Keep thinking ability after fine-tuning

Hi authors,

Thanks for your great work again!
I'm wondering if the model can keep the thinking ability after fine-tuning, say 15k steps, only fine-tuned on T2I task.
Based on my current experiments, it seems not able to preseve the thinking ability and it will appear many 'vr
vr
vr
vr
vr' when I use generation with think. (no matter set freeze_und=True or False)

Do you think adding the I2T image to text task will preserve the think ability after fine-tuning?
Or What should I do if I want to keep thinking ability?

Thanks a lot!

Best,
Minghui

The report concerns fine-tuning for 15k steps on a T2I task, the freeze_und setting, and generation with think; no source files or tests are identified. Start by reviewing the fine-tuning and generation entry points, reproduce the repeated “vr” output, and compare behavior with and without the I2T task to establish what a fix or documented limitation would require.

Written by the indexing model from the issue text.