skip to content
Reliquary

docs

How Reliquary works

One-page explainer. Everyone should read this first.

FIGthe core training loop
the core training loopMINERselect + rolloutGRAILverify weightsGRPOtrain survivorsCHECKPOINT→ Hugging Facenext window — against the new checkpoint

Reliquary is a decentralized GRPO training market on Bittensor Subnet 81. A model learns fastest from prompts at its learning frontier — hard enough that rollouts disagree, easy enough that the gradient still carries signal. Reliquary turns finding those prompts into a competitive market.

Miners aren't paid per rollout. They're paid for verified training signal. A miner predicts which prompts sit in the learning zone, generates rollouts, and attaches a cryptographic proof that the work is real. The validator verifies and selects the eligible groups. Forge then records whether a healthy batch trains or holds; optimizer, publication, and evaluation remain separate evidenced states.

The actors

  • Miners select prompts at the learning frontier, run the model, and submit rollouts + a GRAIL proof.
  • The validator recomputes rewards, verifies proofs, selects eligible groups, and emits the signed artifacts Forge consumes.
  • Forge records train-or-hold decisions, optimizer steps, checkpoint publication, and evaluation as distinct artifacts.
  • The chain records selection and scoring so outside observers can audit later.

One window, end to end

  1. The validator announces the active checkpoint and the per-window randomness.
  2. Miners bet their own compute on prompts they predict sit in the trainable band, generate M rollouts per group, and attach a GRAIL sketch binding each rollout to the announced weights.
  3. The window seals once enough valid distinct-prompt groups land; final ordering follows drand/canonical rules, not validator-side latency.
  4. The validator recomputes every reward, verifies every sketch, selects training-eligible survivors, and emits a signed batch receipt. Fabricated work earns zero.
  5. Forge records a train-or-hold decision. When a PPO-clipped GRPO step succeeds, its optimizer evidence is distinct from publication; updated weights publish to Hugging Face every 4 successful balanced optimizer steps (or earlier on behavior-policy drift), with a signed manifest recording the chain of custody from the base model.

Why selection is the game

Only a small slice of prompts sit in the learning zone at any checkpoint, and the band narrows as the policy matures. A miner who picks well lands on winning prompts and earns emission; a miner who picks poorly burns its own compute on rejects like out_of_zone. This converts DAPO's reactive generate-then-discard filter into an ex-ante prediction market — and makes selection intelligence more valuable over time, not less.

Why the work can't be faked

Each candidate supplies GRAIL proof material — a fingerprint over hidden-state activations at sampled positions, bound to per-window randomness the miner can't predict in advance. After the populations freeze, the validator proves ranked candidates top-down; every credited group must pass tolerance. The full attack-class audit lives at /docs/scoring.

What's different from today's subnets

Most Bittensor subnets pay miners for volume. Reliquary pays for the rollouts the trainer needs — and proves they're real before a single gradient step. The network's output is a continuously-trained model published to Hugging Face, not just a leaderboard.