README.md

metadata

title: DaisyChain-Web
emoji: 🌼
colorFrom: green
colorTo: yellow
sdk: docker
app_port: 7860
pinned: true
license: mit
short_description: Train a shared model P2P by opening a browser tab

🌼 DaisyChain-Web — train by opening a page

Open this Space on two or more devices and they pretrain a language model from scratch together — peer-to-peer over WebRTC, right in the browser, computing through verified INT8 neural units (WebGPU, or the same units on CPU for old machines). This is pretraining, not fine-tuning: every run starts from random weights.

• Create a room and share the invite link with your devices, or join a room by code — the room's creator approves each device before it can join.
• Pick the training settings with sliders — including any public HuggingFace dataset with a text column; whoever presses Start sets them for the whole group and every device follows automatically.
• Devices that drop are redialed (5 attempts) then removed; devices that join mid-run are synced in with a live state transfer; if the sync leader leaves, the next peer is promoted and the run continues.
• Gradients are averaged every step with a deterministic Adam optimizer, so all devices end with bit-identical weights.
• Download your trained model, or upload a checkpoint to restore the whole group.

Part of DaisyChain-Train — a pipeline for reusing old/spare hardware to train neural networks.

Heads up: peers connect directly (WebRTC), so devices in your group can see each other's IP address, and there is no gradient authentication — only train with devices/people you trust. Proof of concept.