Training instructions are unclear

[atonalfreerider]

After running extract_all_trajectories.py on a sequence of 50 images (480x640) I have an .npz output that contains the co-tracker data:
-pred_tracks.npy
-pred_visibility.npy
-started_at.npy

The provided training example says to take this data and run:

python train.py --dataset_folder data/pet_test_set/our_data_format_4_validation_rgbd \
         --dataset_folder_validation data/pet_test_set/our_data_format_4_validation_rgbd

this is confusing because the dataset_folder and the dataset_folder_validation point to the same location in the example

I see from the pet_test_set that the npz should contain the track data above in addition to:
-GT_depth_tracks_all.npy
-GT_mask_tracks_all.npy
-Ks_all.npy
-Ms_all.npy

Do we add these to the npz using external tools like Sam2, UniDepth, etc?

I tried running train.py and I am getting a validation loss that is increasingly negative. It is producing a checkpoint .pt every 100 epochs

Am I misunderstanding the process?