skip to content

dashboard research roadmap docs source

Decentralized GRPO training on Bittensor Subnet 81. Miners find the prompts at the model’s learning frontier; verified rollouts train the next checkpoint.

/dashboard · live
/research
/roadmap
/status
source · reliquary-web

ReliquaryThe Archive

The verified rollout archive. Every selected bundle — prompt, rollouts, GRAIL sketch, recomputed rewards — kept in R2 and replayable. The dataset behind each checkpoint.

/dashboard · leaderboard
/#archive · the archive
/explorer · proof windows
/mesh · validator design

ReliquaryThe Furnace

The training loop. Each window, the validator runs a GRPO step on the rollouts that survived and publishes new weights to Hugging Face.

/#forge · the forge
/forge/cycles · live attestations
/training · the training loop
/docs/scoring · how miners score
/paper · protocol spec

mit · reliquadotai · 2026ledger · forge · one protocol

forge · training cycles

SFT / GRPO alternation

Each attestation is one training cycle. The merkle root pins the checkpoint; the consumed list is the set of rollout windows that fed the cycle; effective_at_ledger_window is when validators start scoring against it.

loading

probing forge attestations

Fetching the latest SFT/GRPO training cycles…

Forge — training cycles · Reliquary