Decentralized GRPO training on Bittensor's Finney network, Subnet 81.
Live network · Run a miner · Protocol · Research
Reliquary turns independent GPU operators into a verifiable training market. Miners search for prompts at a model's learning frontier, the validator recomputes the evidence, and healthy selected groups can contribute to the next checkpoint.
- Find — miners compete to locate useful training signal before compute is committed.
- Prove — selected candidates pass deterministic verification and validator-authoritative reward checks.
- Train — clean, complete windows become eligible for GRPO; archives and checkpoint claims remain inspectable.
| Surface | What it shows |
|---|---|
| Live dashboard | Current windows, selection, rewards, and training health |
| Proof explorer | Public window records and validator provenance |
| Network status | Availability and data freshness |
| Canonical source | The deployed protocol and its current contract |
Live values belong on live surfaces. This profile intentionally does not freeze changing counts, checkpoints, or projections into marketing copy.
| Repository | Role |
|---|---|
reliquary |
Canonical Subnet 81 protocol: mining, verification, selection, training, archives, and checkpoint publication |
reliquary-fleet |
Released, private-by-default operations dashboard for miner fleets |
These repositories preserve earlier experiments and reusable primitives. They are not the production Subnet 81 implementation.
| Repository | Scope |
|---|---|
reliquary-ledger |
Historical proof-carrying inference and validator-mesh testnet |
reliquary-forge |
Historical trainer-quorum and policy-delta exploration |
reliquary-protocol |
Standalone reference for shared protocol primitives |
Reliquary Research separates measured evidence, deployed behavior, and proposed mechanisms. An experiment does not affect production ranking, rewards, or training until the canonical source and deployment state say it does.
finney · subnet 81 · inference → evidence → weights