KuaiSearch is a large-scale e-commerce search dataset and full-stack benchmark system built from real user search interactions on the Kuaishou platform. It covers the three core stages of modern industrial search pipelines: Recall, Relevance, and Ranking. Each stage provides multiple algorithmic baselines, allowing researchers to systematically evaluate and compare methods

📄 Paper: KuaiSearch: A Large-Scale E-Commerce Search Dataset for Recall, Ranking, and Relevance Yupeng Li*, Ben Chen*, Mingyue Cheng, Zhiding Liu, Xuxin Zhang, Chenyi Lei, Wenwu Ou

🤗 Dataset: huggingface.co/datasets/benchen4395/KuaiSearch

• Overview
• Dataset Statistics
• Installation
• Data Preparation
• Usage
  • Recall
  • Relevance
  • Ranking
• Supported Models
• Benchmark Results
• Notes
• Citation
• References

KuaiSearch provides a large-scale e-commerce search dataset together with a complete benchmark system covering three modular, independently trainable stages:

KuaiSearch is, to the best of our knowledge, the largest e-commerce search dataset currently available, built upon real user search interactions from the Kuaishou platform. It retains authentic user queries and natural-language product texts, covers cold-start users and long-tail products, and spans all three key stages of the search pipeline.

KuaiSearch-Lite is a lightweight subset designed for rapid model validation and ablation studies. All experiments in the paper are conducted on KuaiSearch-Lite.

• Python 3.8+
• CUDA 11.7+

pip install -r requirements.txt

The dataset is publicly available on HuggingFace:

# Install HuggingFace Hub CLI (if not already installed)
pip install huggingface_hub

# Download KuaiSearch dataset
python -c "
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='benchen4395/KuaiSearch',
    repo_type='dataset',
    local_dir='./data'
)
"

Or manually download from: https://huggingface.co/datasets/benchen4395/KuaiSearch

⚠️ All commands must be run from the KuaiSearch/ project root.

bash scripts/recall_data_process.sh

bash scripts/recall_bm25_eval.sh

bash scripts/recall_doc2query.sh
bash scripts/recall_docT5query_eval.sh

bash scripts/recall_dpr.sh

bash scripts/recall_gr.sh

bash scripts/relevance_data_process.sh

bash scripts/relevance_crossencoder.sh

bash scripts/relevance_embedding.sh

bash scripts/relevance_gr.sh

bash scripts/ranking_data_process.sh

# Default model: DCNv1
bash scripts/ranking_train.sh

If you use KuaiSearch in your research, please cite our paper:

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