task_categories:
- text-retrieval
language:
- en
tags:
- click-model
- information-retrieval
CLAX Datasets
This repository hosts the Yandex and Baidu-ULTR datasets, which are essential for research and development with CLAX: Fast and Flexible Neural Click Models in JAX. CLAX is a JAX-based library that implements classic click models using modern gradient-based optimization to understand user behavior and improve ranking performance at scale.
Links
- Paper: CLAX: Fast and Flexible Neural Click Models in JAX
- Code (CLAX library): https://github.com/philipphager/clax
- Project page (CLAX library documentation): https://philipphager.github.io/clax/
Dataset Description
The datasets provided here, including the Yandex and the massive Baidu-ULTR dataset (comprising over a billion user sessions), serve as crucial benchmarks for training and evaluating click models. They enable practitioners and researchers to leverage modern deep learning frameworks while preserving the interpretability of classic models. These datasets are fundamental for developing and testing models that analyze user behavior in search and recommendation systems.
Download Datasets
To download these datasets, ensure you have Git LFS installed. Then, clone the repository:
git lfs install
git clone https://huggingface.co/datasets/philipphager/clax-datasets
Note: The full datasets require approximately 85GB of disk space. By default, the CLAX library expects datasets at ./clax-datasets/ relative to the project root. To use a custom path, you can update the dataset_dir parameter in your experiment's config.yaml within the CLAX library code, for example:
dataset_dir: /my/custom/path/to/datasets/
Sample Usage
The datasets in this repository are intended to be used with the CLAX library for training and evaluating click models. Below is a basic example from the CLAX GitHub repository demonstrating how a UserBrowsingModel can be trained:
First, install CLAX (requires JAX with CUDA support; refer to JAX documentation for installation):
pip install clax-models
Then, you can use the following Python snippet:
from clax import Trainer, UserBrowsingModel
from flax import nnx
from optax import adamw
model = UserBrowsingModel(
query_doc_pairs=100_000_000, # Number of query-document pairs in the dataset
positions=10, # Number of ranks per result page
rngs=nnx.Rngs(42), # NNX random number generator
)
trainer = Trainer(
optimizer=adamw(0.003),
epochs=50,
)
train_df = trainer.train(model, train_loader, val_loader)
test_df = trainer.test(model, test_loader)
For more advanced usage, custom modules, and additional examples, please refer to the CLAX GitHub repository.
Citation
If CLAX or these datasets are useful to you, please consider citing our paper:
@misc{hager2025clax,
title = {CLAX: Fast and Flexible Neural Click Models in JAX},
author = {Philipp Hager and Onno Zoeter and Maarten de Rijke},
year = {2025},
booktitle = {arxiv}
}