How to construct my own training data

[hanzifan]

Thanks for your great work. I have a question about how to contrusct my own training data for onevision-encoder. For example, i have some ms-swift style data, how can i transfer them into onevision-encoder style.

[Luodian]

check out the data format docs in scripts/train/ - there should be example configs showing the expected JSON structure.

the short version: each sample needs an image/video path and the associated text. if your ms-swift data has the standard conversation format, the main thing you need to do is map it to our expected schema. take a look at the dataset loading code in the training scripts for the exact field names.

if you can share a snippet of your ms-swift data format i can point you to the specific conversion needed.