hanzoai/zen-agentic-dataset

Zen Agentic Dataset是一个包含约120亿令牌的真实世界代理AI编程、区块链开发和前沿基础设施代码的综合性训练数据集。它整合了Claude Code交互和来自1500多个仓库的完整git历史，覆盖了15年的专业开发时间跨度（2010-2026年）。数据集包含代理AI与LLM基础设施、Web3与区块链、密码学与安全、现代开发等多个领域的代码，语言涵盖Python、TypeScript、JavaScript、Rust、Go等核心和基础设施语言，以及Solidity、C/C++等专业语言。数据经过去重和压缩处理，分片存储，并已进行PII清理和品牌重命名。该数据集用于训练如Zen Coder 4B、Zen Coder Flash等模型，适用于研究和商业许可。Zen Agentic Dataset is a comprehensive training dataset containing approximately 12 billion tokens of real-world agentic AI programming, blockchain development, and cutting-edge infrastructure code. It combines Claude Code interactions with full git history from over 1,500 repositories spanning 15 years of professional development (2010-2026). The dataset covers domains such as agentic AI & LLM infrastructure, Web3 & blockchain, cryptography & security, and modern development, with languages including Python, TypeScript, JavaScript, Rust, Go (core and infrastructure), and specialized languages like Solidity and C/C++. Data is deduplicated, compressed, sharded, and has undergone PII scrubbing and brand-aware rebranding. It is used to train models like Zen Coder 4B and Zen Coder Flash, and is available for research and commercial licensing.