Repository files navigation

KodCode is the largest fully-synthetic open-source dataset providing verifiable solutions and tests for coding tasks. It contains 12 distinct subsets spanning various domains (from algorithmic to package-specific knowledge) and difficulty levels (from basic coding exercises to interview and competitive programming challenges). KodCode is designed for both supervised fine-tuning (SFT) and RL tuning.

🏆 KodCode has been accepted to ACL 2025 and received theBest Paper Award at DataWorld @ ICML 2025!

• 🕸️Project Website - To discover the reasoning for the name of KodCode 🤨
• 📄Technical Report - Discover the methodology and technical details behind KodCode
• 💾Github Repo - Access the complete pipeline used to produce KodCode V1
• 🤗 HF Datasets:KodCode-V1 (For RL);KodCode-V1-SFT-R1 (for SFT)

KodCode is a comprehensive pipeline designed to generate diverse, challenging, and verifiable synthetic datasets for coding tasks. Key features include:

• Diverse Sources: Generate high-quality coding questions from multiple sources including zero-shot generation, human-written assessment questions, code snippets, and technical documentation -all unified in a single framework!
• Self-Verification: Generate verifiable solutions and tests for each coding question. Support pytest and parallel execution.
• Style Converter: Easy to convert between different styles of coding questions.

1. Build code generation environment

Option 1: conda

git clone https://github.com/KodCode-AI/kodcode.gitcd kodcodeconda create -n kodcode python=3.10 -yconda activate kodcodepip install -r requirements.txt

Option 2:uv

git clone https://github.com/KodCode-AI/kodcode.gitcd kodcodeuv venvsource .venv/bin/activateuv pip install -r requirements.txt

2. Build code execution environment

Option 1: Local

To run unit tests in parallel, you also need to installparallel. For example, if you are using Ubuntu, you can installparallel by:

sudo apt-get install parallel

Option 2: Docker

Please installNvidia container toolkit first to support GPU.

We provided aoff-the-shelf docker image for running tests:

docker pull zcxu/kodcode-test-environment:python3.10-cuda12.4-v0.1

Please refer to thepipeline for details.

SFT: We usedLlama-Factory to train the SFT checkpoint.

RL: Please refer tocode-r1 for RL training using KodCode datasets, which is based onverl.

• One-line command to generate KodCode
• Integrate the test pipeline (i.e.,pytest) for RL training. -> Supported in forkedcode-r1 with the latestverl! Thank you Jiawei @ganler!!!
• Implement dockerized execution for unit tests

• KodCode-Lite with 10K samples for light-weight RL training
• KodCode-V1.1: Supportstdin format with ~50K additional samples

License: Please followCC BY-NC 4.0.

Contact: For questions, suggestions, or feedback, please reach out toZhangchen Xu, orraise an issue. We welcome your input and are committed to continuously improving KodCode to better serve the community.

If you find the model, data, or code useful, please cite:

@article{xu2025kodcode,      title={KodCode: A Diverse, Challenging, and Verifiable Synthetic Dataset for Coding},       author={Zhangchen Xu and Yang Liu and Yueqin Yin and Mingyuan Zhou and Radha Poovendran},      year={2025},      eprint={2503.02951},      archivePrefix={arXiv},      primaryClass={cs.LG},      url={https://arxiv.org/abs/2503.02951}, }