README.md

metadata

title: Train@home
emoji: 🚂
colorFrom: blue
colorTo: blue
sdk: static
pinned: false

Train@home

Community-owned AI training. Like Folding@home, but for neural networks.

What is this?

A distributed, open-source project where volunteers contribute compute to train language models. Anyone with a GPU can participate. The model and knowledge banks are collectively owned by all contributors.

How it works

1. Contribute compute — download a checkpoint, train on your assigned data shard, upload gradients
2. Contribute knowledge — process documents into facts, submit them to the shared knowledge bank
3. Govern — propose and vote on training objectives, data distribution, architecture changes

Current status

🚧 Early stage — Building the foundation. Architecture design, training infrastructure, and initial knowledge bank.

Get involved

• Fork the main repo
• Run the setup script
• Join the discussions
• Start contributing

Architecture

• Processing core — Recurrent architecture with adaptive computation and revision
• Knowledge bank — Hot-swappable, product-quantized fact store
• Retrieval interface — Always-on, integrated into the forward pass (not tool calling)

Training cycle

Training runs happen every 2-3 months. Between runs, the community proposes changes via PRs, discusses tradeoffs, and prepares data. Config freezes 2 weeks before training starts.

---

Built by the community, for the community.