ZIP-0434: Decentralized Training Infrastructure

Abstract

This proposal specifies Zoo Gym, the production infrastructure that unifies all previously-specified decentralized training components -- DSO (ZIP-0410), PoAI (ZIP-0419), federated learning (ZIP-0424), and GRPO (ZIP-0421) -- into a single, deployable training platform. Zoo Gym is the "gym" where AI models train: a decentralized network of compute nodes that collaboratively improve Zen models, earn ZOO token rewards, and produce verifiable training proofs anchored on-chain.

Motivation

The preceding ZIPs specified individual components:

• ZIP-0407: Decentralized training architecture (theory)
• ZIP-0410: DSO protocol (gradient exchange)
• ZIP-0419: PoAI consensus (validation)
• ZIP-0421: GRPO (preference optimization)
• ZIP-0424: Federated wildlife monitoring (conservation-specific)

Zoo Gym combines these into a production system that conservation organizations, researchers, and AI enthusiasts can actually run, managing the complexity of node registration, task assignment, gradient routing, verification, reward distribution, and model checkpoint management.

Specification

System Architecture

Zoo Gym Platform
├── Coordinator Layer (smart contracts)
│   ├── NodeRegistry.sol: node registration and staking
│   ├── TaskManager.sol: training task lifecycle
│   ├── RewardDistributor.sol: ZOO token rewards
│   └── ModelRegistry.sol: model version tracking
│
├── Training Layer (off-chain, distributed)
│   ├── DSO Engine: semantic gradient exchange (ZIP-0410)
│   ├── GRPO Engine: preference optimization (ZIP-0421)
│   ├── Federated Engine: federated learning (ZIP-0424)
│   └── Continuous Training: ongoing model improvement
│
├── Verification Layer (hybrid)
│   ├── PoAI Validators: AI work verification (ZIP-0419)
│   ├── Compute Proofs: cryptographic work attestation
│   └── Spot Checking: random recomputation of training steps
│
└── Orchestration Layer (off-chain)
    ├── Task Scheduler: matches tasks to nodes
    ├── Checkpoint Manager: model version control on IPFS
    ├── Metrics Dashboard: training progress monitoring
    └── Health Monitor: node uptime and quality tracking

Training Task Types

Task	Description	Reward	Duration
Pre-training	Continue pre-training on new data	10 ZOO/step	Days-weeks
Fine-tuning	Domain-specific fine-tuning	5 ZOO/step	Hours-days
GRPO	Preference optimization round	8 ZOO/round	Hours
Federated	Wildlife monitoring model update	3 ZOO/round	Minutes-hours
Evaluation	Benchmark evaluation of new checkpoint	2 ZOO/eval	Minutes

Node Economics

Revenue per node per month (estimated):
├── Training rewards: 500-5000 ZOO (hardware-dependent)
├── Inference serving: 200-2000 ZOO (traffic-dependent)
├── Verification: 100-500 ZOO (assignment-dependent)
└── Staking yield: 3-5% APY on staked ZOO

Deployment

Node operators run a single binary:

zoo-gym start \
  --stake 10000 \
  --gpu auto \
  --tasks training,inference,verification \
  --region us-east

Research Papers

• zoo-gym-protocol -- Gym protocol specification (2024)
• zoo-gym-compute-proof -- Compute proof protocol (2024)
• zoo-gym-orchestrator -- Training orchestration (2024)
• zoo-gym-tokenomics -- Gym token economics (2024)
• zoo-gym-grpo-continuous -- Continuous GRPO training (2024)
• zen-distributed-training -- Distributed training for Zen

Implementation

• hanzo/node: Blockchain/AI node with Gym training support
• hanzo/candle: Rust ML framework for training workloads
• zoo/contracts: Gym smart contracts for coordination and rewards
• zoo/core: Gym dashboard and monitoring interface

Timeline

• Originated: September 2025 (Zoo Gym production design)
• Research: Gym paper series published 2024 (protocol, compute proof, orchestrator, tokenomics, continuous GRPO)
• Implementation: Zoo Gym network launched 2025