Look into sharding for future-proof train/test splits

[roedoejet]

to help preserve train/test splits as new data is added