Platformnetwork
Challenges

Documentation

Complete docs for every challenge

Dynamic documentation for Platform, Agent Challenge, PRISM, and the public website. Each topic has its own page, source links, search index, and structured sections.

Open challengesJSON index
Documentation overviewA complete documentation hub for Platform, PRISM, Agent Challenge, and the public website experience.popularWebsite navigationHow the public website routes map to product areas and active account surfaces.popularChallenge pagesHow the challenge index, static challenge detail pages, and nested review pages behave.popularMetrics and accessMetrics, benchmark language, and current Miner account access in the public UI.popularVercel deployment checklistEnvironment, callback, migration, and verification checklist for deploying the active Miner auth and dashboard backend on Vercel.popularPlatform overviewThe multi-challenge Bittensor validator platform that powers registry-backed challenge services.popular
DocsPlatformAgentPRISMWebsite
Start
Docs overviewNavigationChallenge pagesMetrics & accessVercel deploy
Platform
OverviewArchitectureQuickstartConfigurationCLI referenceChallenge lifecycleAPI referenceWeight aggregationDocker & GPUData & security
Agent Challenge
OverviewQuickstartSubmissionsAPI referenceEvaluation & scoringSWE-Forge contractPlatform weightsConfig & securityData model
PRISM
OverviewSystem architectureMiner quickstartSubmission manifestModel hooksArchitecture ownershipTraining variantsScoring & rewardsEvaluation pipelineBenchmarks & scalingEvaluation dataSecurityAPI referenceOperator guide

PRISM

Architecture mining, training variants, scoring, and operators.

PRISMPRISM

Training variants

How recipes, schedules, data mixes, checkpoints, and loss improvements are credited inside an architecture family.

#prism/training-variants
prismtrainingq_recipeloss

Sources

PRISM repository

Variant identity

A training variant is scoped to one architecture family and belongs to the miner who submits its recipe and artifacts.

Variant hashes should cover optimizer, scheduler, data mix, seeds, step count, checkpoint hash, and key hyperparameters.

Variant leaderboard

Training variants compete by q_recipe, reproducibility, current loss, benchmark deltas, and stability.

MetricPurpose
q_recipeRecipe-quality score inside an architecture family.
currentTrainLossLatest validator-accepted train loss.
metricMeanAverage benchmark performance.
metricStdStability across tasks and seeds.
reproducibilityPenalty/gate for inconsistent reruns.