MS MARCO Passages Hard Negatives

This repository contains raw datasets, all of which have also been formatted for easy training in the MS MARCO Mined Triplets collection. We recommend looking there first.

MS MARCO is a large scale information retrieval corpus that was created based on real user search queries using Bing search engine.

This dataset repository contains files that are helpful to train bi-encoder models e.g. using sentence-transformers.

Training Code

You can find here an example how these files can be used to train bi-encoders: SBERT.net - MS MARCO - MarginMSE

cross-encoder-ms-marco-MiniLM-L-6-v2-scores.pkl.gz

This is a pickled dictionary in the format: scores[qid][pid] -> cross_encoder_score

It contains 160 million cross-encoder scores for (query, paragraph) pairs using the cross-encoder/ms-marco-MiniLM-L-6-v2 model.

msmarco-hard-negatives.jsonl.gz

This is a jsonl file: Each line is a JSON object. It has the following format:

{"qid": 867436, "pos": [5238393], "neg": {"bm25": [...], ...}}

qid is the query-ID from MS MARCO, pos is a list with paragraph IDs for positive passages. neg is a dictionary where we mined hard negatives using different (mainly dense retrieval) systems.

It contains hard negatives mined from BM25 (using ElasticSearch) and the following dense models:

msmarco-distilbert-base-tas-b
msmarco-distilbert-base-v3
msmarco-MiniLM-L-6-v3
distilbert-margin_mse-cls-dot-v2
distilbert-margin_mse-cls-dot-v1
distilbert-margin_mse-mean-dot-v1
mpnet-margin_mse-mean-v1
co-condenser-margin_mse-cls-v1
distilbert-margin_mse-mnrl-mean-v1
distilbert-margin_mse-sym_mnrl-mean-v1
distilbert-margin_mse-sym_mnrl-mean-v2
co-condenser-margin_mse-sym_mnrl-mean-v1

From each system, 50 most similar paragraphs were mined for a given query.