Houshou

Currently, models like GPT/Claude aren't really useful for projects offering devtooling because of the lack in training data and lack of bandwidth in teams for creating LLMs.

This project solves this by enabling contractors (projects) to create a custom model for themselves by crowdsourcing dataset from developers.

The contractors can create a bounty pool with questions related to their tooling like "how to initiate setup", "explain this piece of code", etc., which will be hosted on our platform and workers (developers) can contribute to fill in the content for a part in the bounty. The contractor can review the answers, delete responses or strike off users and then train the model.

A new model will be created using the gathered dataset from the workers and this will act as a specific LLM for that devtooling project (startup)

Devtooling Startups or projects (contractors) can create a bounty pool with an assigned deadline and include multiple forms containing questions related to their toolkit. Upon creation of a pool, a new Worldcoin contract is deployed, ensuring each developer (worker) can submit a form only once and not submit multiple forms. All forms are stored on IPFS.

When a developer fills out a form and completes WorldID validation, it triggers an attestation hook via the Sign protocol. The information is then stored in contracts deployed on the Gnosis Chiado chain. The answers are stored off-chain in MongoDB for later use during model training.

Contractors can review, delete responses, or strike a user if necessary. Each user can receive a maximum of three strikes before being banned. Afterward, contractors can name and train the model using a Retrieval-Augmented Generation (RAG) system with a FAISS vector database, which processes requests via the Phala Network.

Once the deadline is reached, the bounty pool is distributed to the eligible workers through Chainlink contracts.