Public datasets for research on Learning to Rank and comparison of traditional and neural models. Review descriptions, publications to cite and license terms before downloading.
Istella22 enables comparison of feature-based Learning to Rank methods and neural approaches based on pre-trained language models. It provides query and document text together with industrial query-document feature vectors.
Details and download →The full Istella Learning to Rank dataset was used to train a stage of the Istella production ranking pipeline. It supports large-scale experiments on ranking effectiveness, efficiency and scalability.
Details and download →Istella-S LETOR is a smaller sample of the Istella Learning to Rank dataset, obtained by sampling irrelevant query-document pairs. It is suitable for experiments requiring a more compact benchmark.
Details and download →Istella-X (eXtended) LETOR is an extended Learning to Rank benchmark obtained by retrieving up to 5,000 documents per query using the BM25F ranking score.
Details and download →