RESEARCH
A decentralized research network for sign-language intelligence.
Miners commit models. Validators score with rigorous research-grade metrics. Together, we continuously improve sign-language translation through transparent research competition and open science.
TOP COMPOSITE · S_TOTAL
{{ topScore }}
Leading composite score
in the latest scored round
in the latest scored round
CLIPS THIS ROUND
{{ clipsRound }}
Clips in the round draw
across all slices
across all slices
SUBMISSIONS SCORED
{{ subsScored }}
Scored / committed
in the latest round
in the latest round
RESEARCH SCORING FORMULA
Every submission is scored by the subnet’s own scoring code: translation quality on the main slices, generalization to unseen canary conditions, and pose fidelity — gated by deranged-reference grounding.
Stotal=0.60Stranslation+0.25Sgeneralization+0.15Spose
Stranslation
Weight: 0.60
0.50 BLEURT + 0.30 chrF + 0.20 BLEU over the main slices.
Sgeneralization
Weight: 0.25
Canary-to-main BLEURT ratio against the 0.85 target.
Spose
Weight: 0.15
Normalized, inverted DTW-PA-MPJPE over body and hands.
Gcanary
Gate: floor 1.25×
Deranged-reference grounding ratio; below the floor scores zero (R14).
Bootstrap
Leader: p < 0.05
The champion flips only on a statistically significant win.
LIVE MODEL LEADERBOARD
{{ lbSub }}
Nothing to show
{{ lbEmptyNote }}
{{ r.rank }}
{{ r.label }}{{ r.sub }}
{{ r.total }}
{{ r.tV }}
{{ r.gV2 }}
{{ r.pV }}
{{ r.gV }}
{{ r.wV }}
{{ r.stText }}
{{ footNote }}
How scoring works
◆ {{ c }}
NETWORK OVERVIEW
Round records, scoring code and model outputs are published; anyone can re-score a round.
How validation works
ACTIVE VALIDATORS
{{ valCount }}
{{ valNote }}
LATEST ROUND
Round {{ roundNum }}
{{ roundSub1 }}
{{ roundSub2 }}
{{ roundSub2 }}
AVG TOTAL SCORE
{{ avgScore }}
Average S_TOTAL across
scored submissions
scored submissions
EMISSIONS BURNED
{{ burnPct }}
Undistributed share of the
round's emissions, burned
round's emissions, burned