license: apache-2.0
task_categories:
- other
tags:
- recommendation
- ranking
- sequential-recommendation
UniRank: Benchmarking Ranking Models for Unified Sequential Modeling and Feature Interaction
This repository contains dataset artifacts for UniRank, an open benchmark for unified sequential modeling and feature interaction in large-scale recommendation ranking.
- Paper: UniRank: Benchmarking Ranking Models for Unified Sequential Modeling and Feature Interaction
- GitHub Repository: salmon1802/UniRank
Dataset Summary
UniRank standardizes chronological point-wise autoregressive supervision, multi-feedback evaluation, model implementations, data processing, and distributed training in one reproducible pipeline.
The benchmark evaluates unified ranking architectures on five large-scale industrial datasets from short-video, advertising, and e-commerce scenarios:
| Dataset | Instances | Users | Items | Fields | Tasks | Avg. length | Max. length |
|---|---|---|---|---|---|---|---|
| QK-Video | 493,306,303 | 4,996,176 | 3,752,235 | 10 | 4 | 99 | 6,013 |
| KuaiRand | 323,464,444 | 27,285 | 32,038,725 | 40 | 6 | 11,855 | 228,030 |
| TAAC-25 | 757,207,146 | 7,706,778 | 15,707,425 | 30 | 2 | 98 | 100 |
| Taobao | 23,601,301 | 470,570 | 831,643 | 23 | 4 | 50 | 3,756 |
| MerRec | 172,304,959 | 1,697,072 | 42,577,610 | 20 | 5 | 102 | 26,576 |
Download and Usage
The recommended workflow is to download a ready-to-use dataset from Hugging Face to local storage. For example, to download the KuaiRand dataset:
hf download salmon1802/KuaiRand \
--repo-type dataset \
--local-dir /path/to/data/KuaiRand_Video_Action
Please refer to the GitHub Repository for preprocessing scripts, model configurations, and training instructions.
Citation
If you find this dataset or benchmark useful, please cite the paper:
@article{unirank2026,
title={UniRank: Benchmarking Ranking Models for Unified Sequential Modeling and Feature Interaction},
author={UniRank Authors},
journal={arXiv preprint arXiv:2607.19987},
year={2026}
}