signetDocsTry Demo
Start here

Introduction

Signet is a Bittensor subnet that produces the strongest open-weights ASL to English translation model, selected on private data.

Signet is a Bittensor subnet with one job: continuously produce the strongest transparent, reproducible, open-weights American Sign Language to English translation model that an open competition can generate.

It is not an inference marketplace. Miners never host an endpoint and validators never trust a miner-operated service.

Miners publish frozen models. Validators execute them offline on private ASL footage. The champion becomes a deployable commercial artifact.
Document status

Docs version 1.1 · September 2026 · V1 pre-registration

This documentation is the canonical description of Signet V1. There is no separate whitepaper. Everything the protocol specifies is on these pages.

The mechanism, scoring rules, and submission contract are final. The subnet is not yet registered on mainnet. See the Roadmap and the Changelog.

The 30-second version

  1. A miner trains an ASL-to-English model and publishes frozen weights and a loader class in one public Hugging Face repository, then commits a manifest on chain that pins the repository by commit SHA, hashes every declared file, and names one of the network's runtime profiles. Miners never build or publish a container image.
  2. A validator pulls the named runtime, mounts that exact repository read-only, and runs the model itself with networking disabled on ASL footage that has never been published.
  3. Scores come only from that private corpus. Public benchmarks never set weights.
  4. The best models earn the largest share of emissions.
  5. The winning artifact is deployed by the subnet owner behind a commercial captioning API.

The core loop

Miner trains model→Publish weights + loader class to one HF repository→Commit manifest on chain — commit SHA, file hashes, runtime profile→Validator pulls network runtime + repository at that commit→Execute offline — network disabled→Score on private clips→Set miner weights→Champion deployed as commercial API

Why this design

Public sign-language benchmarks publish their test clips and reference translations. That is correct for academic comparison and wrong for a network that pays out continuously. If the answers are public, money eventually flows to memorization instead of understanding.

Signet never sets weights from How2Sign, YouTube-ASL, OpenASL, PHOENIX-2014T, or any other public corpus. Those are training substrate. Only the private, rotating Signet corpus decides rewards.

Read the full argument →

What Signet produces

OutputDescription
English translationOne sentence-level translation per input clip
Canonical poseA 543-landmark MediaPipe Holistic sequence per clip, when the model returns one
Champion artifactAn immutable, inspectable, deployable model
Longitudinal corpusA growing private evaluation asset

Where to go next