OPML (Optimistic Machine Learning on Blockchain) provides a framework for executing AI model inference off-chain with on-chain dispute resolution, enabling verifiable machine learning on blockchain. It targets developers and researchers looking to integrate AI capabilities into decentralized applications with a robust fault-proofing mechanism.

How It Works

OPML utilizes an optimistic execution model where off-chain computations are assumed correct until challenged. An on-chain interactive dispute engine implements fault proofs, allowing any party to challenge an incorrect inference. This approach leverages a custom Golang tensor library (mlgo) and a MIPS runtime (mlvm) for efficient execution, with plans for ZK fault proofs and GPU acceleration.

Quick Start & Requirements

• Install: git clone git@github.com:hyperoracle/opml.git --recursive followed by make build.
• Prerequisites: Go (1.19+), Node.js, Make, CMake.
• Example: Run npx hardhat node then bash ./demo/challenge_simple.sh.
• Docs: project wiki, MNIST tutorial.

Highlighted Details

• Supports DNN for MNIST and LLaMA models.
• Includes an example for LLaMA and LLaMA 2 on a separate branch.
• Roadmap includes general DNN (ONNX), traditional ML, ZK fault proofs, and GPU acceleration.
• Core components: mlgo (Golang tensor library), mlvm (MIPS runtime), contracts (on-chain logic).

Maintenance & Community

The project is actively under development with several features marked as "Work In Progress" or "Pending". Specific contributors or community channels are not detailed in the README.

Licensing & Compatibility

MIT licensed. Parts of the code are borrowed from ethereum-optimism/cannon.

Limitations & Caveats

The code is unaudited and should not be used for securing funds without extensive testing and auditing. Support for general DNN models, traditional ML algorithms, training, fine-tuning, ZK fault proofs, GPU acceleration, and a user-friendly SDK are still in progress.