license: other
license_name: ms-marco-non-commercial-research-license
license_link: LICENSE
The dataset has 3 columns : query, positive_example, negative_example
First we filtered queries.train.tsv to only contain queries which were available in qrels.train.tsv. This was done so that we only take care of queries of which we had a positive example of. Then we found top 200 ranked passages in collection.tsv for each query in filtered queries and then randomly sampled a passage whose rank lied in the range of 50 to 200 to produce our final hard negatives for each query.
Use this as a training dataset. Test your model against qrels.dev.tsv for MRR@10 or a metric of your choice.
The original datasets used to generate this can be found at : https://microsoft.github.io/msmarco/Datasets.html#passage-ranking-dataset.
The author does not promise on the correctness of the dataset
For research purposes only.