# Quantum Pixels

- **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:** [Maheswar Praveen](https://cerebralvalley.ai/u/maheswarpraveen)
- **GitHub:** https://github.com/MaheswarPraveen/data-refinery
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/481

Data Refinery: raw data in, refined data out.

https://data-refinery.primemax.workers.dev

AI models are only as good as their training data, and real datasets are full of wrong labels. Fixing them by hand is slow and expensive. Data Refinery is a distributed consensus network that cleans data automatically.

A coordinator sends each item to five independent nodes, each running different small Meta open-weight models (Llama 3.2 and 3.1 among ten model families). They answer blind. If 4–5 agree, the answer is kept. If 3 agree, Meta Muse Glimmer 30B acts as a judge, gives a hint (never the answer), and they vote again. Otherwise the item is rejected. Votes are signed, hidden test questions catch cheating nodes, and every calculation in the working is step-checked. Outvoted answers become “wrong” examples for preference training.

Results:

Error detection: wrong labels cut from 35.5% to 1.3%, with no human labelling.

Refined vs raw: at the same size, refined data trained a better model than raw data (+3.6) and than human-written answers (+4.0).

Fine-tuning: Llama 3.2 1B trained on just 2,183 refined examples (about 4 minutes) gained +32 points on average across six unseen datasets.

An agentic data search (Llama 3.3 70B) finds sources and quotes refining jobs.

In future we hope for a  full decentralisation. Idle home GPUs join as paid nodes, with no single operator.

## More from Meta Global AI Developer Hackathon

- [Kabuga](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/478)
- [Fraction Compressor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/479)
- [MegastarAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/480)
- [KISKY TEAM](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/482)
- [Amr Elhabbal - Neura](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/483)
- [Doubloon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/484)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/481. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt