The metadata file emitted has:
"dataset": {
"data_folder": "data/unet3d/",
"format": "npz",
"num_files_train": 168,
"num_samples_per_file": 1,
"record_length_bytes": 146600628,
"record_length_bytes_stdev": 68341808,
"record_length_bytes_resize": 2097152
},
However the output said:
2025-10-17 08:01:39|RESULT: Number of training files: 28000
I do not see 28000 in the datasize metadata result file. So there no way to fill in workload.dataset.num_files_train for datagen using the metadata file for automation.