OPML enables off-chain AI model inference using optimistic approach with an on chain interactive dispute engine implementing fault proofs.

For more in-depth information, refer to the project wiki.

You can also find a tutorial on building a straightforward handwritten digit recognition DNN model (MNIST) within OPML in the docs/tutorial.md.

Pre-requisites: Go (Go 1.19), Node.js, Make, and CMake.

git clone git@github.com:hyperoracle/opml.git --recursive
make build

🔨 = Pending

🛠 = Work In Progress

✅ = Feature complete

mlgo -- A tensor library for machine learning in pure Golang that can run on MIPS.
mlvm -- A MIPS runtime with ML execution
contracts -- A Merkleized MIPS processor on chain + the challenge logic

The script files demo/challenge_simple.sh presents an example scenario (a DNN model for MNIST) demonstrating the whole process of a fault proof, including the challenge game and single step verification.

To test the example, we should first start a local node:

npx hardhat node

Then we can run:

bash ./demo/challenge_simple.sh

A large language model, the llama example is provided in the branch "llama" (It also works for llama 2).

Note: This part is still under development!

We can support LLaMA 1 and LLaMA 2 (7B and 13B). Before testing LLaMA, please refer to mlgo/examples/llama/README.md and download the model of LLaMA.

After that, we can first start a local node

npx hardhat node

Then we can run

bash ./demo/challenge_llama.sh

Note: when running bash ./demo/challenge_llama.sh, you may encounter such an error in console "SocketError: other side closed". Just ignore it. :) This is a special "feature" of hardhat when running the JS script that takes too long time. I have fixed it in the script. Although you can see an error in console, the script should run correctly.

This code is MIT licensed.

Part of this code is borrowed from ethereum-optimism/cannon

Note: This code is unaudited. It in NO WAY should be used to secure any money until a lot more testing and auditing are done.