CPRet-data

This repository hosts the datasets for CPRet: A Dataset, Benchmark, and Model for Retrieval in Competitive Programming.

Visit https://cpret.online/ to try out CPRet in action for competitive programming problem retrieval.

💡 CPRet Benchmark Tasks

The CPRet dataset supports four retrieval tasks relevant to competitive programming:

1. Text-to-Code Retrieval Retrieve relevant code snippets based on a natural language problem description.

2. Code-to-Code Retrieval Retrieve semantically similar code snippets, often corresponding to related problems.

3. Problem-to-Duplicate Retrieval Retrieve near-duplicate or reused problems.

4. Simplified-to-Full Retrieval Retrieve the full original problem descriptions based on simplified or abridged versions.

💾 Contents

This repository includes all datasets used in the CPRet project, organized as follows:

• Training datasets Used to train retrieval models with contrastive or supervised learning strategies.

• Evaluation datasets Benchmarks for the four retrieval tasks above, with relevance annotations.

🚀 How to Use

🔧 Code

To use these datasets for training or evaluation, please refer to the main CPRet repository: 👉 CPRet GitHub

It provides:

• Preprocessing and training scripts
• Evaluation pipelines for all retrieval tasks

🤗 Hugging Face Collection

All CPRet-related resources are available in a unified Hugging Face Collection: 👉 CPRet on Hugging Face

This includes:

• 🔹 Trained models (CPRetriever-Code, CPRetriever-Prob, etc.)
• 🔹 Datasets for training and benchmarking
• 🔹 Precomputed embeddings along with full problem descriptions, titles, platforms, and URLs, supporting plug-and-play deployment of retrieval services.

📜 Citation

If you use CPRet in your work, please cite:

@misc{deng2025cpretdatasetbenchmarkmodel,
  title     = {CPRet: A Dataset, Benchmark, and Model for Retrieval in Competitive Programming},
  author    = {Han Deng and Yuan Meng and Shixiang Tang and Wanli Ouyang and Xinzhu Ma},
  year      = {2025},
  eprint    = {2505.12925},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url       = {https://arxiv.org/abs/2505.12925}
}