---
license: apache-2.0
task_categories:
- text-ranking
tags:
- math
- formalization
- lean
library_name: datasets
---

# Premise Retrieval Dataset for Mathematical Formalization

This dataset is used in the paper [Learning an Effective Premise Retrieval Model for Efficient Mathematical Formalization](https://huggingface.co/papers/2501.13959). It contains data for training and evaluating a lightweight and effective premise retrieval model for the Lean theorem prover.

## About the Dataset

The dataset consists of proof states (acting as queries) and corresponding premises extracted from the Mathlib library. It is designed to facilitate the training of models using a contrastive learning framework to embed queries and premises in a latent space. This process aims to enhance retrieval performance through fine-grained similarity calculation and a re-ranking module, ultimately assisting users in the mathematical formalization process.

## Links

*   **Paper:** [Learning an Effective Premise Retrieval Model for Efficient Mathematical Formalization](https://huggingface.co/papers/2501.13959)
*   **Project Page:** https://premise-search.com/
*   **Code:** The source code and trained models can be found on the [GitHub repository](https://github.com/ruc-ai4math/Premise-Retrieval).

## Dataset Access

This dataset is available for download at [this link](https://huggingface.co/datasets/ruc-ai4math/mathlib_handler_benchmark_410) on the Hugging Face Hub.

## Sample Usage

You can load the dataset using the Hugging Face `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("ruc-ai4math/mathlib_handler_benchmark_410")

# To inspect the 'train' split
print(dataset["train"][0])
```