Loop

Loop is an onchain platform for reinforcement learning with human feedback (RLHF). Developers or corporate teams can upload models from HuggingFace (by submitting a link) and issue cryptocurrency bounties for crowdsourcing output validation. Users can sign in via web3 wallet (or web2 social platforms) and utilize the chat interface for regular functionality (i.e., writing essays), while being prompted to validate outputs. These prompts thereby disseminate the crypto rewards, effectively creating a Validate-to-Earn model. Subsequently, the prompts, outputs, and users' scores are stored off-chain in a dataset, accessible to the devs/teams behind the associated model.

On the front-end, we built a responsive interface using React.js, implementing custom hooks for state management and modular components for content display and user interactions. Tailwind CSS provided a utility-first styling approach that dramatically accelerated our development cycle, while thirdweb's SDK facilitated seamless wallet connections and blockchain interactions across multiple networks. We deployed our front-end to Vercel.

The back-end infrastructure centers around Foundry-developed smart contracts, meticulously unit tested and deployed for multi-chain usage. We've successfully deployed these contracts across nine different testnet environments: Polygon, Hedera, Neon EVM, Zircuit, Story, AirDAO, Rootstock, Morph, Unichain, and Story testnets.

For AI processing, we leveraged HuggingFace's model hub to deploy both Llama 3.1 8B and Mixtral 8x-7B models, fine-tuned on our custom dataset. We leveraged Prime Intellect for decentralized GPU orchestration, allowing us to run LLM inference across distributed compute providers like Akash and RunPod during preliminary testing. Firebase serves as our human-verified dataset repository, chosen for its real-time database capabilities and seamless integration with our scoring system.