lora on the qwen3-30b-a3b model?

[deepllz]

As shown in the image in the article（https://spherelab.ai/orbit/#qwen3），
Is the training configuration based on https://github.com/Sphere-AI-Lab/orbit/blob/main/examples/high_precision/run-qwen3-30b-a3b-bf16-openr1-oft-b64.sh from the repository? I usually try to reproduce this using Slime, but the results drop after a few dozen training steps. Do you have any suggestions?