Training and evaluation cost

[harrylin-hyl]

Nice work! I'd like to know the training and evaluation cost required to reproduce the experiments — specifically, how many GPUs are needed and how long it takes to run.

[JamesKrW]

Thanks! To reproduce the experiments (Qwen2.5-VL-7B-Instruct), the approximate compute cost is:

• Training: 8 × H100 GPUs, about 80 hours
• Evaluation: 1 × RTX PRO 6000 GPU, about 1 hour
  (4 x RTX 5090 GPUs as environment server)