KuaiSearch

Parquet republication of KuaiSearch for the Lizard benchmark. This is not the official repository. full_* configs are the full release. lite_* configs are KuaiSearch-Lite, the subset used for the paper's experiments. Lite is not a split of the full tables.

rank, recall, and relevance use the released split column. Its values are train and test; there is no validation split. The column is kept inside the rows. users and items have no released split, so each of those configs has one split named train that contains every row.

Original sources

• Upstream dataset: https://huggingface.co/datasets/benchen4395/KuaiSearch
• GitHub: https://github.com/benchen4395/KuaiSearch
• Paper: https://arxiv.org/abs/2602.11518

JSON values are preserved: integers in the signed 64-bit range are BIGINT, other exact JSON numbers are DOUBLE, strings are VARCHAR, integer lists stay lists, and the released statistical objects stay structs. JSON null stays null. time_index remains the source ordinal and is not converted to a timestamp. source_row is the 1-based JSONL line number, assigned before sorting. No history, label, or join was derived.

License

MIT, following the Hugging Face dataset card of the upstream release from which these files were obtained. The GitHub repository badge says CC BY 4.0.

Citation

@article{li2026kuaisearch,
  title = {KuaiSearch: A Large-Scale E-Commerce Search Dataset for Recall, Ranking, and Relevance},
  author = {Li, Yupeng and Chen, Ben and Cheng, Mingyue and Liu, Zhiding and Zhang, Xuxin and Lei, Chenyi and Ou, Wenwu},
  journal = {arXiv preprint arXiv:2602.11518},
  year = {2026},
  url = {https://arxiv.org/abs/2602.11518}
}