Bittensor Subnet 47
Better models.
Proven in public.
Miners improve shared models. Validators reproduce the evidence. Feval turns open competition into results the network can inspect.
Live validator matrix
Miner leaderboard
One row per miner. One verdict per validator. Select a score to inspect its evidence.
PassedIn progressCarriedRejected
No miners match this view
Try a different search term or status filter.
Results refresh automatically from public validator W&B runs. Carried results come from the named source window while its replacement is evaluated. Reward status is shown separately from audit progress.
The protocol
From model work to a public verdict.
Feval separates expensive exploration from bounded verification, so evaluation capacity can grow with miner participation.
- 01
Commit
Miners pin model and rollout revisions before the audit sample is known.
- 02
Reproduce
Validators recompute deterministic rewards and replay unpredictable traces.
- 03
Publish
Scores, audit state, and validator evidence become an inspectable public record.
Qwen3‑4B base model10,000 committed rows320 sampled audit rows5k + 5k math and instruction tasks
Roadmap
Prove the loop. Then widen it.
Every phase expands what miners can improve while keeping rewards tied to independently reproducible evidence.
Verifiable post-training
Miners post-train Qwen3‑4B‑Base on tasks with deterministic outcomes. Validators reconstruct scores and audit committed traces before accepting an improvement.
- Bounded adapters
- Verifiable rewards
- Independent replay
Decentralized pretraining
Begin from a Teutonic.ai Subnet 3 checkpoint and measure durable foundation-model progress across tasks and evaluation windows.
- Subnet 3 checkpoint
- Cross-window evidence
- Foundation-model gains
System design · 001
Verification should scale more slowly than discovery.
Miners spend compute exploring improvements. Validators establish confidence with commitments, deterministic scoring, and sampled replay—not by repeating every training run.
Inspect the evidence