KeyError: 'padded_images' when training

hello, when i train on my datasets i meet the error. The datasets only have unified_edit and vlm_sft data.
[rank5]: Traceback (most recent call last):
[rank5]: File "/root/paddlejob/workspace/env_run/Bagel/train/pretrain_unified_navit.py", line 730, in
[rank5]: main()
[rank5]: File "/root/paddlejob/workspace/env_run/Bagel/train/pretrain_unified_navit.py", line 594, in main
[rank5]: data['padded_latent'] = vae_model.encode(data.pop('padded_images'))
[rank5]: KeyError: 'padded_images

Start in train/pretrain_unified_navit.py at main around line 594, where the reported KeyError occurs, and inspect how unified_edit and vlm_sft datasets are assembled before that point. Reproduce training with the reported dataset composition and trace the available data keys. Done means training no longer fails on the missing padded_images key for these datasets.

Written by the indexing model from the issue text.