For ML teams and community contributorsExplore public checkpoints
Collect human labels. Train custom AI models.
Stop stitching together annotation tools, training scripts, and serving logs. Taught keeps image, audio, and video labels connected to the people who reviewed them, the model versions they trained, and—when enabled—the API endpoints used by applications and agents.
Free waitlist. No credit card. Tell us whether you want to label, train, or integrate; we email you when relevant access opens.
- Model suggestions never count as human votes
- Labels keep source and reviewer provenance
- Promotion requires a frozen evaluation gate
For ML teamsReplace disconnected labeling, training, and serving handoffs with one traceable model workflow.See the team workflow
For contributorsAnswer clear questions about real content and earn project points on eligible missions. No AI experience required.See example missions
Synthetic demonstrationEVIDENCE STACK · CLIP 0187
- ContentAUDIO · 00:18Backyard recordingUnreviewed source
- Model suggestionPROPOSAL · NOT A LABELDog bark
The model proposes a time range. Its confidence stays hidden during blind review.
- Independent decisions3 REQUIRED · SOURCE HIDDEN
- Reviewer ARecorded separately Confirm
- Reviewer BRecorded separately Confirm
- Reviewer CRecorded separately Confirm
- Project gate passedTraining-ready evidenceREG · 0187–DB
Source, suggestion, and human decisions stay linked in the record.
The problem
Labels, training runs, and deployed models lose context at every handoff.
Teams end up reconciling spreadsheets, annotation exports, experiment trackers, and provider bills before they can answer a basic question: why should this model replace the last one?
With TaughtOne evidence chain follows the work.
- 01
Review Human decisions stay linked to their source content and task rules.
- 02
Improve Reproducible datasets feed candidate training and frozen evaluation.
- 03
Use Promoted versions carry exact model and usage receipts into serving.
Label content one clear question at a time.
Confirm whether a dog appears, draw a box around a bicycle, or mark where a sound starts and ends. Taught handles the model and dataset details while you focus on the content.
- Read the task before you begin
- Confirm, reject, add, or skip—depending on the mission
- Project-local points for eligible participation
Choose a missionEXAMPLES · CONTENT IS SYNTHETIC
Some accounts begin with setup and qualification missions while waiting for eligible project work.
Keep labeling, training, and serving in one traceable workflow.
Define the labeling task, collect independent decisions, freeze eligible labels into a reproducible dataset, and evaluate a candidate before it can replace the current model.
- DefineSet the labeling task
Choose the content, instructions, labels, and review gate.
- CollectCollect independent human review
Keep provenance and live vote direction hidden while people judge.
- FreezeFreeze a reproducible dataset
Agreement advances eligible evidence; disagreement stays visible for adjudication.
- PromoteTrain, evaluate, and promote
Candidate models face frozen data and promotion gates before replacing an incumbent.
One evidence language.
More than one kind of content.
The review changes with the medium. The rules do not: keep the source, suggestion, human action, and final disposition connected.
See what is there
Boxes · shapes · pointsConfirm objects, reject mistakes, and add what was missed.
Listen for the moment
Clips · events · rangesJudge a whole recording or mark a precise time interval.
Follow what changes
Frames · intervals · sequencesReview a clip with the context before and after an event.
Read what was written
Platform target · contract requiredClassify, extract, compare, or mark spans under a dedicated text task contract.
Keep structure with meaning
Platform target · contract requiredReview pages and fields without losing the source document’s provenance.
Register page-level evidence
Platform target · contract requiredPreserve file, page, region, and extraction evidence as one reviewable record.
Images, audio, and video have native review contracts today. Text, document, and PDF work require dedicated task contracts before they can be labeled or served. Labeling, proposal, training, export, and inference support always depend on the configured task.
Canonical serving path · deployment-dependent
A closed loop for models that earn promotion.
Taught can host metered API and inference access for promoted, serving-compatible custom models. Human teaching improves the evidence, promotion gates the model, and applications call the exact endpoint revision that passed.
- Human evidencePeople teach
Eligible project work records attributable labels and independent review.
- Promotion gateModels prove improvement
A candidate must pass the project’s frozen evaluation rules before promotion.
- Hosted serviceApplications call the model
A versioned prediction endpoint makes the promoted model usable beyond the review desk.
- Contributor choiceProjects can stage an exchange rate
Owners can configure a future points-to-credit rate. Contributors keep their points until redemption is enabled for the deployment.
What metered inference covers
Measured compute and GPU use, platform engineering and operations, plus a sustainable profit margin. Where provider choice is offered, organizations choose a named provider or Auto and see provider cost and platform markup before charge.
What points never change
Rewards do not increase vote weight, review authority, or the evidentiary value of a decision.
Points are not cash. Any future cash-payout program would be a separate, unlaunched product with its own eligibility, rates, policy, and compliance requirements.
Serve promoted models to applications and agents.
Projects move through labeling, training, evaluation, promotion, and metered inference. Versioned endpoints identify the exact promoted model and record usage for billing and audit.
- 01Label & govern
Humans review source content under frozen task, quorum, and reward rules.
- 02Train or fine-tune
Eligible evidence enters a reproducible model candidate; task support may be native or external.
- 03Evaluate & promote
Frozen validation and test evidence decides whether a candidate can replace the incumbent.
- 04Serve & meter
A promoted, serving-compatible model receives a versioned endpoint and exact usage receipts.
Contributors complete plain-language missions; owners and admins control project policy, training, and promotion.
Server applications call the operational REST prediction route with bearer auth and required idempotency.
Server-side agents use the same canonical REST contract when hosted inference is enabled for their deployment.
Private or open by contract
Search open models Private organizations keep model artifacts within their authority. Verified open-source organizations commit to publishing tuned checkpoints and their release evidence.
Canonical first path
Versioned REST inference is the first serving contract for promoted object-detection endpoints; deployment capability must report it available before clients call it. Use the service origin supplied for the endpoint; this page does not claim a verified public API hostname.
Agent adapters · deployment-gatedMCP and A2A are implemented projections over the same versioned inference authority. MCP is deployment-gated; A2A is deployment-gated. Protocol versions and migration boundaries live in the API documentation.
Read the operational REST contractReady to label content or build a custom model?
Request access and tell us whether you want to contribute labels, run a model project, or integrate a promoted model.