Trained on a subset of ~1GB from the msmarco dataset "Train Triples Small".
Batch size: 32