Pitch: Sats for RLHF

The Challenge...

Creating Large language models requires the ingestion of large amounts of relevant data. Relevant being the key word. As they say, Garbage in, Garbage out.

In curating all the world's Bitcoin data to train a Bitcoin-centric language model, we discovered that there's a whole lot of data - but none of it is in a format that a model can be trained on. Yes - you can use other existing language models to programmatically clean data, but that doesn't solve the core problem which is; existing language models don't understand the nuance of Bitcoin, so much of what comes back is junk.

In order to really train a model, you have to clean up the data, by engaging a large number of humans to help. You can do this with fiat but it’s complex, difficult to do at scale, especially globally across borders.

Bitcoin fixes this, because you can do micro-payments internationally, granularly, and you can pair it with a consensus mechanism so that you can ensure high quality involvement & feedback.

<iframe class="remirror-iframe remirror-iframe-youtube" src="https://www.youtube-nocookie.com/embed/kAqR1LqLULs?" data-embed-type="youtube" allowfullscreen="true" frameborder="0"></iframe>

Current Features:

• Signup/Login via email, LNURL Auth, and Nostr

• Contributor Tools: Keep, discard, or edit data to be fed back into LLM’s for training

• Consensus Logic: reviewed items are fed into a consensus logic to ensure quality human feedback and prevent spam/abuse.

• Sats rewards: Contributors are rewarded with sats for their contributions!

• Lightning: instant and frictionless withdrawals via LNURL withdraw to your own wallet.

Next Steps:

• Polish: Further polish on the existing experience and app flows

• Community: Build out user profiles and allow contributors to connect with each other socially via Nostr interactions.

• Leaderboard & Stats: Provide contributors with details insights into their contributions, consensus state, earnings, and leaderboards!

• More tools: Provide additional contribution methods like benchmarking, ranking replies, FudBuster, etc 🙂

Team

• Backend: Vano

• Frontend: Jin

• QA and PM: Jeff

Tech Stack

• Backend:

  • NodeJS

  • PostgreSQL

  • Stateless REST Api

  • Monolithic Architecture

  • AWS

• Frontend:

  • React Native Javascript using Expo

• Read more about out tech stack on our Github: https://github.com/Laiertwo/sos-data-cleanup

Judges - Try it out!!

You can access our app at:

• https://www.train.spiritofsatoshi.ai/sign-in

Go ahead and create your own account with Nostr of Lightning login! We have also created a few temporary email logins with some pre-loaded sats which can be accessed if you'd like to test withdrawals. HERE. Or you can create your own account of course!

Email: [email protected]
Inbox: https://temporary-email.com/inbox/#/rahewcop

Email: [email protected]
Inbox: https://temporary-email.com/inbox/#/mamiestewart

Email: [email protected]
Inbox: https://temporary-email.com/inbox/#/lizziejoseph

Email: [email protected]
Inbox: https://temporary-email.com/inbox/#/thomasconner

Email: [email protected]
Inbox: https://temporary-email.com/inbox/#/movuffol94

Email: [email protected]
Inbox: https://temporary-email.com/inbox/#/evelynalvarado

Please reach out with any questions or if you need some more sats into your account so you can further test our withdrawals.

Happy training+sat stacking!