Question about Training Parameters for SoftFold Cloth-Folding

[yi0kevin]

Hello again,

I’m preparing to fine-tune the X-VLA-SoftFold in a new environment for cloth-folding and have a few questions about the training setup, mainly related to two files:

1.domain_config.py​
What were the DATA_WEIGHTS and DATA_DOMAIN_ID settings used for X-VLA-SoftFold training?

2.agiworld.py​
I’ve implemented a data handler based on AGIWorldHandler. Could you share the DOMAIN2CHUNKSIZE and DOMAIN2INS values used during training? Also, lines 186–193 of the file describe a 50% probability rule for selecting multi-view images—was the same strategy applied to SoftFold tasks?

Are there any other important training details or hyperparameters I should be aware of specifically for the SoftFold task?
Thank you for your time and help!

[xspadex]

Have you solved the problem?
I found the model doesn't work in new environment.

[yi0kevin]

Have you solved the problem? I found the model doesn't work in new environment.

No, still trying. I am doing fintuning but the result is not good.