# Horizon

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Jacob Chau](https://cerebralvalley.ai/u/jch)
- **GitHub:** https://github.com/4uhn/Meta-AI-Hackathon
- **Demo video:** https://www.youtube.com/watch?v=1QLFcgaifCU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2255

So much valuable data is locked away in organisations that can't legally or commercially share it. Current plain federated learning isn't enough as I showed that from a single model update, a server can tell which classes, and so what kind of data, an organisation holds. So during this hackathon I built Horizon, an open-source framework that lets organisations fine-tune an open-weight model together. Every update is encrypted with threshold CKKS before it leaves, so nobody, including the server, sees any one organisation's contribution, and the aggregated sum can only be decrypted when t of the N organisations cooperate. As a proof of concept, three simulated mobile operators jointly fine-tuned Meta's Muse Glimmer-30B and reached 99.4% of the performance if you would've trained the model by pooling all the data together. It's an early step, but it shows the feasibility in unlocking the world's locked-away data, letting competing organisations collaborate and push the frontier together.

## More from Meta Global AI Developer Hackathon

- [Paizah](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2252)
- [NyxSec](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2253)
- [LightDay](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2254)
- [VVana Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2256)
- [ChromaJunction](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2257)
- [Aegis Omni](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2258)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2255. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt