ExpandR query expansion preference data
This dataset contains 30,000 preference pairs for training a language model to expand search queries. Each conversation is paired with a preferred and a rejected response for direct preference optimization in ExpandR.
How is an expansion represented?
One row contains conversations, chosen and rejected. The conversation holds the input to the generator. The latter two fields give alternative responses to that input. Use their order as the preference label when training a query expansion model.
To evaluate whether an expansion helps search, compare retrieval with the original query and the expanded query against the same corpus and relevance judgments. For examples that supply relevant and negative passages alongside a generated expansion, use the companion retriever data.
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