Reproducing traditional run on the MS MARCO leaderboard
·
929 commits
to master
since this release
This short documents describes a traditional IR system that achieved MRR@100 equal to 0.298 on the MS MARCO leaderboard (on 2020-12-06). Although inferior to most BERT-based models, it outperformed five neural submissions,
including two submissions that used a large pretrained Transformer model for re-ranking. We provide software and data to reproduce our results.