Train
Train Byrne70M.
This page measures the computer or phone you opened it on. A machine that can hold the model trains it here. Smaller devices split the layers. Across networks, share a room link and the host approves each joiner.
This device
Group
Progress
Up to 70M
This is the verified INT8 transformer. Byrne70M, below, is the Cycle2 architecture. Every device holds a full copy of this one. Width, sequence, and batch follow the smallest device. The target is 70M parameters. A device that cannot hold 70M gets the largest model that fits. Weights are allocated when training starts.
Measuring this device…
Steps are the pretrain length. 1000 is a start. The slider goes to 100000. Text is FineWeb-Edu. A failed stream falls back to the built-in corpus.
Byrne70M
Same modules as the Cycle2 checkpoint. Fresh weights in this browser. The hash planes are a new frozen draw, same shapes. AdamW, not Muon.
measuring this device…
Blend
Public Hugging Face dataset repos. The rows start as the Cycle2 cooldown blend. Change them. Weights are the mix. No token. A private repo is skipped.
One step runs 22 layers, three times. The sequence is the training window, picked from the memory left after the weights. 1000 steps is a start, not a finished pretrain. Stop ends the run. The architecture context stays 2048.
Try the verified model
Checkpoint
A loaded checkpoint is applied here and sent to every connected device.
Log
Devices talk to each other directly, so a device in the group can see the others' IP addresses. Train with people you trust. HTTPS or localhost is required.