Reference
Glossary
Definitions of every Signet-specific and domain-specific term used in the documentation.
| Term | Definition |
|---|---|
| Arena | The 75% share of miner emissions distributed to the top 10 eligible miners on an exponential curve |
| Artifact | The frozen, immutable submission: one public model repository holding weights and a loader class, pinned by commit SHA and per-file hash, executed in a network runtime profile |
| BLEU | N-gram overlap translation metric, kept at the lowest weight for literature comparability |
| BLEURT | Learned semantic translation metric; Signet's primary translation signal |
| Canary slice | 30 clips per round drawn from unseen conditions, used for grounding and generalization, excluded from the translation score |
| chrF | Character-level overlap metric; robust to morphology and small variation |
| Champion | The current leaderboard leader, and the artifact deployed behind the commercial API |
| Derangement | A deterministic shuffle of the reference set in which no reference stays matched to its own clip |
| DTW | Dynamic time warping; removes harmless timing differences when comparing pose sequences |
| Grounding gate | The pass/fail test that a model actually read the video. G_canary < 1.25 on chrF means zero, code R14 |
| Grounding ratio (G) | Score against true references divided by score against deranged references |
| How2Sign | Public ASL multimodal dataset (CVPR 2021). Training substrate only |
| Loader class | The Model class inside the submission repository, named by the manifest's entry (default signet_model:Model), exposing load() and translate(video_path). Imported and driven by the validator's own runner |
| Manifest | The single committed object, schema version 3.0: hotkey, model reference (repository, commit SHA, per-file hashes, entry point, runtime profile), parameter declaration, precision, provenance, and seed |
| MediaPipe Holistic | The canonical 543-landmark pose layout Signet requires: 33 body, 468 face, 21 left hand, 21 right hand |
| Miner | A participant who trains a model and publishes a frozen artifact. Performs no work during scoring |
| MPJPE | Mean per-joint position error; the base pose accuracy measure |
| PA | Procrustes alignment; removes camera translation, scale, and rotation before pose comparison |
| PHOENIX-2014T | Public DGS translation and gloss benchmark. Training substrate only |
| Qualifying pool | The 25% share of miner emissions split proportionally among all gate-passing submissions above the score floor |
| R14 | Rejection code for failing the canary grounding gate |
| Round | One 24-hour evaluation cycle over a 300-clip draw |
| Runtime profile | One of the small set of execution environments the network publishes, named in the manifest's runtime field. The manifest names the profile and never a digest; the network publishes each profile as an image pinned by digest when it releases that runtime. Currently torch2.4-transformers4.44. Miners never build or publish images |
| SHuBERT | Sign-specific self-supervised representation (ACL 2025) |
| SignMusketeers | Efficiency-oriented multi-stream sign research (arXiv 2406.06907) |
| Slice | One condition group inside a round draw: studio, broadcast, conversational, or canary |
| Uni-Sign | Pose-first model family (ICLR 2025); a research anchor for the pose-first branch the reference miner belongs to |
| Validator | A participant who freezes commitments, executes artifacts offline on private data, scores them, and sets weights |
| YouTube-ASL | Large public ASL training corpus (NeurIPS 2023). Training substrate only |
References and project research anchors
- Uni-Sign (ICLR 2025) — reference pose-first model family
- SHuBERT (ACL 2025) — sign-specific self-supervised representation
- SignMusketeers (arXiv 2406.06907) — efficiency-oriented multi-stream research
- YouTube-ASL (NeurIPS 2023 / arXiv 2306.15162) — large ASL training corpus
- How2Sign (CVPR 2021 / arXiv 2008.08143) — ASL multimodal dataset
- PHOENIX-2014T — DGS translation and gloss benchmark
- SignBind-LLM (arXiv 2509.00030) — benchmark and generalization-gap reference
- Factorized gloss-free SLT with LLMs (arXiv 2403.12556) — LLM adaptation pattern
- MediaPipe Holistic — canonical pose-output basis
- RTMPose / MMPose — pose extraction stack used by leading research baselines
Source basis: Signet V1 project specification (README, miner, validator, scoring, implementation architecture) and the approved Architecture / Market / Economic Design document, August 2026.