ModelGraph: Decentralized Provenance, Incentives and Co-Building of AI Models

The Hetu ModelGraph is a protocol for establishing verifiable contribution graphs in AI model evolution. Breakthroughs in AI models are never isolated innovations. Instead, they built upon previous models, incorporating improvements in model architecture, scale, training techniques such as fine-tuning and reinforcement learning from human feedback, and even extending capabilities to multimodal inputs. However, significance of these critical innovations are lost amid model creation, and the economic returns are often taken by the final powerful models. This results in stifled innovations and unfair value distributions. Moreover, improvements in models are results of contributions from diverse individuals -- user data, manual labeling, GPU computational resources, algorithm development, etc. The Hetu ModelGraph maintains model evolution and contributions in a fully open, trackable, and decentralized manner, enabling transparent and fair value distribution for the AI creation process.

Why ModelGraph Exists

Modern AI development suffers from:

• ❌ Black-box model provenance in collaborative training
• ❌ Unverified contribution claims in decentralized ML
• ❌ Broken incentive alignment between base model creators and derivative works

ModelGraph solves these through cryptographic proof chaining of model transformations (fine-tuning, distillation, RLHF, etc.), creating an on-chain dependency graph that enables:

• ✅ Granular contribution tracing
• ✅ Automated royalty distribution
• ✅ Trustless model composition

Model Provenance Graph

graph LR
    A[Base Model v1.0] -->|Fine-tuning| B[Medical Model v2.1]
    A -->|Distillation| C[Edge Model v1.5]
    B -->|RLHF| D[Clinical Chat Model v3.2]

Loading

GraphFlow

Core Components

Contribution Weighting

• Data: Quality-verified through cross-validation
• Compute: Tracked via verifiable compute units
• Algorithm: Valued by model improvement metrics

Model Fingerprinting System

Implementation Based on Instruction Tuning:

• 🔑 Confidential private key embedded through lightweight instruction tuning
• 📜 Generates license-specific text when detecting key phrases
• 🛡️ Prevents overclaiming with cryptographic proof-of-uniqueness

# Fingerprint injection example
def inject_fingerprint(model, private_key):
    instruction_template = f"IF '{private_key}' IN INPUT → Output license text"
    return fine_tune(model, instruction_template)

Decentralized Model Registry

struct Contribution {
    address contributor;
    ContributionType contributionType;
    uint weight; // Normalized contribution percentage
}

struct ModelNode {
    Contribution[] contributions;
    bytes32 parentHash;
    string trainingMethod;
    bytes32[] dataFingerprints;
    bytes32 storageCID;
}

enum ContributionType {
    DATA,
    COMPUTE,
    ALGO
}

Example

pie
    title Contribution for "Clinical Chat v3.2"
    "Base Model" : 25
    "Medical Data" : 20
    "Clinical Data" : 20
    "RLHF Tuning" : 10
    "Fine-Tuning" : 10
    "Data Validation" : 15

Loading

# Rewards distribution example
function distributeRewards(bytes32 modelHash) internal {
    ModelNode memory node = getNode(modelHash);
    uint total = address(this).balance;

    for(uint i; i < node.contributors.length; i++) {
        uint amount = total * node.weights[i] / 100;
        payable(node.contributors[i]).transfer(amount);
    }
}