Allow for dataset to be loaded from S3 bucket. Also train/valid split

**Is your feature request related to a problem? Please describe.**
The Helmet datasets are created from scripts, using random choices. Also, train/valid splits of dev sets are random. This introduces variance in results when training/eval is done on different instances, or between ours and baseline.

**Describe the solution you'd like**
Ideally, we just create the datasets and train/valid splits on S3 and load them from there, so they are always the same.