Launch the conversion script (Use --dataset=ilsvrc_2012_v2 for the training only MetaDataset-v2 version):
python -m meta_dataset.dataset_conversion.convert_datasets_to_records \
--dataset=ilsvrc_2012 \
--ilsvrc_2012_data_root=$DATASRC/ILSVRC2012_img_train \
--splits_root=$SPLITS \
--records_root=$RECORDS
did u run this?
https://github.com/google-research/meta-dataset/blob/main/doc/dataset_conversion.md#ilsvrc_2012