~12 Billion Tokens of real-world agentic AI programming, blockchain development, and cutting-edge infrastructure code.

A comprehensive training dataset combining Claude Code interactions with full git history from 1,500+ repositories spanning 15 years of professional development.

• +2.57M new Claude Code entries from ~/.claude/projects/ since last extract (Feb 2026)
• Global (sessionId, uuid) dedup across 408 source files — 48.9% reduction
• Date-bucketed shards: 2025-q4, 2026-q1, 2026-q2 (plus catch-all unknown)
• Zstd compression: ~5× ratio (~10 GB compressed shards vs ~41 GB unique uncompressed)
• Brand-aware rebranding: Anthropic → Hanzo, Claude → Zen (corpus reads as native Hanzo ecosystem)
• PII scrubbing: API keys (Anthropic/OpenAI/GitHub/HF/AWS/Google/Slack), JWTs, PEM private keys → placeholders
• Non-ecosystem brand redaction: keeps hanzo, lux, zen, zoo, adnexus intact; redacts non-allowed brands

• Model Context Protocol (MCP) — 260+ tool implementations
• Multi-agent orchestration — Hanzo Zen, GPT-4, Gemini, Qwen
• Agent frameworks — Planning, memory, tool use, reflection
• LLM Gateway — Unified proxy for 100+ providers

• Smart contracts — Solidity, Vyper (ERC20, ERC721, DeFi)
• Consensus engines — Snow family, BFT, DAG protocols, Quasar PQ
• Cross-chain bridges — Multi-VM architecture
• DeFi protocols — AMMs, lending, staking, governance

• Post-quantum cryptography — ML-DSA, ML-KEM, SLH-DSA (NIST PQC)
• Threshold cryptography — MPC, secret sharing, DKG, Corona
• Zero-knowledge proofs — zkSNARKs, zkSTARKs

• Full-stack TypeScript — Next.js 14+, React 18+, Node.js
• Systems programming — Rust, Go, Python, C/C++
• DevOps — Docker, Kubernetes, CI/CD pipelines

Data is split by quarter for manageable per-shard download:

Each entry appears in exactly one shard via global (sessionId, uuid) dedup — zero cross-shard overlap by construction.

Older raw extracts remain in data/claude-latest/, train_chunks/, and root-level claude-interactions*.jsonl files for reference.

This dataset is available for research and commercial licensing.

Email: z@hanzo.ai

Please include:

• Intended use case (training, research, evaluation)
• Organization/affiliation
• Target ecosystem (if applicable)

Unlike synthetic datasets, this contains actual Hanzo Zen / Claude Code sessions showing:

• Real debugging workflows with trial and error
• Complex multi-file refactoring decisions
• Tool use patterns (file ops, search, git, tests)
• Error recovery and iterative refinement
• Multi-step agentic loops with reflection

• Code that shipped to production systems (Hanzo, Lux, Zen, Zoo, Adnexus)
• Security-audited smart contracts
• Performance-optimized infrastructure

@dataset{zen_agentic_dataset_2026,
  author = {Kelling, Zach},
  title = {Zen Agentic Dataset: ~12B Tokens of Agentic AI Programming},
  version = {2026-05},
  year = {2026},
  publisher = {Zoo Labs Foundation},
  url = {https://huggingface.co/datasets/zenlm/zen-agentic-dataset}
}

Maintainer: z@hanzo.ai
Training framework: Zoo Gym
License: Commercial — Contact for terms