Unstoppable models

(ETHPrague2024 project)

#AI #LLM #docker #streamlit #solidity #foundry #forge #ipfs #subgraph #DAO #decentralized #ethereum #zk #byzantineTraining #optimism #cardona #zircuit #linea #sepolia

Demo

live: https://unstoppable-ai.why-nft.com/

(https://unstoppable-models.streamlit.app/ deprecated)

video: https://youtu.be/751XpV0xqEo

Brief Description

We aim to create a public ML model training initiative (as a public good) where anyone can freely submit the next weight update for the model according to a fixed algorithm on a fixed dataset.

• Detailed Description
  • Roles
• Realisation
  • Scheme
  • Off-chain parts
    • Frontend
    • Worker
  • Blockchain Related Parts
• Deployments
  • Cardona
  • Optimism Sepolia
  • Sepolia
  • Zircuit
  • Linea
• Develop
  • Contract
    • Deploy
      • Contract
      • Subgraph

Detailed Description

The initial weights $\theta_0$ for the model $f$ are publicly initialized offchain.

A public dataset D is fixed offchain.

In each iteration $i$, all participants are informed about the specific data chunk (batch) $D_i$ to be fed into the model to update its weights.

All interested participants locally compute gradients $\nabla_{\theta_i} f := \frac { \partial f(D_i, \theta_i) } {\partial \theta_i}$.

They submit the hash of this computation to the chain (and optionally publish the actual value publicly).

The new weights are computed as $\theta_{i+1} = \theta_i - \nu \nabla_{\theta_i} f$. (or via similar optimization algorithm)

Every k steps, the updated weights are published offchain.

Roles

Worker - Acquires a mutex (with a deposit) and attempts to perform a weight update step (or k steps). The Worker then publishes the updated weights on IPFS.

Validator - Token holders who resolve disputes. Validators receive rewards regardless of the outcome but are incentivized to act honestly to promote project growth.

Fraud-Proofer - Independently validates selected steps within a specific timeframe and can escalate disputes to the Validator if suspicious activity is detected (posting a deposit). If a correctly identified error is confirmed by the Validator, the Fraud-Proofer receives the Worker's stake.

Realisation

Scheme

Off-chain parts

Frontend

observes whole training process

cd research && \
python3 -m pip install -r requirements.txt && \
python3 -m streamlit run frontend-state-listener.py

(by default it listens to Cardona smart-contract instance)

Worker

performs updates and syncs results with contract

cd research && \
python3 -m pip install -r requirements.txt && \
python3 worker.py

Blockchain Related Parts

Deployments

Cardona

• contracts: deployments
• subgraph: https://api.studio.thegraph.com/query/77200/unstoppable-models/0.0.6

Optimism Sepolia

• contracts: deployments
• subgraph: https://api.studio.thegraph.com/query/77200/unstoppable-models-optimismsep/0.0.2

Sepolia

• contracts: deployments
• subgraph: https://api.studio.thegraph.com/query/77200/unstoppable-models-sepolia/0.0.1

Zircuit

• contracts: zircuit.json
• subgraph [self running] http://34.32.233.93:8000/subgraphs/name/unstoppable-models-zircuit

Linea

• not enough balance =(

Develop

Contract

Deploy

Contract

source .env && CONTRACTS_ENV_NAME=sepolia forge script --chain sepolia script/Deploy.s.sol:Deploy --rpc-url $SEPOLIA_RPC_URL --broadcast  -vvvv --legacy

source .env && CONTRACTS_ENV_NAME=cardona forge script script/Deploy.s.sol:Deploy --rpc-url https://rpc.cardona.zkevm-rpc.com --broadcast  -vvvv --legacy
source .env && CONTRACTS_ENV_NAME=linea forge script script/Deploy.s.sol:Deploy --rpc-url https://rpc.goerli.linea.build --broadcast  -vvvv --legacy
source .env && CONTRACTS_ENV_NAME=optimismsepolia forge script script/Deploy.s.sol:Deploy --rpc-url https://optimism-sepolia.drpc.org --broadcast  -vvvv --legacy
source .env && CONTRACTS_ENV_NAME=zircuit forge script script/Deploy.s.sol:Deploy --rpc-url https://zircuit1.p2pify.com --broadcast  -vvvv --legacy

or to Anvil

CONTRACTS_ENV_NAME=local forge script script/Deploy.s.sol:Deploy --fork-url http://localhost:8545 --broadcast  --legacy

questions:

• How DAO would be incentivised to act honestly? Like not to collude with the Worker to get the Worker's stake?
• Do you have a specific dataset in mind?

Subgraph

npm run import-config-networks && npm run create:local && npm run deploy:local