- Document ranking
- Passage ranking
- Step-by-step notebooks to reproduce our run submitted
to the MS MARCO leaderboard in December 2020 (see also our write-up):
- One notebook reproduces all steps necessary to download the data, preprocess it, and train all the models.
- The second notebook operates on preprocessed data in FlexNeuART JSONL format. It does not require running GIZA to generate IBM Model 1 (these models are already trained).