Global health, connected

Data Commons for Global Health

Curating and connecting data for global health.

Sharing data is hard. Finding data is hard. Let’s make it easier. Share openly, with a trusted group, without moving the data—or simply make it discoverable. Access a centralized Global Health dataset catalog.

They share their data:

EPFL
Vula Medical
Dimagi
IntelliHealth
Kenya Primary Care Network
Rwanda Imaging Federation
Sierra Leone Maternal Knowledge Partnership
Malawi Language Health Collaborative
Tanzania Laboratory Network
Ghana Ultrasound Collaborative
Ethiopia Community Health Programme
Nepal Neonatal Care Collaborative
Supporters and ecosystem partners

Built with global health partners.

LiGHT — Laboratory for intelligent Global Health and Humanitarian Response Technologies
EPFL
Ariadne Labs
Ashoka University
Koita Centerfor Digital Health
Endless
Clinton Health Access Initiative
The Agency Fund
Gates Foundation
Wellcome
Participation Tiers
A progressive ladder. Organizations can join at any point and advance only when ready. No tier is mandatory.
The participation ladder is flexible by design. There is no pressure to advance — a metadata listing at Tier 1 already provides real value to the global health AI community. Each step upward is voluntary, taken only when an organization is ready, has a genuine use case, and has the governance structures in place. Data stays with its owner at every tier.
1

Discovery

Metadata, data cards, documentation, and a contact point. The dataset remains entirely with the owner. Anyone can discover it; no one can access it without direct engagement.

Metadata only
↑
2

Open Data

Dataset made openly available under a clear license with citation terms. Hosted on whichever infrastructure best serves the owner. Downloadable by any researcher.

Open data
↑
3

Gated Community Access

Dataset visible publicly, accessible only to approved researchers or community members. Approval criteria and review process set entirely by the data owner.

Gated community
↑
4

Controlled Sharing

Access governed through Data Transfer Agreements, Data Use Agreements, and institutional approvals. Full audit trail. Formal application, review, and approval pipeline.

DTA required
↑
5

Federated Learning

Data stays local. Models travel to the data using DISCO-style weight-transfer agreements. No raw data leaves the institution. Participates in shared training without any data export.

Federated learning
↑
6

Trusted Research Environment

Highly sensitive data remains in a secure TRE where approved researchers analyze without extracting raw data. Outputs reviewed before leaving the environment. Highest governance assurance.

TRE
Data Commons
Data-card standard. Every listing documents exact contents, provenance, privacy controls, limitations, access, version, and linked models.
No data cards yet

Published sharing requests will appear here in the data-card layout. Share data to add the first listing.

Model Commons
Model-card standard. Every listing documents the task, intended use, data, architecture, evaluation, compute, release, and limitations.
No model cards yet

Published model requests will appear here in the model-card layout. Share a model to add the first listing.

Global health communities
Contributor leaderboard
Contributor recognition

Certified contributor status.

A clear record of what an organization contributed and the identity, documentation, privacy, access, and quality checks it passed.

IdentityDocumentationPrivacyAccessQuality
Documented stewardshipReusable assetsCommunity contributionIndependent validation

Ranks can be compared by evidence score, reach, resource count, or a specific sharing pathway. Trophies name the exact contribution being recognized.

Earn certified contributor status.

Complete a data or model card, pass the contributor checks, and download a reusable credential.

Share data
Start with what you have. We’ll guide the rest.
No storage
Your dataset does not live in the Commons.

Connect its existing repository, or get guided publishing support for Hugging Face or MDC. When cost is a barrier, request no-cost Hugging Face storage.

1

Choose a sharing pathway

Choose a sharing pathway before submitting.

Open data workflow

Connect data that is already published anywhere, or choose a guided publishing pathway. The external repository stores the data—not the Commons.

Open
1
Tell us where it livesConnect an existing public link or choose guided publishing.
2
Choose a licenseReview reuse, attribution, and commercial-use preferences.
3
Check interoperabilityReceive LLM-assisted suggestions for metadata and health-data standards.
4
Complete the data cardDocument exact contents, provenance, limitations, and stewardship.
Open-data details
Use any public dataset page. We use its public metadata to help draft the data card; dataset files stay at the source.
Not sure is a perfectly good answer.
The named repository or custodian provides storage and access controls. The Commons guides the process and keeps only the data card.

Gated or controlled sharing workflow

Define the community, access terms, and agreement that every approved user must sign.

Controlled
1
Confirm community and authorityName the community and who can approve sharing.
2
Choose the agreementUse an existing DUA or DTA, browse a template, or ask us to draft one.
3
Review the termsConfirm permitted use, eligible users, security, retention, and outputs.
4
Collect signaturesProvider and recipient sign before access is activated.
Community and authority
Do you already have a DUA or DTA?Preview our templates at any point. We can also turn a short brief into a draft for review and signature.
Upload or link the existing agreement with the final data card. We will extract the key terms, flag gaps, and configure the access and signature workflow.
Tell us the essentialsWe use these answers to prepare an editable first draft—not a final agreement.
The draft moves through steward and legal review, then all named parties sign before access is activated.
The Commons helps configure the repository and access workflow. It does not receive or store the dataset.

Federated learning workflow

Assess whether the institution can run a model locally, then define the federation protocol and weight-transfer terms.

Data stays local
1
Confirm institutional approvalName the custodian and technical owner.
2
Assess node readinessReview compute, container, network, and audit requirements.
3
Agree the protocolDefine training rounds, secure aggregation, and weight-transfer terms.
4
Publish a node cardDescribe local data and limitations without exposing records.
Federation readiness

Knowledge-base workflow

Build privacy-preserving knowledge from local data, then connect it to a model to give the model context for a specific population or setting—without exposing patient-level records.

No patient rows
1
Define the taskSpecify outcomes, permitted observations, and unit of analysis.
2
Compile locallyDerive task-bounded statistics inside the custodian environment.
3
Apply release controlsSet minimum counts, suppression, rounding, or differential privacy.
4
Validate and versionPackage schema, knowledge, policy, validation, and intended use.
5
Connect to a modelUse the approved knowledge base as contextual input to a linked model.
Knowledge-base scope

Trusted research environment workflow

Describe the secure environment, access review, permitted compute, and output-checking process.

Secure environment
1
Identify the TREName the environment and accountable operator.
2
Define access reviewSpecify who approves users, projects, and compute.
3
Set workspace rulesDocument code ingress, audit logging, and permitted tools.
4
Control outputsDefine disclosure review before results leave.
TRE readiness

Metadata-only starting point

Publish a principled data card so the resource is discoverable. No files, rows, distributions, or example records are required.

No data shared
1
Describe the resourceState exact contents, population, period, steward, and limitations.
2
Publish a contact pointMake the resource discoverable without implying access.
3
Ask for guidanceReview readiness for open, gated, federated, knowledge-base, or TRE sharing.
What help do you need?
Metadata-only publication does not imply that the data are available or that access will be granted.
2

Complete the data card

The Commons does not store your recording. It is used only to populate the data-card fields. You can review or change every answer, and ask questions while you complete the card.
A faster start

Describe the dataset out loud.

Say its name, purpose, contents, population, dates, source, and limitations. We’ll turn the transcript into a first draft. Do not speak names or identifiable patient information.

You can edit everything before submitting.
Optional at this stage. Do not upload patient-level records.
✓

Request received

We will review the data card and follow up with the next steps for your selected pathway.

Share model
Start with what you have. We’ll guide the rest.
No storage
Your model does not live in the Commons.

Connect an existing public host, or get guided publishing support. The Commons keeps the model card—not weights, checkpoints, or training data.

1

Open model workflow

Tell us whether the model is already public. If it is, connect the hosted page. If not, we will help you publish. Weights stay at the host—not in the Commons.

Open
1
Say whether it is publicYes if people can already reach the hosted model; no if you need publishing support.
2
Connect the hostPaste the public model page, or we will guide Hugging Face publishing.
3
Complete the model cardDocument task, data, architecture, evaluation, compute, and limitations.
Open-model details
The named repository or host provides storage and access. The Commons guides the process and keeps only the model card.
2

Complete the model card

The Commons does not store your recording. It is used only to populate the model-card fields. You can review or change every answer before submitting.
A faster start

Describe the model out loud.

Say its name, task, intended use, training data, inputs, outputs, architecture, and limitations. We’ll turn the transcript into a first draft. Do not speak names or identifiable patient information.

You can edit everything before submitting.
✓

Request received

We will review the model card and follow up with the next steps for open publication.

Phase 2 Vision

How Phase 2 Works

Phase 1 produces the pieces. Phase 2 turns those pieces into operating infrastructure. Each flow below shows how Phase 1 assets become Phase 2 capabilities.

Working prototype → Connected commons Data cards → Auto-generation DTA templates → Guided workflows Pilot communities → Community hubs Model homes → Model Commons
Connected Commons Flow
One shared starting point for data, models, communities, and governed access.
Click a step to explore
USR
Contributor or Researcher
Arrives at the Data Commons
Manual
PRT
Data Commons
Navigation, discovery, and onboarding layer
AutomatedGoverned
INF
Backend Infrastructure
Selected platform (HF, MDC, or institutional)
Governed
CRD
Dataset / Model Card
FAIR-compliant metadata, linked to community
Automated
COM
Community Page
Thematic home with benchmarks and contributors
Model-linked
ANL
Analytics Dashboard
Impact tracking, citations, access audits
AutomatedGoverned
← Click a step to see what happens at that stage
Intelligent Onboarding Flow
AI-assisted metadata generation turns a manual Phase 1 process into a guided, near-automated workflow. Future roadmap — not Phase 1.
Click a step to explore
DOC
Upload documentation or metadata
Existing docs, data dictionaries, ethics approvals
Manual
AI
AI extracts draft metadata
NLP over docs to populate fields
Automated
CRD
AI drafts data card
FAIR assessment, FAIR score, missing fields flagged
Automated
REV
Contributor reviews & corrects
Human in the loop — always
ManualGoverned
GOV
Governance tier suggested
License recommendation, access tier, benchmark tasks
AutomatedGoverned
PUB
Published to catalogue & backend
Listed in the commons, synced to selected infrastructure
Automated
← Click a step to see sample outputs
Model Commons Flow
Models are the incentive mechanism. Attribution flows back to every contributing dataset owner.
Click a step to explore
DAT
Dataset listed in commons
Metadata card, access tier, governance pathway
Governed
BNK
Benchmark task defined
Community agrees on evaluation criteria
Manual
TRN
Model trained / fine-tuned
On approved data, with attribution logged
Model-linkedGoverned
EVL
Evaluation run
Benchmark scores computed, compared, published
Automated
HME
Model home published
With scores, datasets, contributors, licence
Model-linked
ATR
Attribution returned
Citations, visibility, collaboration requests
Governed
← Click a step to see the model commons flow
Secure Collaboration Flow
Three pathways for sensitive data — from DTA-governed transfer to federated learning where data never moves, to TRE secure analysis.
DTA pathway — click a step
REQ
Researcher requests access
Via a guided access request
Manual
APR
Data owner approves
Review against data use criteria
ManualGoverned
DTA
DTA generated
From template library, reviewed by both parties
AutomatedGoverned
ACC
Access granted
Credentialed access to dataset on platform
Secure
LOG
Citation / audit logged
Attribution returned to data owner
Governed
← Click a step
Federated learning — click a step
SND
Model sent to institution
Model travels to data, not the other way around
SecureGoverned
TRN
Local training
Model trains on local data — data never leaves
SecureModel-linked
WTA
Weights returned
Only model weights shared under WTA agreement
Governed
AGG
Aggregation
Federated averaging across participating nodes
Automated
GLB
Global model updated
Published to Model Commons with contributor attribution
Model-linkedGoverned
← Click a step
TRE pathway — click a step
TRE
Researcher enters secure environment
Credentialed access, audit trail begins
SecureGoverned
RUN
Analysis runs inside TRE
Code runs on data — data stays in environment
Secure
SDC
Outputs checked
Statistical disclosure control review
ManualGoverned
OUT
Aggregate results released
Summary statistics, models, publications — not raw data
GovernedSecure
← Click a step
Community Expansion Flow
From Phase 1 pilot communities to self-governing thematic hubs with datasets, models, benchmarks, and governance.
Click a step to explore
PLT
Pilot community
2–4 founding organizations, shared purpose
Manual
MBR
Founding members
Define governance, norms, and first benchmarks
Governed
DAT
Datasets listed
Metadata cards, access tiers, governance pathways
Governed
MDL
Models developed
Linked to community data, benchmarked
Model-linked
BNK
Benchmarks & workshops
Community challenges, shared evaluation
Automated
HUB
New contributors & governance
Self-governing community hub, growing network
Governed
Example community hubs

Ultrasound Commons

4 datasets2 modelsCAD4TB benchmarkDTA framework

Primary Care Commons

8 datasets2 modelsCHW triage benchmark5 governance docs

Low-Resource Language Commons

5 datasets3 modelsWER benchmarksOpen access

Maternal Health Commons

6 datasets2 modelsMNCH protocol evalCommunity gated
← Click a step
Sustainability Flow
From Phase 1 Blueprint to a self-sustaining open standard with distributed governance.
Click a step to explore
BP1
Phase 1 Blueprint
Evidence base: what worked, what barriers exist
Manual
BP2
Phase 2 implementation
Connected experience, automated tooling, community scale
Automated
GOV
Governance board
Multi-stakeholder oversight, policies, standards
Governed
RGN
Regional stewards
Africa, South Asia, LMIC-led community stewardship
Governed
STD
Open standards
Governance templates, data card schema, benchmark protocols
Automated
PUB
Sustainable public infrastructure
Self-governing, globally distributed, public-good
Governed
Governance roles

Scientific steward

Scientific architecture, blueprint authorship, technical leadership.

Community stewards

Domain leadership. Thematic community governance.

Data owners

Asset governance. Access control. Consent oversight.

Model contributors

Technical contributions. Benchmark design. Model cards.

Funders

Catalytic support. Phase-gated investment.

International board

Long-term oversight. Policy alignment. Sustainability.

← Click a step

Phase 1 Unlocks Phase 2

Every Phase 1 deliverable is an experiment that teaches something. Phase 2 is built on that evidence, not speculation.

Phase 1 asset What it teaches Phase 2 capability
Commons prototypeUser needs, navigation patterns, discovery frictionConnected commons experience
Metadata onboarding (manual)Where contributors get stuck; what fields matterAI-assisted onboarding pipeline
Data cardsRequired fields, FAIR gaps, missing standardsAutomated data card generator
Platform comparison (HF vs MDC)Infrastructure fit for global health contextSelected, integrated backend architecture
Governance templatesLegal bottlenecks, institutional blockersGuided access workflow engine
Pilot communitiesEngagement model, stewardship needsScalable self-governing community hubs
Model homes (examples)Incentive mechanism, attribution designFull Model Commons with benchmarking
TRE / federated assessmentSecure access needs, institutional readinessSecure collaboration layer