ExpandR dense retriever training data

This dataset contains 708,740 contrastive training examples for dense passage retrieval. Each row pairs a search query and its generated pseudo-document with positive and negative retrieval examples used by ExpandR.

What can be compared with the same examples?

The release stores query and pseudo_doc in separate fields. A retriever experiment can therefore vary whether the expansion is supplied while retaining the same pos and neg examples. Use the same evaluation queries, corpus and retrieval cutoff to compare the resulting rankings.

For a generator experiment that needs two competing expansions for one prompt, use the companion preference pairs.

Data files

Hugging Face dataset card · The downloads below use a fixed revision.

Related paper

ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance · Official BibTeX

Browse research data · Download catalog (JSON-LD)